acceptodds
Under review as a conference paper at ICLR 2027

Test-Time Compositional Generalization in Diffusion Models via Prototype Discovery

21 views

Abstract

Compositional generalization asks a model to produce what it has never seen by recombining what it knows. Diffusion methods for this problem typically assume that the reusable concepts are known in advance, provided through labels, prompts, or a learned concept library. We ask whether a pretrained diffusion model can instead discover its own reusable prototypes. A prototype is a mode of a noisy marginal . We perform gradient ascent on the learned score from a datum to recover a mode and fit a local Gaussian around it, and the noise level sets the level of abstraction. On pretrained CIFAR-10 and CelebA-HQ DDPMs, the prototypes recovered from different data are distinct at fine noise levels and become less separated as noise increases, while a datum's own prototype changes sharply across levels. A single datum therefore induces several distinct, reusable prototypes rather than one prototype under increasing blur. With these discovered prototypes, at test time we select those that explain a single out-of-distribution query and compose them into a product of experts whose closed-form score guides the composable diffusion sampler. On held-out compositions of ColorMNIST and CelebA, the samples generated from one query generalize to other instances of the unseen class and outperform the same sampler conditioned on trained concept tokens. A pretrained diffusion model thus already contains prototypes at many levels of abstraction, and they can be recovered and composed without supervision.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
compositional generalization, OOD generalization, prototype discovery, diffusion models

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_uMCMOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

Compositional generalization tasks may require models to reuse a learned concept library. Do pretrained diffusion models (DM) internally acquire such a concept library? This work tackles the problem with gradient-ascent-based prototype discovery. The prototypes are collected by estimating the local argmax of the implicit density function of the generative distribution, which is modeled by Gaussian mixture models per time step. The discovered prototypes are leveraged to compose score functions in the style of a weighted product of experts for better out-of-distribution generalization. Experiments on the compositional settings in Color-MNIST, CelebA, and CIFAR-10 reveal that the proposed method achieved higher OOD-sample generation quality than the baselines, including editing-based SDEdit.

Critical strengths
  • The idea of prototype composition is natural and attracting to explain the extraordinaly OOD generalizability of modern diffusion models.
  • The experiments are small but well designed to observe compositional generalization behavior of the models
Critical weaknesses
  • The limited experimental scale and range: Although the small controlled experiments are good proof of concept, I would like to see whether larger scale, product-ready diffusion models also show similar prototyping behavior.

  • Why the discovered mode peaks can be called "prototypes" is not deeply discussed. While it is possible that they are internal prototypes, it is also possible that they are simply cluster centers of natural images. The results that the number of prototypes differs is partially insightful, but the number of clusters is also externally controllable in existing clustering algorithms, and I do not think that this is conclusive.

Questions for authors
  • CIFAR and CelebA include almost the same size objects. Is the proposed method expected to work for multi-scale objects? It is questionable how the prototype represents difference scales of objects.

  • Minor comments:

    • Eqn. 5: "arg local max" is understandable but a bit strange. I prefer argmax_{x \in local}
1 comment
AuthorsOpenReview ↗

We thank the reviewer for finding the idea of prototype composition natural and the experiments well designed.

“The limited experimental scale and range: Although the small controlled experiments are good proof of concept, I would like to see whether larger scale, product-ready diffusion models also show similar prototyping behavior.”

They do. Section 4 of the revised version runs its analysis on four pretrained models. Next to the two DDPMs it has DiT-XL/2 (Peebles & Xie, 2023), a 0.68B-parameter model trained on ImageNet at 256×256256\times256, and the text-to-image latent diffusion model of Rombach et al. (2022), a 1.45B-parameter model trained on LAION-400M and the architecture that Stable Diffusion is built on. For each of the two we take 256 of its own samples, from 16 classes or 16 prompts. We report the number of distinct prototypes that the images reach at each noise level tt, starting from the number of images (Figure 2 and Table 2 in the revised version, more examples in Figure 5):

ModelParametersDistinct prototypes as tt grows
CIFAR-10 DDPM36M512 → 139 (t=300t=300) → 14 (t=380t=380) → 1 (t=460t=460)
CelebA-HQ DDPM114M48 → 21 (t=260t=260) → 4 (t=380t=380) → 2 (t=460t=460)
DiT-XL/20.68B256 → 131 (t=400t=400) → 37 (t=460t=460) → 1 (t=580t=580)
Text-to-image LDM1.45B256 → 125 (t=500t=500) → 21 (t=660t=660) → 6 (t=740t=740) → 1 (t=860t=860)

Both large models show the prototyping behavior of Section 4. Every image has its own prototype at low noise, and more and more images end at the same prototype as the noise grows, until a single one remains. Images with the same content share a prototype first. On DiT-XL/2, 44% of the pairs of images from the same ImageNet class share a prototype at t=440t=440, against 21% of the pairs from different groups of classes (for example a dog and a car). On the text-to-image model, 8.7% of the pairs from the same prompt share a prototype at t=580t=580, against 4.2% of the pairs from unrelated prompts (Appendix B.2 and Figure 6 in the revised version).

“Why the discovered mode peaks can be called "prototypes" is not deeply discussed. While it is possible that they are internal prototypes, it is also possible that they are simply cluster centers of natural images. The results that the number of prototypes differs is partially insightful, but the number of clusters is also externally controllable in existing clustering algorithms, and I do not think that this is conclusive.”

A cluster center is a property of the images, and a prototype is a property of the model. Our ColorMNIST model is trained on 12 of the 16 combinations of digit color and background color. We ran mode ascent from digits in a combination that was never in training and from the same digits recolored to a combination that was in training, with the same background, so the two images differ only in the digit color. At t=300t=300 the prototype still has the colors of its image in 89.8% of the runs for the combination seen in training and in 9.7% for the unseen one (40 digits, 32 runs each, Table 3 and Figure 7 