PreprintPDF Available

GAUDI: A Neural Architect for Immersive 3D Scene Generation

Preprints and early-stage research may not have been peer reviewed yet.

Abstract and Figures

We introduce GAUDI, a generative model capable of capturing the distribution of complex and realistic 3D scenes that can be rendered immersively from a moving camera. We tackle this challenging problem with a scalable yet powerful approach, where we first optimize a latent representation that disentangles radiance fields and camera poses. This latent representation is then used to learn a generative model that enables both unconditional and conditional generation of 3D scenes. Our model generalizes previous works that focus on single objects by removing the assumption that the camera pose distribution can be shared across samples. We show that GAUDI obtains state-of-the-art performance in the unconditional generative setting across multiple datasets and allows for conditional generation of 3D scenes given conditioning variables like sparse image observations or text that describes the scene.
Content may be subject to copyright.
GAUDI: A Neural Architect for
Immersive 3D Scene Generation
Miguel Angel BautistaPengsheng GuoSamira Abnar Walter Talbott
Alexander Toshev Zhuoyuan Chen Laurent Dinh Shuangfei Zhai Hanlin Goh
Daniel Ulbricht Afshin Dehghan Josh Susskind
We introduce
, a generative model capable of capturing the distribution of
complex and realistic 3D scenes that can be rendered immersively from a moving
camera. We tackle this challenging problem with a scalable yet powerful approach,
where we first optimize a latent representation that disentangles radiance fields
and camera poses. This latent representation is then used to learn a generative
model that enables both unconditional and conditional generation of 3D scenes.
Our model generalizes previous works that focus on single objects by removing
the assumption that the camera pose distribution can be shared across samples.
We show that GAUDI obtains state-of-the-art performance in the unconditional
generative setting across multiple datasets and allows for conditional generation of
3D scenes given conditioning variables like sparse image observations or text that
describes the scene.
1 Introduction
In order for learning systems to be able to understand and create 3D spaces, progress in generative
models for 3D is sorely needed. The quote "The creation continues incessantly through the media of
humans." is often attributed to Antoni Gaudí, who we pay homage to with our method’s name. We
are interested in generative models that can capture the distribution of 3D scenes and then render
views from scenes sampled from the learned distribution. Extensions of such generative models to
conditional inference problems could have tremendous impact in a wide range of tasks in machine
learning and computer vision. For example, one could sample plausible scene completions that are
consistent with an image observation, or a text description (see Fig. 1 for 3D scenes sampled from
GAUDI). In addition, such models would be of great practical use in model-based reinforcement
learning and planning [12], SLAM [39], or 3D content creation.
Recent works on generative modeling for 3D objects or scenes [
] employ a Generative
Adversarial Network (GAN) where the generator explicitly encodes radiance fields a parametric
function that takes as input the coordinates of a point in 3D space and camera pose, and outputs a
density scalar and RGB value for that 3D point. Images can be rendered from the radiance field
generated by the model by passing the queried 3D points through the volume rendering equation
to project onto any 2D camera view. While compelling on small or simple 3D datasets (e.g. single
denotes equal contribution. Corresponding email:
Preprint. Under review.
arXiv:2207.13751v1 [cs.CV] 27 Jul 2022
sample 3D scene
and poses
conditioned on image
conditioned on text
Figure 1: GAUDI allows to model both conditional and unconditional distributions over complex 3D
scenes. Sampled scenes and poses from (left) the unconditional distribution, and (right) a distribution
conditioned on an image observation or a text prompt.
objects or a small number of indoor scenes), GANs suffer from training pathologies including mode
collapse [
] and are difficult to train on data for which a canonical coordinate system does not
exist, as is the case for 3D scenes [
]. In addition, one key difference between modeling distributions
of 3D objects vs. scenes is that when modeling objects it is often assumed that camera poses are
sampled from a distribution that is shared across objects (i.e. typically over
), which is not true
for scenes. This is because the distribution of valid camera poses depends on each particular scene
independently (based on the structure and location of walls and other objects). In addition, for scenes
this distribution can encompass all poses over the
group. This fact becomes more clear when
we think about camera poses as a trajectory through the scene(cf. Fig. 3(b)).
In GAUDI, we map each trajectory (i.e. a sequence of posed images from a 3D scene) into a latent
representation that encodes a radiance field (e.g. the 3D scene) and camera path in a completely
disentangled way. We find these latent representations by interpreting them as free parameters and
formulating an optimization problem where the latent representation for each trajectory is optimized
via a reconstruction objective. This simple training process is scalable to thousands of trajectories.
Interpreting the latent representation of each trajectory as a free parameter also makes it simple to
handle a large and variable number of views for each trajectory rather than requiring a sophisticated
encoder architecture to pool across a large number of views. After optimizing latent representations
for an observed empirical distribution of trajectories, we learn a generative model over the set of
latent representations. In the unconditional case, the model can sample radiance fields entirely from
the prior distribution learned by the model, allowing it to synthesize scenes by interpolating within the
latent space. In the conditional case, conditional variables available to the model at training time (e.g.
images, text prompts, etc.) can be used to generate radiance fields consistent with those variables.
Our contributions can be summarized as:
We scale 3D scene generation to thousands of indoor scenes containing hundreds of thousands of
images, without suffering from mode collapse or canonical orientation issues during training.
We introduce a novel denoising optimization objective to find latent representations that jointly
model a radiance field and the camera poses in a disentangled manner.
Our approach obtains state-of-the-art generation performance across multiple datasets.
Our approach allows for various generative setups: unconditional generation as well as conditional
on images or text.
2 Related Work
In recent years the field has witnessed outstanding progress in generative modeling for the 2D image
domain, with most approaches focusing either on adversarial [
] or auto-regressive models
]. More recently, score matching based approaches [
] have gained popularity. In
particular, Denoising Diffusion Probabilistic Models (DDPMs) [
] have emerged as
strong contenders to both adversarial and auto-regressive approaches. In DDPMs, the goal is to learn
a step-by-step inversion of a fixed diffusion Markov Chain that gradually transforms an empirical data
distribution to a fixed posterior, which typically takes the form of an isotropic Gaussian distribution.
In parallel, the last couple of years have seen a revolution in how 3D data is represented within neural
networks. By representing a 3D scene as a radiance field, NeRF [
] introduces an approach to
optimize the weights of a MLP to represent the radiance of 3D points that fall inside the field-of-view
of a given set of posed RGB images. Given the radiance for a set of 3D points that lie on a ray
shot from a given camera pose, NeRF [
] uses volumetric rendering to compute the color for the
corresponding pixel and optimizes the MLP weights via a reconstruction loss in image space.
A few attempts have also been made at incorporating a radiance field representation within generative
models. Most approaches have focused on the problem of single objects with known canonical
orientations like faces or Shapenet objects with shared camera pose distributions across samples in a
dataset [
]. Extending these approaches from single objects to completely
unconstrained 3D scenes is an unsolved problem. One paper worth mentioning in this space is GSN
], which breaks the radiance field into a grid of local radiance fields that collectively represent a
scene. While this decomposition of radiance fields endows the model with high representational
capacity, GSN still suffers from the standard training pathologies of GANs, like mode collapse [
which are exacerbated by the fact that unconstrained 3D scenes do not have a canonical orientation.
As we show in our experiments (cf. Sect. 4), these issues become prominent as the training set size
increases, impacting the capacity of the generative model to capture complex distributions. Separately,
a line of recent approaches have also studied the problem of learning generative models of scenes
without employing radiance fields [
]. These works assume that the model has access to
room layouts and a database of object CAD models during training, simplifying the problem of scene
generation to a selection of objects from the database and pose predictions for each object.
Finally, approaches that learn to predict a target view given a single (or multiple) source view and
relative pose transformation have been recently proposed [
]. The pure reconstruction
objective employed by these approaches forces them to learn a deterministic conditional function that
maps a source image and a relative camera transformation to a target image. The first is that this scene
completion problem is ill-posed (e.g. given a single source view of a scene there are multiple target
completions that are equally likely). Attempts at modeling the problem in a probabilistic manner have
been proposed [
]. However, these approaches suffer from inconsistency in predicted scenes
because they do not explicitly model a 3D consistent representation like a radiance field.
Our goal is to learn a generative model given an empirical distribution of trajectories over 3D scenes.
denote a collection of examples defining an empirical distribution, where
each example
is a trajectory. Every trajectory
is defined as a variable length sequence of
corresponding RGB, depth images and 6DOF camera poses (see Fig. 3).
We decompose the task of learning a generative model in two stages. First, we obtain a latent
z= [zscene,zpose ]
for each example
that represents the scene radiance field
and pose in separate disentangled vectors. Second, given a set of latents
we learn
the distribution p(Z).
3.1 Optimizing latent representations for radiance fields and camera poses
We now turn to the task of finding a latent representation
for each example
(i.e. for each
trajectory in the empirical distribution). To obtain this latent representation we take an encoder-less
view and interpret
’s as free parameters to be found via an optimization problem [
]. To map
to trajectories
, we design a network architecture (i.e. a decoder) that disentangles camera
poses and radiance field parameterization. Our decoder architecture is composed of 3 networks
(shown in Fig. 2):
camera pose decoder
(parameterized by
), is responsible for predicting camera
at the normalized temporal position
in the trajectory, conditioned on
which represents the camera poses for the whole trajectory. To ensure that the output of
a valid camera pose (e.g. an element of
), we output a
D vector representing a normalized
quaternion qsfor the orientation and a 3D translation vector ts.
<latexit sha1_base64="B06wqsKluLjelLKASviSbUT2vTo=">AAAB6HicbVBNS8NAEJ3Ur1q/qh69LBbBU0lE1GNBBI8t2A9oQ9lsJ+3azSbsboQS+gu8eFDEqz/Jm//GbZuDtj4YeLw3w8y8IBFcG9f9dgpr6xubW8Xt0s7u3v5B+fCopeNUMWyyWMSqE1CNgktsGm4EdhKFNAoEtoPx7cxvP6HSPJYPZpKgH9Gh5CFn1FipEfbLFbfqzkFWiZeTCuSo98tfvUHM0gilYYJq3fXcxPgZVYYzgdNSL9WYUDamQ+xaKmmE2s/mh07JmVUGJIyVLWnIXP09kdFI60kU2M6ImpFe9mbif143NeGNn3GZpAYlWywKU0FMTGZfkwFXyIyYWEKZ4vZWwkZUUWZsNiUbgrf88ippXVS9q6rXuKzU7vI4inACp3AOHlxDDe6hDk1ggPAMr/DmPDovzrvzsWgtOPnMMfyB8/kDzdCM8w==</latexit>
<latexit sha1_base64="Oforsn3qXl3pFlZU0Yiyb1ETWWY=">AAAB/3icbVBNS8NAEN34WetXVPDiZbEIHqQkItpjwYvHCvYDmlA22027dDcJuxOxxBz8K148KOLVv+HNf+O2zUFbHww83pthZl6QCK7Bcb6tpeWV1bX10kZ5c2t7Z9fe22/pOFWUNWksYtUJiGaCR6wJHATrJIoRGQjWDkbXE799z5TmcXQH44T5kgwiHnJKwEg9+9DLPM0HkpxhD9gDBGFGci/v2RWn6kyBF4lbkAoq0OjZX14/pqlkEVBBtO66TgJ+RhRwKlhe9lLNEkJHZMC6hkZEMu1n0/tzfGKUPg5jZSoCPFV/T2REaj2WgemUBIZ63puI/3ndFMKan/EoSYFFdLYoTAWGGE/CwH2uGAUxNoRQxc2tmA6JIhRMZGUTgjv/8iJpnVfdy6p7e1Gp14o4SugIHaNT5KIrVEc3qIGaiKJH9Ixe0Zv1ZL1Y79bHrHXJKmYO0B9Ynz8s7JYu</latexit>
<latexit sha1_base64="9Vfm4HmMOO/t/Aw5dEEp3fZWGo0=">AAAB6HicbVBNS8NAEJ3Ur1q/qh69LBbBU0mkqMeCCB5bsLXQhrLZTNq1m03Y3Qil9Bd48aCIV3+SN/+N2zYHbX0w8Hhvhpl5QSq4Nq777RTW1jc2t4rbpZ3dvf2D8uFRWyeZYthiiUhUJ6AaBZfYMtwI7KQKaRwIfAhGNzP/4QmV5om8N+MU/ZgOJI84o8ZKzbBfrrhVdw6ySrycVCBHo1/+6oUJy2KUhgmqdddzU+NPqDKcCZyWepnGlLIRHWDXUklj1P5kfuiUnFklJFGibElD5urviQmNtR7Hge2MqRnqZW8m/ud1MxNd+xMu08ygZItFUSaIScjsaxJyhcyIsSWUKW5vJWxIFWXGZlOyIXjLL6+S9kXVu6zWmrVK/TaPowgncArn4MEV1OEOGtACBgjP8ApvzqPz4rw7H4vWgpPPHMMfOJ8/y76M9A==</latexit>
<latexit sha1_base64="B4F/f67BKYEZVOAvDxXEu4giNAw=">AAAB/XicbVDJSgNBFOyJW4zbuNy8NAbBU5iRoB6DXjxGNAskQ+jpvCRNeha634hxCP6KFw+KePU/vPk3dpI5aGJBQ1H1in6v/FgKjY7zbeWWlldW1/LrhY3Nre0de3evrqNEcajxSEaq6TMNUoRQQ4ESmrECFvgSGv7wauI37kFpEYV3OIrBC1g/FD3BGRqpYx+0ER4wveUQApUMIUQ97thFp+RMQReJm5EiyVDt2F/tbsSTwKS5ZFq3XCdGL2UKBZcwLrQTDTHjQ9aHlqEhC0B76XT7MT02Spf2ImVeiHSq/k6kLNB6FPhmMmA40PPeRPzPayXYu/BSEcaJOYvPPuolkmJEJ1XQrlDAUY4MYVwJsyvlA6YYR1NYwZTgzp+8SOqnJfesVL4pFyuXWR15ckiOyAlxyTmpkGtSJTXCySN5Jq/kzXqyXqx362M2mrOyzD75A+vzBwrKlZ4=</latexit>
Scene latents
<latexit sha1_base64="NHeYyMj21rU7fOCAq1xqGvdJFGc=">AAAB/HicbVDLSgNBEOyNrxhfqzl6GQyCp7ArQT0GvXiMYB6QLGF2MpsMmX0w0yuGJf6KFw+KePVDvPk3TpI9aGLBQFHVxXSXn0ih0XG+rcLa+sbmVnG7tLO7t39gHx61dJwqxpsslrHq+FRzKSLeRIGSdxLFaehL3vbHNzO//cCVFnF0j5OEeyEdRiIQjKKR+na5h/wRs0asOZEUeYR62rcrTtWZg6wSNycVyNHo21+9QczS0KSZpFp3XSdBL6MKBZN8WuqlmieUjemQdw2NaMi1l82Xn5JTowxIECvzIiRz9Xcio6HWk9A3kyHFkV72ZuJ/XjfF4MrLRJSk5iy2+ChIJcGYzJogA6E4QzkxhDIlzK6EjaiiDE1fJVOCu3zyKmmdV92Lau2uVqlf53UU4RhO4AxcuIQ63EIDmsBgAs/wCm/Wk/VivVsfi9GClWfK8AfW5w9aoZU9</latexit>
Pose latents
<latexit sha1_base64="b9efxyx94Kmxhn7kIWr/5HWKRiE=">AAACAnicbVBNS8NAEN3Ur1q/qp7ES7AInkoiRT0WvXisYD+gLWWznbRLN5uwOxFrCF78K148KOLVX+HNf+Om7UFbHww83pthZp4XCa7Rcb6t3NLyyupafr2wsbm1vVPc3WvoMFYM6iwUoWp5VIPgEurIUUArUkADT0DTG11lfvMOlOahvMVxBN2ADiT3OaNopF7xoBNQHHp+8pD2kg7CPSaagYQ07RVLTtmZwF4k7oyUyAy1XvGr0w9ZHIBEJqjWbdeJsJtQhZwJSAudWENE2YgOoG2opAHobjJ5IbWPjdK3/VCZkmhP1N8TCQ20Hgee6cwO1vNeJv7ntWP0L7oJl1GMINl0kR8LG0M7y8PucwUMxdgQyhQ3t9psSBVlaFIrmBDc+ZcXSeO07J6VKzeVUvVyFkeeHJIjckJcck6q5JrUSJ0w8kieySt5s56sF+vd+pi25qzZzD75A+vzB+t9mGw=</latexit>
<latexit sha1_base64="ZgJJTUJklOpDQOWM8P89PPK7Ft4=">AAAB9HicbVBNSwMxEJ31s9avqkcvwSJ40LIrRT0WvXisYD9gu5Rsmm1Ds8maZAtl6e/w4kERr/4Yb/4b03YP2vpg4PHeDDPzwoQzbVz321lZXVvf2CxsFbd3dvf2SweHTS1TRWiDSC5VO8SaciZowzDDaTtRFMchp61weDf1WyOqNJPi0YwTGsS4L1jECDZWCjTqMIH8C+8ceUG3VHYr7gxomXg5KUOOerf01elJksZUGMKx1r7nJibIsDKMcDopdlJNE0yGuE99SwWOqQ6y2dETdGqVHoqksiUMmqm/JzIcaz2OQ9sZYzPQi95U/M/zUxPdBBkTSWqoIPNFUcqRkWiaAOoxRYnhY0swUczeisgAK0yMzaloQ/AWX14mzcuKd1WpPlTLtds8jgIcwwmcgQfXUIN7qEMDCDzBM7zCmzNyXpx352PeuuLkM0fwB87nD0lhkH4=</latexit>
<latexit sha1_base64="dLTR5EUYNU8gMh6V547gIwc7bSU=">AAAB83icbVBNS8NAEJ3Ur1q/qh69LBbBU0lKUY8FLx4r2A9oQtlstu3SzSbsToRa+je8eFDEq3/Gm//GbZuDtj4YeLw3w8y8MJXCoOt+O4WNza3tneJuaW//4PCofHzSNkmmGW+xRCa6G1LDpVC8hQIl76aa0ziUvBOOb+d+55FrIxL1gJOUBzEdKjEQjKKV/Kd+RHwUMTdE9csVt+ouQNaJl5MK5Gj2y19+lLAs5gqZpMb0PDfFYEo1Cib5rORnhqeUjemQ9yxV1K4JpoubZ+TCKhEZJNqWQrJQf09MaWzMJA5tZ0xxZFa9ufif18twcBNMhUoz5IotFw0ySTAh8wBIJDRnKCeWUKaFvZWwEdWUoY2pZEPwVl9eJ+1a1buq1u/rlUYtj6MIZ3AOl+DBNTTgDprQAgYpPMMrvDmZ8+K8Ox/L1oKTz5zCHzifP4z3kVQ=</latexit>
<latexit sha1_base64="R7u0OTHX7R25bJQnDAvr4GeMunk=">AAACBHicbVBNS8NAEN34WetX1GMvi0XwVJIq6rHgxWMV+wFNLZvtpl262YTdiVhCD178K148KOLVH+HNf+OmzUFbHww83pthZp4fC67Bcb6tpeWV1bX1wkZxc2t7Z9fe22/qKFGUNWgkItX2iWaCS9YADoK1Y8VI6AvW8keXmd+6Z0rzSN7COGbdkAwkDzglYKSeXfKAPYAfpPEEe1xiLyQw9P30ZnJ30rPLTsWZAi8SNydllKPes7+8fkSTkEmggmjdcZ0YuilRwKlgk6KXaBYTOiID1jFUkpDpbjp9YoKPjNLHQaRMScBT9fdESkKtx6FvOrMb9byXif95nQSCi27KZZwAk3S2KEgEhghnieA+V4yCGBtCqOLmVkyHRBEKJreiCcGdf3mRNKsV96ziXJ+Wa9U8jgIqoUN0jFx0jmroCtVRA1H0iJ7RK3qznqwX6936mLUuWfnMAfoD6/MHu02YFw==</latexit>
<latexit sha1_base64="Eyd63NoWcXzHECFQXtVg1Nbm8e4=">AAAB/3icbVDLSsNAFJ3UV62vquDGTbAIrkoiRV0W3bisYB/QljKZ3rRDJ5MwcyPWmIW/4saFIm79DXf+jZO2C209MHA4517umeNFgmt0nG8rt7S8srqWXy9sbG5t7xR39xo6jBWDOgtFqFoe1SC4hDpyFNCKFNDAE9D0RleZ37wDpXkob3EcQTegA8l9zigaqVc86AQUh56fPKS9DsI9JlGoIe0VS07ZmcBeJO6MlMgMtV7xq9MPWRyARCao1m3XibCbUIWcCUgLnVhDRNmIDqBtqKQB6G4yyZ/ax0bp236ozJNoT9TfGwkNtB4HnpnM0up5LxP/89ox+hfdhMsoRpBsesiPhY2hnZVh97kChmJsCGWKm6w2G1JFGZrKCqYEd/7Li6RxWnbPypWbSql6OasjTw7JETkhLjknVXJNaqROGHkkz+SVvFlP1ov1bn1MR3PWbGef/IH1+QNiVJb/</latexit>
<latexit sha1_base64="f2Q76/2rshYrT+GfRna54t8/VoY=">AAAB+XicbVDLSsNAFL3xWesr6tLNYBFclaQUdVlw47KCfUAbwmQ6aYdOHsxMqjHkT9y4UMStf+LOv3HSZqGtBwYO59zLPXO8mDOpLOvbWFvf2NzaruxUd/f2Dw7No+OujBJBaIdEPBJ9D0vKWUg7iilO+7GgOPA47XnTm8LvzaiQLArvVRpTJ8DjkPmMYKUl1zSHAVYTz88ecjd7TJ9y16xZdWsOtErsktSgRNs1v4ajiCQBDRXhWMqBbcXKybBQjHCaV4eJpDEmUzymA01DHFDpZPPkOTrXygj5kdAvVGiu/t7IcCBlGnh6ssgpl71C/M8bJMq/djIWxomiIVkc8hOOVISKGtCICUoUTzXBRDCdFZEJFpgoXVZVl2Avf3mVdBt1+7LevGvWWo2yjgqcwhlcgA1X0IJbaEMHCMzgGV7hzciMF+Pd+FiMrhnlzgn8gfH5A53zlEM=</latexit>
<latexit sha1_base64="nU9dxt1jdC9rB96D2aENDeN3xws=">AAAB+HicbVDLSsNAFL2pr1ofjbp0EyyCq5JIUZdFNy4r2Ae0IUymk3boZBJmJmIM+RI3LhRx66e482+ctFlo64GBwzn3cs8cP2ZUKtv+Nipr6xubW9Xt2s7u3n7dPDjsySgRmHRxxCIx8JEkjHLSVVQxMogFQaHPSN+f3RR+/4EISSN+r9KYuCGacBpQjJSWPLM+CpGa+kHWz73sMc09s2E37TmsVeKUpAElOp75NRpHOAkJV5ghKYeOHSs3Q0JRzEheGyWSxAjP0IQMNeUoJNLN5sFz61QrYyuIhH5cWXP190aGQinT0NeTRUy57BXif94wUcGVm1EeJ4pwvDgUJMxSkVW0YI2pIFixVBOEBdVZLTxFAmGlu6rpEpzlL6+S3nnTuWi27lqN9nVZRxWO4QTOwIFLaMMtdKALGBJ4hld4M56MF+Pd+FiMVoxy5wj+wPj8AYzrk68=</latexit>
<latexit sha1_base64="W9vlVYQi9fJgNTIWlP4lj5hL8aw=">AAACKXicbVDLSsNAFJ34rPUVdelmsAgupCRS1I1QdOOygn1AGsJkOmmHTh7MTKRpyO+48VfcKCjq1h9x0gasrQcGzpxzL/fe40aMCmkYn9rS8srq2nppo7y5tb2zq+/tt0QYc0yaOGQh77hIEEYD0pRUMtKJOEG+y0jbHd7kfvuBcEHD4F4mEbF91A+oRzGSSnL0etdHcuB6aTuDV9D6/TnpKMlO4ayQjOeE0TizHb1iVI0J4CIxC1IBBRqO/trthTj2SSAxQ0JYphFJO0VcUsxIVu7GgkQID1GfWIoGyCfCTieXZvBYKT3ohVy9QMKJOtuRIl+IxHdVZb6nmPdy8T/PiqV3aac0iGJJAjwd5MUMyhDmscEe5QRLliiCMKdqV4gHiCMsVbhlFYI5f/IiaZ1VzfNq7a5WqV8XcZTAITgCJ8AEF6AObkEDNAEGj+AZvIF37Ul70T60r2npklb0HIA/0L5/AOpHqFA=</latexit>
W=[Wxy,Wyz ,Wxz]
<latexit sha1_base64="USWVdL0K/kjkYXofrI11Wh/I3ko=">AAAB+HicbVDLSsNAFL2pr1ofjbp0M1gEVyURUZdFNy4r2Ae0IUymk3boZBJmJmIb+iVuXCji1k9x5984abPQ1gMDh3Pu5Z45QcKZ0o7zbZXW1jc2t8rblZ3dvf2qfXDYVnEqCW2RmMeyG2BFORO0pZnmtJtIiqOA004wvs39ziOVisXiQU8S6kV4KFjICNZG8u1qP8J6FIRZZ+ZnT9OZb9ecujMHWiVuQWpQoOnbX/1BTNKICk04VqrnOon2Miw1I5zOKv1U0QSTMR7SnqECR1R52Tz4DJ0aZYDCWJonNJqrvzcyHCk1iQIzmcdUy14u/uf1Uh1eexkTSaqpIItDYcqRjlHeAhowSYnmE0MwkcxkRWSEJSbadFUxJbjLX14l7fO6e1m/uL+oNW6KOspwDCdwBi5cQQPuoAktIJDCM7zCmzW1Xqx362MxWrKKnSP4A+vzB45wk7A=</latexit>
<latexit sha1_base64="rW4YXcW3C/NVVJyOZc+yWWsWYLA=">AAAB+HicbVDLSsNAFL2pr1ofjbp0EyyCq5JIUZdFNy4r2Ae0IUymk3boZBJmJkIa8iVuXCji1k9x5984abPQ1gMDh3Pu5Z45fsyoVLb9bVQ2Nre2d6q7tb39g8O6eXTck1EiMOniiEVi4CNJGOWkq6hiZBALgkKfkb4/uyv8/hMRkkb8UaUxcUM04TSgGCkteWZ9FCI19YOsn3tZOs89s2E37QWsdeKUpAElOp75NRpHOAkJV5ghKYeOHSs3Q0JRzEheGyWSxAjP0IQMNeUoJNLNFsFz61wrYyuIhH5cWQv190aGQinT0NeTRUy56hXif94wUcGNm1EeJ4pwvDwUJMxSkVW0YI2pIFixVBOEBdVZLTxFAmGlu6rpEpzVL6+T3mXTuWq2HlqN9m1ZRxVO4QwuwIFraMM9dKALGBJ4hld4M+bGi/FufCxHK0a5cwJ/YHz+AI/2k7E=</latexit>
<latexit sha1_base64="rl8UgLe6+YP+XOsJk2KEgugJIWI=">AAACF3icbVDLSsNAFJ3UV62vqEs3wSK4ComIuhGKblxW6AvaECbTSTt08nDmRigxf+HGX3HjQhG3uvNvnLQpaOuBgXPOvZe593gxZxIs61srLS2vrK6V1ysbm1vbO/ruXktGiSC0SSIeiY6HJeUspE1gwGknFhQHHqdtb3Sd19v3VEgWhQ0Yx9QJ8CBkPiMYlOXqZm+IIe0FGIaenzayzJWX3Zm8U+phJkAJx9WrlmlNYCwSuyBVVKDu6l+9fkSSgIZAOJaya1sxOCkWwAinWaWXSBpjMsID2lU0xAGVTjq5KzOOlNM3/EioF4IxcX9PpDiQchx4qjNfUs7XcvO/WjcB/8JJWRgnQEMy/chPuAGRkYdk9JmgBPhYEUwEU7saZIgFJqCirKgQ7PmTF0nrxLTPzNPb02rtqoijjA7QITpGNjpHNXSD6qiJCHpEz+gVvWlP2ov2rn1MW0taMbOP/kD7/AGtBqGD</latexit>
Figure 2: Architecture of the decoder model that disentangles camera poses from 3D geometry and
appearance of the scene. Our decoder is composed by 3 submodules. A decoder
that takes as input
a latent code representing the scene
and produces a factorized representation of 3D space via
a tri-plane latent encoding
. A radiance field network
that takes as input points
is conditioned on
to predict a density
and a signal
to be rendered via volumetric rendering
(Eq. 1). Finally, we decode the camera poses through a network
that takes as input a normalized
temporal position
and is conditioned on
which represents camera poses for the
whole trajectory xto predict the camera pose ˆ
scene decoder
(parameterized by
), is responsible for predicting a conditioning
variable for the radiance field network
. This network takes as input a latent code that represents the
and predicts an axis-aligned tri-plane representation [
. Which
correspond to 3 feature maps
[Wxy,Wxz ,Wy z]
of spatial dimension
channels, one
for each axis aligned plane: xy,xz and yz.
radiance field decoder
(parameterized by
), is tasked with reconstructing image
level targets using the volumetric rendering equation in Eq. 1. The input to
and the
tri-plane representation
W= [Wxy,Wxz ,Wy z]
. Given a
D point
p= [i, j, k]
for which radiance
is to be predicted, we orthogonally project
into each plane in
and perform bi-linear sampling.
We concatenate the 3 bi-linearly sampled vectors into
wxyz = [Wxy (i, j),Wxz (j, k ),Wyz(i, k)]
, which is used to condition the radiance field function
. We implement
as a MLP that outputs
a density value
and a signal
. To predict the value
of a pixel, the volumetric rendering equation
is used (cf. Eq. 1) where a 3D point is expressed as ray direction
(corresponding with the pixel
location) at particular depth u.
v(r,W) = Zuf
T r(u)σ(r(u),wxyz )a(r(u),d,wxyz )du
T r(u) = exp Zu
σ(r(u),wxyz )du.(1)
We formulate a denoising reconstruction objective to jointly optimize for
shown in Eq. 2. Note that while latents
are optimized for each example
independently, the
parameters of the networks
are amortized across all examples
. As opposed to
previous auto-decoding approaches [
], each latent
is perturbed during training with additive
noise that is proportional to the empirical standard deviation across all latents,
inducing a contractive representation [
]. In this setting,
controls the trade-off between the entropy
of the distribution
and the reconstruction term, with
β= 0
the distribution of
’s becomes a
collection of indicator functions, whereas non-trivial structure in latent space arises for
β > 0
. We
use a small
β > 0
value to enforce a latent space in which interpolated samples (or samples that
contain small deviations from the empirical distribution, as the ones that one might get from sampling
a subsequent generative model) are included in the support of the decoder.
θdfc,Z ExXLscene(xim
s,zscene,Ts) + λLpose (Ts,zpose , s)(2)
We optimize parameters
θd, θf, θc
and latents
with two different losses. The first loss function
measures the reconstruction between the radiance field encoded in
and the images in
the trajectory
denotes the normalized temporal position of the frame in the trajectory),
given ground-truth camera poses
required for rendering. We use an
loss for RGB and
. The second loss function
measures the camera pose reconstruction error between
the poses
encoded in
and the ground-truth poses. We employ an
loss on translation and
loss for the normalized quaternion part of the camera pose. Although theoretically normalized
quaternions are not necessarily unique (e.g.
) we do not observe any issues empirically
during training.
3.2 Prior Learning
Given a set of latents
resulting from minimizing the objective in Eq. 2, our goal is to learn a
generative model
that captures their distribution (i.e. after minimizing the objective in Eq. 2
we interpret
as examples from an empirical distribution in latent space). In order to model
p(Z)we employ a Denoising Diffusion Probabilistic Model (DDPM) [15], a recent score-matching
] based model that learns to reverse a diffusion Markov Chain with a large but finite number of
timesteps. In DDPMs [
] it is shown that this reverse process is equivalent to learning a sequence of
denoising auto-encoders with tied weights. The supervised denoising objective in DDPMs makes
simple and scalable. This allows us to learn a powerful generative model that enables
both unconditional and conditional generation of 3D scenes. For training our prior
we take
the objective function in [
] defined in Eq. 3. In Eq. 3
denotes the timestep,
the noise and
is a noise magnitude parameter with a fixed scheduling. Finally,
denotes the
denoising model.
θpEt,zZ,∼N (0,I)kθp(¯αtz+1¯αt, t)k2(3)
At inference time, we sample
by following the inference process in DDPMs. We start by
zT N(0,I)
and iteratively apply
to gradually denoise
, thus reversing the diffusion
Markov Chain to obtain
. We then feed
as input to the decoder architecture (cf. Fig. 2) and
reconstruct a radiance field and a camera path.
If the goal is to learn a conditional distribution of the latents
, given paired data
{zZ, y
, the denoising model
is augmented with a conditioning variable
, resulting in
θp(z, t, y)
implementation details about how the conditioning variable is used in the denoising architecture can
be found in the appendix C.
4 Experiments
In this section we show the applicability of GAUDI to multiple problems. First, we evaluate
reconstruction quality and performance of the reconstruction stage. Then, we evaluate the performance
of our model in generative tasks including unconditional and conditional inference, in which radiance
fields are generated from conditioning variables corresponding to images or text prompts. Full
experimental settings and details can be found in the appendix B.
4.1 Data
We report results on 4 datasets: Vizdoom [
], Replica [
], VLN-CE [
] and ARKit Scenes [
which vary in number of scenes and complexity (see Fig. 3 and Tab. 1).
]: Vizdoom is a synthetic simulated environment with simple texture and geometry. We
use the data provided by [
] to train our model. It is the simplest dataset in terms of number of scenes
and trajectories, as well as texture, serving as a test bed to examine GAUDI in the simplest setting.
]: Replica is a dataset comprised of
realistic scenes from which trajectories are
rendered via Habitat [55]. We used the data provided by [7] to train our model.
]: VLN-CE is a dataset originally designed for vision and language navigation in
continuous environments. This dataset is composed of
K trajectories of an agent navigating
between two points in a 3D scene from the 3D dataset [
]. We render observations via Habitat [
Notably, this dataset contains also textual descriptions of the trajectories taken by an agent. In Sect.
4.5 we train GAUDI in a conditional manner to generate 3D scenes given a description.
1We obtain depth predictions by aggregating densities across a ray as in [29]
<latexit sha1_base64="o2hvLYVsBn1LjhLmQoNUQGSR9wI=">AAAB8XicbVDLSgNBEJz1GeMr6tHLYBDiJexKUI9BLx4jmAcmS5iddJIhs7PLTK8YlvyFFw+KePVvvPk3TpI9aGJBQ1HVTXdXEEth0HW/nZXVtfWNzdxWfntnd2+/cHDYMFGiOdR5JCPdCpgBKRTUUaCEVqyBhYGEZjC6mfrNR9BGROoexzH4IRso0RecoZUeOghPmJbY2aRbKLpldwa6TLyMFEmGWrfw1elFPAlBIZfMmLbnxuinTKPgEib5TmIgZnzEBtC2VLEQjJ/OLp7QU6v0aD/SthTSmfp7ImWhMeMwsJ0hw6FZ9Kbif147wf6VnwoVJwiKzxf1E0kxotP3aU9o4CjHljCuhb2V8iHTjKMNKW9D8BZfXiaN87J3Ua7cVYrV6yyOHDkmJ6REPHJJquSW1EidcKLIM3klb45xXpx352PeuuJkM0fkD5zPH1GqkLI=</latexit>
<latexit sha1_base64="+Ut9sqvjAWv9B9ydcKbrsgVfHvw=">AAAB8XicbVDLSgNBEJz1GeMr6tHLYBDiJexKUI9BLx4jmAcmS5iddJIhs7PLTK8YlvyFFw+KePVvvPk3TpI9aGJBQ1HVTXdXEEth0HW/nZXVtfWNzdxWfntnd2+/cHDYMFGiOdR5JCPdCpgBKRTUUaCEVqyBhYGEZjC6mfrNR9BGROoexzH4IRso0RecoZUeOghPmJaCs0m3UHTL7gx0mXgZKZIMtW7hq9OLeBKCQi6ZMW3PjdFPmUbBJUzyncRAzPiIDaBtqWIhGD+dXTyhp1bp0X6kbSmkM/X3RMpCY8ZhYDtDhkOz6E3F/7x2gv0rPxUqThAUny/qJ5JiRKfv057QwFGOLWFcC3sr5UOmGUcbUt6G4C2+vEwa52Xvoly5qxSr11kcOXJMTkiJeOSSVMktqZE64USRZ/JK3hzjvDjvzse8dcXJZo7IHzifP1MwkLM=</latexit>
Figure 3: (a) Examples of the 4 datasets we use in this paper (from left to right): Vizdoom [
Replica [
], VLN-CE [
], ARKitScenes [
]. (b) Top-down views of 2 different camera paths in
]. Blue and red dots represent start-end positions and the camera path is highlighted in
]: ARKitScenes is a dataset of scans of indoor spaces. This dataset contains more
K scans of about
K different indoor spaces. As opposed to the previous datasets where
RGB, depth and camera poses are obtained via rendering in a simulation (i.e. either Vizdoom [
] or
Habitat [
]), ARKitScenes provides raw RGB and depth of the scans and camera poses estimated
using ARKit SLAM. In addition, whereas trajectories from the previous datasets are point-to-point,
as typically done in navigation, the camera trajectories for ARKitScenes resembles a natural scan a
of full indoor space. In our experiments we use a subset of
K scans from ARKitScenes to train our
4.2 Reconstruction
We first validate the hypothesis that the optimization problem described in Eq. 2 can find latent codes
that are able reconstruct the trajectories in the empirical distribution in a satisfactory way. In Tab. 1
we report reconstruction performance of our model across all datasets. Fig. 4 shows reconstructions
of random trajectories for each dataset. For all our experiments we set the dimension of
to 2048 and
β= 0.1
unless otherwise stated. During training, we normalize camera poses for
each trajectory so that the middle frame in a trajectory becomes the origin of the coordinate system.
See appendix E for ablation experiments.
Figure 4: Qualitative reconstruction results of random trajectories on different datasets (one for each
column): Vizdoom [
], Replica [
], VLN-CE [
]and ARKitScenes [
]. For each pair of images
the left is ground-truth and right is reconstruction.
#sc-#tr-#im l1PSNR SSIM Rot Err. Trans. Err
Vizdoom [21] 1-32-1k 0.004 44.42 0.98 0.01 1.26
Replica [60] 18-100-1k 0.006 38.86 0.99 0.03 0.01
VLN-CE [23] 90-3.6k-600k 0.031 25.17 0.73 0.30 0.02
ARKitScenes [1] 300-1k-600k 0.039 24.51 0.76 0.16 0.04
Table 1: Reconstruction results of the optimization process described in Eq. 2. The first column
shows the number of scenes (#sc), trajectories (#tr) and images (#im) per dataset. Due to the large
number of images on VLN-CE [
] and ARKitScenes [
] datasets we sample 10 random images per
trajectory to compute the reconstruction metrics.
4.3 Interpolation
In addition, to evaluate the structure of the latent representation obtained from minimizing the
optimization problem in Eq. 2, we show interpolation results between pairs of latents
Fig. 5. To render images while interpolating the scene we place a fixed camera at the origin of the
coordinate system. We observe a smooth transition of scenes in both geometry (walls, ceilings) and
texture (stairs, carpets). More visualizations are included in the appendix E.1.
<latexit sha1_base64="MazdHGmt0ni1F4NWKhF8/6B9jMg=">AAAB83icbVDLSgMxFL2pr1pfVZdugkVwVWakqMuiG5cV7AM6Q8mkmTY0kxmSjFCH/oYbF4q49Wfc+Tdm2llo64HA4Zx7uScnSATXxnG+UWltfWNzq7xd2dnd2z+oHh51dJwqyto0FrHqBUQzwSVrG24E6yWKkSgQrBtMbnO/+8iU5rF8MNOE+REZSR5ySoyVPC8iZhyE2dNswAfVmlN35sCrxC1IDQq0BtUvbxjTNGLSUEG07rtOYvyMKMOpYLOKl2qWEDohI9a3VJKIaT+bZ57hM6sMcRgr+6TBc/X3RkYiradRYCfzjHrZy8X/vH5qwms/4zJJDZN0cShMBTYxzgvAQ64YNWJqCaGK26yYjoki1NiaKrYEd/nLq6RzUXcv6437Rq15U9RRhhM4hXNw4QqacActaAOFBJ7hFd5Qil7QO/pYjJZQsXMMf4A+fwCFPJID</latexit>
<latexit sha1_base64="46flhQNvRHXdoMAYU5TQvvWVxeQ=">AAAB83icbVDLSgMxFL1TX7W+qi7dBIvgqsyIqMuiG5cV7AM6Q8mkmTY2kwxJRqhDf8ONC0Xc+jPu/Bsz7Sy09UDgcM693JMTJpxp47rfTmlldW19o7xZ2dre2d2r7h+0tUwVoS0iuVTdEGvKmaAtwwyn3URRHIecdsLxTe53HqnSTIp7M0loEOOhYBEj2FjJ92NsRmGUPU37D/1qza27M6Bl4hWkBgWa/eqXP5AkjakwhGOte56bmCDDyjDC6bTip5ommIzxkPYsFTimOshmmafoxCoDFEllnzBopv7eyHCs9SQO7WSeUS96ufif10tNdBVkTCSpoYLMD0UpR0aivAA0YIoSwyeWYKKYzYrICCtMjK2pYkvwFr+8TNpnde+ifn53XmtcF3WU4QiO4RQ8uIQG3EITWkAggWd4hTcndV6cd+djPlpyip1D+APn8weGwJIE</latexit>
Figure 5: Interpolation of 3D scenes in latent space (e.g. interpolating the encoded radiance field) for
the VLN-CE dataset [23]. Each row corresponds to a different interpolation path.
4.4 Unconditional generative modeling
Given latent representations
that can reconstruct samples
with high accuracy as shown
in Sect. 4.2, we now evaluate the capacity of the prior
to capture the empirical distribution
x X
by learning the distribution of latents
. To do so we sample
by following
the inference process in DDPMs, and then feed
through the decoder network, which results in
trajectories of RGB images that are then used for evaluation. We compare our approach with the
following baselines: GRAF [
-GAN [
] and GSN [
]. We sample
k images from predicted
and target distributions for each model and dataset and report both FID [
] and SwAV-FID [
scores. We report quantitative results in Tab. 2, where we can see that GAUDI obtains state-of-the-art
performance across all datasets and metrics. We attribute this performance improvement to the fact
that GAUDI learns disentangled yet corresponding latents for radiance fields and camera poses, which
is key when modeling scenes (see ablations in the appendix E). We note that to obtain these great
empirical results GAUDI needs to simultaneously find latents with high reconstruction fidelity while
also efficiently learning their distribution.
VizDoom [21] Replica [60] VLN-CE [23] ARKitScenes [1]
GRAF [56] 47.50 ±2.13 5.44 ±0.43 65.37 ±1.64 5.76 ±0.14 90.43 ±4.83 8.65 ±0.27 87.06 ±9.99 13.44 ±0.26
π-GAN[5] 143.55 ±4.81 15.26 ±0.15 166.55 ±3.61 13.17 ±0.20 151.26 ±4.19674 14.07 ±0.56 134.80 ±10.60 15.58 ±0.13
GSN [7] 37.21 ±1.17 4.56 ±0.19 41.75 ±1.33 4.14 ±0.02 43.32 ±8.86 6.19 ±0.49 79.54 ±2.60 10.21 ±0.15
GAUDI 33.70 ±1.27 3.24 ±0.12 18.75 ±0.63 1.76 ±0.05 18.52 ±0.11 3.63 ±0.65 37.35 ±0.38 4.14 ±0.03
Table 2: Generative performance of state-of-the-art approaches for generative modelling of radiance
fields on 4 scene datasets: Vizdoom [
], Replica [
], VLN-CE [
] and ARKitScenes [
], according
to FID [14] and SwAV-FID [31] metrics.
In Fig. 6 we show samples from the unconditional distribution learnt by GAUDI for different datasets.
We observe that GAUDI is able to generate diverse and realistic 3D scenes from the empirical
distribution which can be rendered from the sampled camera poses.
4.5 Conditional Generative Modeling
In addition to modeling the distribution
, with GAUDI we can also tackle conditional gen-
erative problems
, where a conditioning variable
is given to modulate
. For
all conditioning variables
we assume the existence of paired data
{z, y}
to train the conditional
model [
]. In this section we show both quantitative and qualitative results for conditional
inference problems. The first conditioning variable we consider are textual descriptions of trajectories.
Second, we consider a conditional model where randomly sampled RGB images in a trajectory act as
conditioning variables. Finally, we use a categorical variable that indicates the 3D environment (i.e.
Figure 6: Different scenes sampled from unconditional GAUDI (one sample per row) and rendered
from their corresponding sampled camera poses (one dataset per column): Vizdoom [
], Replica
], VLN-CE [
]and ARKitScenes [
]. The resolutions are
64 ×64
64 ×64
128 ×128
64 ×64 respectively.
the particular indoor space) from which each trajectory was obtained (i.e. a one-hot vector). Tab. 3
shows quantitative results for the different conditional inference problems.
Text Conditioning Image Conditioning Categorical Conditioning
18.50 3.75 19.51 3.93 18.74 3.61
Avg. Per-Environment
50.79 4.10
Table 3: Quantitative results of Conditional Generative Modeling on VLN-CE [
] dataset. GAUDI
is able to produce high-quality scene renderings with low FID and SwAV-FID scores. In the right
table we show the difference in average per-environment FID score between the conditional and
unconditional models.
4.5.1 Text Conditioning
We tackle the challenging task of training a text conditional model for 3D scene generation. We
use the navigation text descriptions provided in VLN-CE [
] to condition our model. These text
descriptions contain high level information about the scene as well as the navigation path (i.e."Walk
out of the bedroom and into the living room","Exit the room through the swinging doors and
then enter the bedroom"). We employ a pre-trained RoBERTa-base [
] text encoder and use its
intermediate representation to condition the diffusion model. Fig. 7 shows qualitative results of
GAUDI for this task. To the best of our knowledge, this is the first model that allows for conditional
3D scene generation from text in an amortized manner (i.e. without distilling CLIP [
] through a
costly optimization problem [17, 28]).
Go through the hallway Go up the stairs Walk into the kitchen
Go down the stairs
Figure 7: Text conditional 3D scene generation using GAUDI (one sample per row). Our model is
able to capture the conditional distributions of scenes by generating multiple plausible scenes and
camera paths that match the given text prompts.
4.5.2 Image Conditioning
We now analyze whether GAUDI is able to pick up information from the RGB images to predict a
distribution over
. In this experiment we randomly pick images in a trajectory
and use it as a
conditioning variable
. For this experiment we use trajectories in the VLN-CE dataset [
]. During
each training iteration we sample a random image for each trajectory
and use it as a conditioning
variable. We employ a pre-trained ResNet-18 [
] as an image encoder. During inference, the
resulting conditional GAUDI model is able to sample radiance fields where the given image is
observed from a stochastic viewpoint. In Fig. 8 we show samples from the model conditioned on
different RGB images.
Figure 8: Image conditional 3D scene generation using GAUDI (one sample per row). Given a
conditioned image (top row), our model is able to sample scenes where the same or contextually
similar view is observed from a stochastic viewpoint.
Environment ID: 1 Environment ID: 2 Environment ID: 3
Figure 9: Samples from the GAUDI model conditioned on a categorical variable denoting the indoor
scene (one sample per row).
4.5.3 Categorical Conditioning
Finally, we analyze how GAUDI performs when conditioned on a categorical variable that indi-
cates the underlying 3D indoor environment in which each trajectory was recorded. We perform
experiments in the VLN-CE [
] dataset, where we employ a trainable embedding layer to learn a rep-
resentation for categorical variables indicating each environment. We compare the per-environment
FID score of conditional model with its unconditional counterpart. This per-enviroment FID score
is computed only on real images of the same indoor environment that the model is conditioned on.
Our hypothesis is that if the model efficiently captures the information in the conditioning variable
it should capture the environment specific distribution better than its unconditional counterpart
trained on the same data. In Tab. 3 the last column shows difference (e.g. the
) on the average
per-environment FID score between the conditional and unconditional model on VLN-CE dataset.
We observe that the conditional model consistently obtains a better FID score than the unconditional
model across all indoor environments, resulting in a sharp reduction of average FID and SwAV-FID
scores. In addition, in Fig. 9 we show samples from the model conditioned on a given categorical
5 Conclusion
We have introduced GAUDI, a generative model that captures distributions of complex and realistic
3D scenes. GAUDI uses a scalable two-stage approach which first involves learning a latent repre-
sentation that disentangles radiance fields and camera poses. The distribution of disentangled latent
representations is then modeled with a powerful prior. Our model obtains state-of-the-art performance
when compared with recent baselines across multiple 3D datasets and metrics. GAUDI can be used
both for conditional and unconditional problems, and enabling new tasks like generating 3D scenes
from text descriptions.
Gilad Baruch, Zhuoyuan Chen, Afshin Dehghan, Yuri Feigin, Peter Fu, Thomas Gebauer, Daniel Kurz,
Tal Dimry, Brandon Joffe, Arik Schwartz, et al. Arkitscenes: A diverse real-world dataset for 3d indoor
scene understanding using mobile rgb-d data. In Thirty-fifth Conference on Neural Information Processing
Systems Datasets and Benchmarks Track (Round 1), 2021.
Piotr Bojanowski, Armand Joulin, David Lopez-Paz, and Arthur Szlam. Optimizing the latent space of
generative networks. arXiv preprint arXiv:1707.05776, 2017.
[3] Marcus Carter and Ben Egliston. Ethical implications of emerging mixed reality technologies. 2020.
Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio
Gallo, Leonidas Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3d generative
adversarial networks. arXiv preprint arXiv:2112.07945, 2021.
Eric R Chan, Marco Monteiro, Petr Kellnhofer, Jiajun Wu, and Gordon Wetzstein. pi-GAN: Periodic
implicit generative adversarial networks for 3d-aware image synthesis. arXiv preprint arXiv:2012.00926,
Angel Chang, Angela Dai, Thomas Funkhouser, Maciej Halber, Matthias Niessner, Manolis Savva, Shuran
Song, Andy Zeng, and Yinda Zhang. Matterport3d: Learning from rgb-d data in indoor environments.
arXiv preprint arXiv:1709.06158, 2017.
Terrance DeVries, Miguel Angel Bautista, Nitish Srivastava, Graham W Taylor, and Joshua M Susskind.
Unconstrained scene generation with locally conditioned radiance fields. ICCV, 2021.
Emilien Dupont, Miguel Angel Bautista, Alex Colburn, Aditya Sankar, Carlos Guestrin, Josh Susskind,
and Qi Shan. Equivariant neural rendering. In International Conference on Machine Learning, pages
2761–2770. PMLR, 2020.
Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image
synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition,
pages 12873–12883, 2021.
Jiatao Gu, Lingjie Liu, Peng Wang, and Christian Theobalt. Stylenerf: A style-based 3d-aware generator
for high-resolution image synthesis. arXiv preprint arXiv:2110.08985, 2021.
Pengsheng Guo, Miguel Angel Bautista, Alex Colburn, Liang Yang, Daniel Ulbricht, Joshua M Susskind,
and Qi Shan. Fast and explicit neural view synthesis. In Proceedings of the IEEE/CVF Winter Conference
on Applications of Computer Vision, pages 3791–3800, 2022.
[12] David Ha and Jürgen Schmidhuber. World models. arXiv preprint arXiv:1803.10122, 2018.
Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition.
In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770–778, 2016.
Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans
trained by a two time-scale update rule converge to a local nash equilibrium. arXiv preprint
arXiv:1706.08500, 2017.
Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural
Information Processing Systems, 33:6840–6851, 2020.
Aapo Hyvärinen and Peter Dayan. Estimation of non-normalized statistical models by score matching.
Journal of Machine Learning Research, 6(4), 2005.
Ajay Jain, Ben Mildenhall, Jonathan T Barron, Pieter Abbeel, and Ben Poole. Zero-shot text-guided object
generation with dream fields. arXiv preprint arXiv:2112.01455, 2021.
Niharika Jain, Alberto Olmo, Sailik Sengupta, Lydia Manikonda, and Subbarao Kambhampati. Imperfect
imaganation: Implications of gans exacerbating biases on facial data augmentation and snapchat selfie
lenses. arXiv preprint arXiv:2001.09528, 2020.
Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial
networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition,
pages 4401–4410, 2019.
Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and
improving the image quality of stylegan. In Proceedings of the IEEE/CVF Conference on Computer Vision
and Pattern Recognition, pages 8110–8119, 2020.
Michał Kempka, Marek Wydmuch, Grzegorz Runc, Jakub Toczek, and Wojciech Ja´
skowski. Vizdoom:
A doom-based ai research platform for visual reinforcement learning. In 2016 IEEE Conference on
Computational Intelligence and Games (CIG), pages 1–8. IEEE, 2016.
Adam R Kosiorek, Heiko Strathmann, Daniel Zoran, Pol Moreno, Rosalia Schneider, Sona Mokrá, and
Danilo Jimenez Rezende. Nerf-vae: A geometry aware 3d scene generative model. In International
Conference on Machine Learning, pages 5742–5752. PMLR, 2021.
Jacob Krantz, Erik Wijmans, Arjun Majundar, Dhruv Batra, and Stefan Lee. Beyond the nav-graph:
Vision and language navigation in continuous environments. In European Conference on Computer Vision
(ECCV), 2020.
Zihang Lai, Sifei Liu, Alexei A Efros, and Xiaolong Wang. Video autoencoder: self-supervised disentan-
glement of static 3d structure and motion. In Proceedings of the IEEE/CVF International Conference on
Computer Vision, pages 9730–9740, 2021.
Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt. Neural sparse voxel fields.
arXiv preprint arXiv:2007.11571, 2020.
Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis,
Luke Zettlemoyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining approach. arXiv
preprint arXiv:1907.11692, 2019.
Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy
networks: Learning 3d reconstruction in function space. In Proceedings of the IEEE Conference on
Computer Vision and Pattern Recognition, pages 4460–4470, 2019.
Oscar Michel, Roi Bar-On, Richard Liu, Sagie Benaim, and Rana Hanocka. Text2mesh: Text-driven neural
stylization for meshes. arXiv preprint arXiv:2112.03221, 2021.
Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng.
Nerf: Representing scenes as neural radiance fields for view synthesis. arXiv preprint arXiv:2003.08934,
Yisroel Mirsky and Wenke Lee. The creation and detection of deepfakes: A survey. ACM Computing
Surveys (CSUR), 54(1):1–41, 2021.
Stanislav Morozov, Andrey Voynov, and Artem Babenko. On self-supervised image representations for
GAN evaluation. In International Conference on Learning Representations, 2021.
Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya
Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided
diffusion models. arXiv preprint arXiv:2112.10741, 2021.
Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In
International Conference on Machine Learning, pages 8162–8171. PMLR, 2021.
Michael Niemeyer and Andreas Geiger. Giraffe: Representing scenes as compositional generative neural
feature fields. arXiv preprint arXiv:2011.12100, 2020.
Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf:
Learning continuous signed distance functions for shape representation. In Proceedings of the IEEE/CVF
Conference on Computer Vision and Pattern Recognition, pages 165–174, 2019.
Despoina Paschalidou, Amlan Kar, Maria Shugrina, Karsten Kreis, Andreas Geiger, and Sanja Fidler.
Atiss: Autoregressive transformers for indoor scene synthesis. Advances in Neural Information Processing
Systems, 34, 2021.
Songyou Peng, Michael Niemeyer, Lars Mescheder, Marc Pollefeys, and Andreas Geiger. Convolutional
occupancy networks. arXiv preprint arXiv:2003.04618, 2, 2020.
Ethan Perez, Florian Strub, Harm De Vries, Vincent Dumoulin, and Aaron Courville. Film: Visual reason-
ing with a general conditioning layer. In Proceedings of the AAAI Conference on Artificial Intelligence,
volume 32, 2018.
Benjamin Planche, Xuejian Rong, Ziyan Wu, Srikrishna Karanam, Harald Kosch, YingLi Tian, Jan Ernst,
and Andreas Hutter. Incremental scene synthesis. In Advances in Neural Information Processing Systems,
pages 1668–1678, 2019.
Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish
Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from
natural language supervision. In International Conference on Machine Learning, pages 8748–8763. PMLR,
Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and
Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning,
pages 8821–8831. PMLR, 2021.
Ali Razavi, Aaron Van den Oord, and Oriol Vinyals. Generating diverse high-fidelity images with vq-vae-2.
Advances in neural information processing systems, 32, 2019.
Daniel Rebain, Mark Matthews, Kwang Moo Yi, Dmitry Lagun, and Andrea Tagliasacchi. Lolnerf: Learn
from one look. arXiv preprint arXiv:2111.09996, 2021.
Christian Reiser, Songyou Peng, Yiyi Liao, and Andreas Geiger. Kilonerf: Speeding up neural radiance
fields with thousands of tiny mlps. In Proceedings of the IEEE/CVF International Conference on Computer
Vision, pages 14335–14345, 2021.
Xuanchi Ren and Xiaolong Wang. Look outside the room: Synthesizing a consistent long-term 3d scene
video from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern
Recognition (CVPR), 2022.
Salah Rifai, Grégoire Mesnil, Pascal Vincent, Xavier Muller, Yoshua Bengio, Yann Dauphin, and Xavier
Glorot. Higher order contractive auto-encoder. In Joint European conference on machine learning and
knowledge discovery in databases, pages 645–660. Springer, 2011.
Daniel Ritchie, Kai Wang, and Yu-an Lin. Fast and flexible indoor scene synthesis via deep convolu-
tional generative models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern
Recognition, pages 6182–6190, 2019.
Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution
image synthesis with latent diffusion models. arXiv preprint arXiv:2112.10752, 2021.
Robin Rombach, Patrick Esser, and Björn Ommer. Geometry-free view synthesis: Transformers and no 3d
priors, 2021.
Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical
image segmentation. In International Conference on Medical image computing and computer-assisted
intervention, pages 234–241. Springer, 2015.
Negar Rostamzadeh, Emily Denton, and Linda Petrini. Ethics and creativity in computer vision. arXiv
preprint arXiv:2112.03111, 2021.
Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed
Ghasemipour, Burcu Karagol Ayan, S. Sara Mahdavi, Rapha Gontijo Lopes, Tim Salimans, Jonathan Ho,
David J Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language
understanding, 2022.
Mehdi S. M. Sajjadi, Henning Meyer, Etienne Pot, Urs Bergmann, Klaus Greff, Noha Radwan, Suhani
Vora, Mario Lucic, Daniel Duckworth, Alexey Dosovitskiy, Jakob Uszkoreit, Thomas Funkhouser, and
Andrea Tagliasacchi. Scene Representation Transformer: Geometry-Free Novel View Synthesis Through
Set-Latent Scene Representations. CVPR, 2022.
Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, and Xi Chen. Improved
techniques for training gans. NeurIPS’16, page 2234–2242, Red Hook, NY, USA, 2016. Curran Associates
Manolis Savva, Abhishek Kadian, Oleksandr Maksymets, Yili Zhao, Erik Wijmans, Bhavana Jain, Julian
Straub, Jia Liu, Vladlen Koltun, Jitendra Malik, et al. Habitat: A platform for embodied ai research. In
Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9339–9347, 2019.
Katja Schwarz, Yiyi Liao, Michael Niemeyer, and Andreas Geiger. GRAF: Generative radiance fields for
3d-aware image synthesis. arXiv preprint arXiv:2007.02442, 2020.
Edward J Smith and David Meger. Improved adversarial systems for 3d object generation and reconstruction.
In Conference on Robot Learning, pages 87–96. PMLR, 2017.
Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised
learning using nonequilibrium thermodynamics. In International Conference on Machine Learning, pages
2256–2265. PMLR, 2015.
Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint
arXiv:2010.02502, 2020.
Julian Straub, Thomas Whelan, Lingni Ma, Yufan Chen, Erik Wijmans, Simon Green, Jakob J Engel, Raul
Mur-Artal, Carl Ren, Shobhit Verma, et al. The replica dataset: A digital replica of indoor spaces. arXiv
preprint arXiv:1906.05797, 2019.
Hoang Thanh-Tung and Truyen Tran. Catastrophic forgetting and mode collapse in gans. In 2020
International Joint Conference on Neural Networks (IJCNN), pages 1–10. IEEE, 2020.
Patrick Tinsley, Adam Czajka, and Patrick Flynn. This face does not exist... but it might be yours! identity
leakage in generative models. In Proceedings of the IEEE/CVF Winter Conference on Applications of
Computer Vision, pages 1320–1328, 2021.
Arash Vahdat, Karsten Kreis, and Jan Kautz. Score-based generative modeling in latent space. Advances
in Neural Information Processing Systems, 34:11287–11302, 2021.
Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural
information processing systems, 30, 2017.
Xinpeng Wang, Chandan Yeshwanth, and Matthias Nießner. Sceneformer: Indoor scene generation with
transformers. In 2021 International Conference on 3D Vision (3DV), pages 106–115. IEEE, 2021.
Daniel Watson, William Chan, Jonathan Ho, and Mohammad Norouzi. Learning fast samplers for diffusion
models by differentiating through sample quality. In International Conference on Learning Representations,
Daniel Watson, Jonathan Ho, Mohammad Norouzi, and William Chan. Learning to efficiently sample from
diffusion probabilistic models. arXiv preprint arXiv:2106.03802, 2021.
Alex Yu, Sara Fridovich-Keil, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa.
Plenoxels: Radiance fields without neural networks. arXiv preprint arXiv:2112.05131, 2021.
Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or
few images. arXiv preprint arXiv:2012.02190, 2020.
Peng Zhou, Lingxi Xie, Bingbing Ni, and Qi Tian. Cips-3d: A 3d-aware generator of gans based on
conditionally-independent pixel synthesis. arXiv preprint arXiv:2110.09788, 2021.
A Limitations, Future Work and Societal Impact
Although GAUDI represents a step forward in generative models for 3D scenes, we would like to clearly discuss
the limitations. One current limitation of our model is the fact that inference is not real-time. The reason for this
is two fold: (i) sampling from the DDPM prior is slow even if it is amortized for the whole 3D scene. Techniques
for improving inference efficiency in DDPMs have been recently proposed [
] and can complement
GAUDI. (ii) Rendering from a radiance field is not as efficient as rendering other 3D structures like meshes.
Recent work have also tackled this problem [
] and could be applied to our approach. In addition, many
of the latest image generative models [
] use multiple stages of up-sampling through diffusion models
to render high-res images. These up-sample stages could be directly applied to GAUDI. In addition, one could
considering studying efficient encoders to replace the optimization process to find latents. While attempts have
been made at using transformers [
] for short trajectories (5-10 frames) it is unclear how to scale to thousands
of images per trajectory like the ones in [
]. Finally, the main limitation for a model like GAUDI to exhibit
improved generation and generalization abilities is the lack of massive-scale and open-domain 3D datasets. In
particular ones with other associated modalities like textual descriptions.
When considering societal impact of generative models a few aspects that need attention are the use generative
models for creating disingenuous data, e.g. "DeepFakes" [
], training data leakage and privacy [
], and
amplification of the biases present in training data [
]. One specific ethical consideration that applies to GAUDI
is the impact that a model which can easily create immersive 3D scenes can have on future generations and their
detachment of reality [
]. For an in-depth review of ethical considerations in generative modeling we refer the
reader to [51].
B Experimental Settings and Details
In this section we describe details about data and model hyper-parameters. For all experiments our latents
dimensions. In the first stage, when latents are optimized via Eq. 2,
gets reshaped to
feature map before feeding it to the scene decoder network. In the second stage, when training the
DDPM prior we reshape
latent and leverage the power of a UNet [
] denoising
For each dataset, trajectories have different length, physical scale, as well as near and far planes for rendering,
which we adjust accordingly in our model.
]: In Vizdoom, trajectories contains
steps on average. In each step the camera is allowed to
move forward
game units or rotate left or right by
degrees. We set the unit length of an element in the
tri-plane representation as
game units (meaning each latent code
represents a volume of space of
cubic game units). The near plane is at
game units and the far plane at
game units. We use the data
and splits provided by [7].
]: In Replica, all trajectories contain 100 steps. In each step, the camera can either rotate left or
right by
degrees or move forward
centimeters. We set the unit length of an element in the tri-plane
representation as
centimeters (meaning each latent code
represents a volume of space of
centimeters). The near plane is at
meters and the far plane at
meters. We use the data and splits provided
by [7].
]: in VLN-CE trajectories contain a variable number of steps between
, approximately.
In each step, the camera can either rotate left or right by
degrees or move forward
centimeters. We set the
unit length of an element in the tri-plane representation as
centimeters. The near plane is at
meters and
the far plane at 12 meters. We use the data and training splits provided by [23].
]: in ARKitScenes trajectories contain a number of steps around
on average. In these
trajectories the camera is able to move continuously in any direction and orientation. We set the unit length of an
element in the tri-plane representation as
centimeters. The near plane is at
meters and the far plane at
meters. We use the 3DOD split of data provided by [1]
C Decoder Architecture Design and Details
In this section we describe the decoder model in Fig. 2 in the main paper. The decoder network is composed of
3 modules: scene decoder,camera pose decoder and radiance field decoder.
scene decoder
network follows the architecture of the VQGAN decoder [
], parameterized with
convolutional architecture that contains a self-attention layers at the end of each block. The output
<latexit sha1_base64="rl8UgLe6+YP+XOsJk2KEgugJIWI=">AAACF3icbVDLSsNAFJ3UV62vqEs3wSK4ComIuhGKblxW6AvaECbTSTt08nDmRigxf+HGX3HjQhG3uvNvnLQpaOuBgXPOvZe593gxZxIs61srLS2vrK6V1ysbm1vbO/ruXktGiSC0SSIeiY6HJeUspE1gwGknFhQHHqdtb3Sd19v3VEgWhQ0Yx9QJ8CBkPiMYlOXqZm+IIe0FGIaenzayzJWX3Zm8U+phJkAJx9WrlmlNYCwSuyBVVKDu6l+9fkSSgIZAOJaya1sxOCkWwAinWaWXSBpjMsID2lU0xAGVTjq5KzOOlNM3/EioF4IxcX9PpDiQchx4qjNfUs7XcvO/WjcB/8JJWRgnQEMy/chPuAGRkYdk9JmgBPhYEUwEU7saZIgFJqCirKgQ7PmTF0nrxLTPzNPb02rtqoijjA7QITpGNjpHNXSD6qiJCHpEz+gVvWlP2ov2rn1MW0taMbOP/kD7/AGtBqGD</latexit>
<latexit sha1_base64="Eyd63NoWcXzHECFQXtVg1Nbm8e4=">AAAB/3icbVDLSsNAFJ3UV62vquDGTbAIrkoiRV0W3bisYB/QljKZ3rRDJ5MwcyPWmIW/4saFIm79DXf+jZO2C209MHA4517umeNFgmt0nG8rt7S8srqWXy9sbG5t7xR39xo6jBWDOgtFqFoe1SC4hDpyFNCKFNDAE9D0RleZ37wDpXkob3EcQTegA8l9zigaqVc86AQUh56fPKS9DsI9JlGoIe0VS07ZmcBeJO6MlMgMtV7xq9MPWRyARCao1m3XibCbUIWcCUgLnVhDRNmIDqBtqKQB6G4yyZ/ax0bp236ozJNoT9TfGwkNtB4HnpnM0up5LxP/89ox+hfdhMsoRpBsesiPhY2hnZVh97kChmJsCGWKm6w2G1JFGZrKCqYEd/7Li6RxWnbPypWbSql6OasjTw7JETkhLjknVXJNaqROGHkkz+SVvFlP1ov1bn1MR3PWbGef/IH1+QNiVJb/</latexit>
<latexit sha1_base64="ZgJJTUJklOpDQOWM8P89PPK7Ft4=">AAAB9HicbVBNSwMxEJ31s9avqkcvwSJ40LIrRT0WvXisYD9gu5Rsmm1Ds8maZAtl6e/w4kERr/4Yb/4b03YP2vpg4PHeDDPzwoQzbVz321lZXVvf2CxsFbd3dvf2SweHTS1TRWiDSC5VO8SaciZowzDDaTtRFMchp61weDf1WyOqNJPi0YwTGsS4L1jECDZWCjTqMIH8C+8ceUG3VHYr7gxomXg5KUOOerf01elJksZUGMKx1r7nJibIsDKMcDopdlJNE0yGuE99SwWOqQ6y2dETdGqVHoqksiUMmqm/JzIcaz2OQ9sZYzPQi95U/M/zUxPdBBkTSWqoIPNFUcqRkWiaAOoxRYnhY0swUczeisgAK0yMzaloQ/AWX14mzcuKd1WpPlTLtds8jgIcwwmcgQfXUIN7qEMDCDzBM7zCmzNyXpx352PeuuLkM0fwB87nD0lhkH4=</latexit>
<latexit sha1_base64="89J+JuC5FxMnvUsR7pwslvrJLdI=">AAAB6HicbVBNS8NAEJ3Ur1q/qh69LBbBU0mkqMeiF48t2FpoQ9lsJ+3azSbsboQS+gu8eFDEqz/Jm//GbZuDtj4YeLw3w8y8IBFcG9f9dgpr6xubW8Xt0s7u3v5B+fCoreNUMWyxWMSqE1CNgktsGW4EdhKFNAoEPgTj25n/8IRK81jem0mCfkSHkoecUWOlJuuXK27VnYOsEi8nFcjR6Je/eoOYpRFKwwTVuuu5ifEzqgxnAqelXqoxoWxMh9i1VNIItZ/ND52SM6sMSBgrW9KQufp7IqOR1pMosJ0RNSO97M3E/7xuasJrP+MySQ1KtlgUpoKYmMy+JgOukBkxsYQyxe2thI2ooszYbEo2BG/55VXSvqh6l9Vas1ap3+RxFOEETuEcPLiCOtxBA1rAAOEZXuHNeXRenHfnY9FacPKZY/gD5/MHyVOM8A==</latexit>
<latexit sha1_base64="ctajA+Ew79AFF9gq5hDForI4aIg=">AAAB6HicbVBNS8NAEJ3Ur1q/qh69LBbBU0lE1GPRi8cW7Ae0oWy2k3btZhN2N0IJ/QVePCji1Z/kzX/jts1BWx8MPN6bYWZekAiujet+O4W19Y3NreJ2aWd3b/+gfHjU0nGqGDZZLGLVCahGwSU2DTcCO4lCGgUC28H4bua3n1BpHssHM0nQj+hQ8pAzaqzUCPvlilt15yCrxMtJBXLU++Wv3iBmaYTSMEG17npuYvyMKsOZwGmpl2pMKBvTIXYtlTRC7WfzQ6fkzCoDEsbKljRkrv6eyGik9SQKbGdEzUgvezPxP6+bmvDGz7hMUoOSLRaFqSAmJrOvyYArZEZMLKFMcXsrYSOqKDM2m5INwVt+eZW0LqreVdVrXFZqt3kcRTiBUzgHD66hBvdQhyYwQHiGV3hzHp0X5935WLQWnHzmGP7A+fwBzOmM8A==</latexit>
<latexit sha1_base64="bzcJZNxuvyerPnnijwidinEzigo=">AAAB/3icbVBNS8NAEN34WetXVPDiZbEIHqQkIuqx6MVjBfsBTSib7aZdupuE3YlYYg7+FS8eFPHq3/Dmv3Hb5qCtDwYe780wMy9IBNfgON/WwuLS8spqaa28vrG5tW3v7DZ1nCrKGjQWsWoHRDPBI9YADoK1E8WIDARrBcPrsd+6Z0rzOLqDUcJ8SfoRDzklYKSuve9lnuZ9SU6wB+wBgjAjuZd37YpTdSbA88QtSAUVqHftL68X01SyCKggWndcJwE/Iwo4FSwve6lmCaFD0mcdQyMimfazyf05PjJKD4exMhUBnqi/JzIitR7JwHRKAgM9643F/7xOCuGln/EoSYFFdLooTAWGGI/DwD2uGAUxMoRQxc2tmA6IIhRMZGUTgjv78jxpnlbd86p7e1apXRVxlNABOkTHyEUXqIZuUB01EEWP6Bm9ojfryXqx3q2PaeuCVczsoT+wPn8AL+6WOA==</latexit>
<latexit sha1_base64="kB6OhSJt9NPMyJMATvw7es7jHZ0=">AAAB+XicbVDLSsNAFJ3UV62vqEs3g0VwVZJS1GXBjcsK9gFtCJPppB06eTBzUxtD/sSNC0Xc+ifu/BunbRbaeuDC4Zx7ufceLxZcgWV9G6WNza3tnfJuZW//4PDIPD7pqCiRlLVpJCLZ84higoesDRwE68WSkcATrOtNbud+d8qk4lH4AGnMnICMQu5zSkBLrmkOgM3A87PH3M1m6VPumlWrZi2A14ldkCoq0HLNr8EwoknAQqCCKNW3rRicjEjgVLC8MkgUiwmdkBHraxqSgCknW1ye4wutDLEfSV0h4IX6eyIjgVJp4OnOgMBYrXpz8T+vn4B/42Q8jBNgIV0u8hOBIcLzGPCQS0ZBpJoQKrm+FdMxkYSCDquiQ7BXX14nnXrNvqo17hvVZr2Io4zO0Dm6RDa6Rk10h1qojSiaomf0it6MzHgx3o2PZWvJKGZO0R8Ynz/IKZRe</latexit>
<latexit sha1_base64="lcRJa1ZSaskFUFSM9oVw6UIPCOQ=">AAACBHicbVC7TsMwFL3hWcqrwNjFokJiqhKogLGChbEg+pCaUDmu01p1nMh2kKooAwu/wsIAQqx8BBt/g9tmgJYjWTo+517de48fc6a0bX9bS8srq2vrhY3i5tb2zm5pb7+lokQS2iQRj2THx4pyJmhTM81pJ5YUhz6nbX90NfHbD1QqFok7PY6pF+KBYAEjWBupVyq7IdZDP0jjDLlMoNnXT2+z+9NeqWJX7SnQInFyUoEcjV7py+1HJAmp0IRjpbqOHWsvxVIzwmlWdBNFY0xGeEC7hgocUuWl0yMydGSUPgoiaZ7QaKr+7khxqNQ49E3lZEc1703E/7xuooMLL2UiTjQVZDYoSDjSEZokgvpMUqL52BBMJDO7IjLEEhNtciuaEJz5kxdJ66TqnFVrN7VK/TKPowBlOIRjcOAc6nANDWgCgUd4hld4s56sF+vd+piVLll5zwH8gfX5A5YGmBA=</latexit>
(a) (b)
Figure 10: (a) Architecture of the camera pose decoder network. (b) Architecture of the radiance
field network.
of the scene decoder is a feature map of shape
64 ×64 ×768
. To obtain the tri-plane representation
W= [Wxy,Wxz ,Wy z ]
we split the channel dimension of the output feature map in 3 chunks of
equal size 64 ×64 ×256.
camera pose decoder
is implemented as an MLP with
conditional batch normalization (CBN)
blocks with residual connections and hidden size of
, as in [
]. The conditional batch normaliza-
tion parameters are predicted from
. We apply positional encoding to the inputs the camera pose
encoder (s[1,1]). Fig. 10(a) shows the architecture of the camera pose decoder module.
radiance field decoder
is implemented as an MLP with
linear layers with hidden dimension
and LeakyReLU activations. We apply positional encoding to the inputs the radiance field
decoder (
) and concatenate the conditioning variable
to the output of every other layer
in the MLP starting from the input layer (e.g. layers 0, 2, 4, and 6). To improve efficiency, we render a
small resolution feature map of
channels (two times smaller than the output resolution) instead of
an RGB image and use a UNet [
] with additional deconvolution layers to predict the final image
[7, 34]. Fig. 10(b) shows the architecture of the radiance field decoder module.
For training we initialize all latents
z= 0
and train them jointly with the parameters of the 3 modules. We use
the Adam optimizer and a learning rate of
for latents and
for model parameters. We train our
model on 8 A100 NVIDIA GPUs for 2-7 days (depending on dataset size), with a batch size of
where we randomly sample 2images per trajectory.
D Prior Architecture Design and Details
We employ a Denoising Diffusion Probabilistic Model (DDPM) [
] to learn the distribution
. Specifically,
we adopt the UNet architecture from [
] to denoise the latent at each timestep. During training, we sample
t {1, ..., T }
uniformly and take the gradient descent step on
from Eq. 3. Different from [
], we keep the
original DDPM training scheme with fixed time-dependent covariance matrix and linear noise schedule. During
inference, we start from sampling latent from zero-mean unit-variance Gaussian distribution and perform the
denoising step iteratively. To accelerate the sampling efficiency, we leverage DDIM [
] to denoise only 50
steps by modeling the deterministic non-Markovian diffusion processes.
For conditional generative modelling tasks, the conditioning mechanism should be general to support conditioning
inputs from diverse modalities (i.e. text, image, categorical class, etc.). To fulfill this requirement, we first
project the conditional inputs into an embedding representations
via a modality-specific encoder. For text
conditioning, we employ a pre-trained RoBERTa-base [
]. For image conditioning, we employ a ResNet-18
] pre-trained on ImageNet. For categorical conditioning, we employ a trainable per-environment embedding
layer. We freeze the encoders for text and image inputs to avoid over-fitting issues. We borrow the cross attention
module from LDM [
] to fuse the conditioning representation
with the intermediate activations at multiple
levels in the UNet [
]. The cross-attention module implements an attention mechanism with key and value
generated from
while the query generated from the intermediate activations in the UNet architecture (we refer
readers to [48] for more details).
For training the DDPM prior, we use the Adam optimizer and learning rate of
. We train our model on 1
A100 NVIDIA GPU for 1-3 days for unconditional prior learning and 3-5 days for conditional prior learning
experiments (depending on dataset size), with a batch size of 256 and 32 respectively. For the hyper-parameters
of the DDPM model, we set the number diffusion steps to 1000, noise schedule as linearly decreasing from
0.0195 to 0.0015, base channel size to 224, attention resolutions at [8, 4, 2, 1], and number of attention heads to
E Ablation Study
We now provide additional ablations studies for the critical components in GAUDI. First, we analyze how the
dimensionality of the latent code
and the magnitude of
affect the optimization problem defined in Eq. 2.
Tab. 4 shows reconstruction metrics for both RGB images and camera poses for a subset of
trajectories in
the VLN-CE dataset [
]. We observe a clear trend where increasing the magnitude of
makes it harder to find
latent codes with high reconstruction accuracy. This drop in accuracy is expected since
controls the amount of
noise in latent codes during training. Finally, we observe that reconstruction performance starts to degrade when
the latent code dimensionality grows past 2048.
l1PSNR SSIM Rot Err. Trans. Err
β= 0.1zd= 2048 7.63e-3 39.12 0.984 4.61e-3 2.90e-3
β= 0.1zd= 4096 7.89e-3 38.55 0.982 4.91e-3 2.76e-3
β= 0.1zd= 8192 9.02e-3 36.33 0.978 5.62e-3 3.36e-3
β= 1.0zd= 2048 1.00e-2 34.82 0.972 6.32e-3 3.77e-3
β= 1.0zd= 4096 1.11e-2 34.46 0.965 7.27e-3 5.69e-3
β= 1.0zd= 8192 1.54e-2 32.28 0.916 1.11e-2 7.13e-3
β= 10.0zd= 8192 3.89e-2 24.89 0.799 7.59e-2 3.61e-2
β= 10.0zd= 4096 9.25e-2 17.52 0.499 1.56e-1 6.30e-2
β= 10.0zd= 2048 1.35e-1 12.74 0.275 5.25e-1 1.29e-1
Table 4: Ablation experiment for the critical parameters of the optimization process described in Eq.
In addition, we also provide ablation experiments for the second stage of our model where we learn the prior
. In particular, we ablate critical factors of our model: the importance of learning corresponding scene and
pose latents, the width of the denoising network in the DDPM prior, and the noise scale parameter
. In Tab.
5 we show results for each factor. In particular, in the first two rows of Tab. 5 we show the result of training
the prior while breaking the correspondence of
z= [zpose,zscene ]
. We break this correspondence by forming
random pairs of
z= [zpose,zscene ]
after optimizing the latent representations, and then training the prior on
these random pairs. We observe that training the prior to render scenes from a random pose latent impacts
both the FID and SwAV-FID scores substantially, which provides support for our claim that the distribution of
valid camera poses depends on the scene. In addition, we can see how the width of the denoising model affects
performance. By increasing the number of channels, the DDPM prior is able to better capture the distribution
of latents. Finally, we also show how different noise scales
impact the capacity of the generative model to
capture the distribution of scenes. All results Tab 5 are performed on the full VLN-CE dataset [23].
VLN-CE [23]
GAUDI 18.52 3.63
GAUDI w. Random Pose 83.66 10.73
Base Channel Size = 64 104.27 13.21
Base Channel Size = 128 22.04 4.35
Base Channel Size = 192 18.61 3.79
Base Channel Size = 224 18.52 3.63
Noise Scale β= 0.0 18.48 3.68
Noise Scale β= 0.1 (same as 1st stage) 18.52 3.63
Noise Scale β= 0.2 18.48 3.67
Noise Scale β= 0.5 20.20 4.11
Table 5: Ablation study for different design choice of GAUDI.
In Tab. 6 we report ablation results for modulation on the denoising architecture in the DDPM prior. We
compare cross-attention style conditioning as in LDM [48] with FiLM style conditioning [38]. For FiLM style
conditioning, we take the mean of the conditioning representation
across spatial dimension and project it into
the same space as denoising timestep embedding. After that, we take the sum of the conditioning and timestep
embedding and predict the scaling and shift factors of the affine transformation applied to the UNet intermediate
activations. We compare the performance of the two conditioning mechanisms in Tab. 6. We observe that the
cross-attention style conditioning performs better than the FiLM style across all our conditional generative
modeling experiments.
E.1 Additional Visualizations
In this section we provide additional visualizations for both figures in this appendix and videos that can be
found attached in the supplementary material. In Fig. 11 we provide additional interpolations between random
Text Conditioning Image Conditioning Categorical Conditioning
FiLM Module [38] 20.99 4.11 21.01 4.21 18.75 3.63
Cross Attention [48] 18.50 3.75 19.51 3.93 18.74 3.61
Table 6: Ablation study for conditioning mechanism of GAUDI.
pairs of latents obtained for VLN-CE dataset [
], where each row represents a interpolation path between
a random pair of latents (i.e. rightmost and leftmost columns). We can see how the model tends to produce
smoothly changing interpolation paths which align similar scene content. In addition we refer readers to the
in which videos of interpolations can be found where for each interpolated scene we
immersively navigate it by moving the camera forwards and rotating left and right.
In addition, we provide more visualization of samples from the unconditional GAUDI model in Fig. 12 for
], Fig. 13 for ARKitScenes [
] and Fig. 14 for Replica [
]. In all these figures, each row represents
a sample from the prior that is rendered from its corresponding sampled camera path. We note how these
qualitative results reinforce the fidelity and variability of the distribution captured by GAUDI, which is also
reflected in the quantitative results in Tab. 2 of the main paper. In addition, the folder
contains videos of more samples from the unconditional GAUDI model for all datasets.
Finally, the folder
contains a video showing samples from GAUDI conditioned on different
modalities like text, images or categorical variables. These visualizations corresponds to the results in Sect. 4.5
of the main paper.
F License
Due to licensing issues we cannot release the VLN-CE [
] raw trajectory data and we refer the reader to
and to the license of the Matterport3D data
Figure 11: Additional interpolation of 3D scenes in latent space for the VLN-CE dataset [
]. Each
row corresponds to a different interpolation path between random pairs of latent representations
Figure 12: Additional visualizations of scenes sampled from unconditional GAUDI for VLN-CE
dataset [23]. Each row to a scene rendered from camera poses sampled from the prior.
Figure 13: Additional visualizations of scenes sampled from unconditional GAUDI for ARKitScenes
dataset [1]. Each row corresponds to a scene rendered from camera poses sampled from the prior.
Figure 14: Additional visualizations of scenes sampled from unconditional GAUDI for Replica
dataset [60]. Each row corresponds to a scene rendered from camera poses sampled from the prior.
ResearchGate has not been able to resolve any citations for this publication.
In this paper, we show that popular Generative Adversarial Network (GAN) variants exacerbate biases along the axes of gender and skin tone in the generated data. The use of synthetic data generated by GANs is widely used for a variety of tasks ranging from data augmentation to stylizing images. While practitioners celebrate this method as an economical way to obtain synthetic data to train data-hungry machine learning models or provide new features to users of mobile applications, it is unclear whether they recognize the perils of such techniques when applied to real world datasets biased along latent dimensions. Although one expects GANs to replicate the distribution of the original data, in real-world settings with limited data and finite network capacity, GANs suffer from mode collapse. First, we show readily-accessible GAN variants such as DCGANs ‘imagine’ faces of synthetic engineering professors that have masculine facial features and fair skin tones. When using popular GAN architectures that attempt to address mode-collapse, we observe that these variants either provide a false sense of security or suffer from other inherent limitations due to their design choice. Second, we show that a conditional GAN variant transforms input images of female and nonwhite faces to have more masculine features and lighter skin when asked to generate faces of engineering professors. Worse yet, prevalent filters on Snapchat end up consistently lightening the skin tones in people of color when trying to make face images appear more feminine. Thus, our study is meant to serve as a cautionary tale for practitioners and educate them about the side-effect of bias amplification when applying GAN-based techniques.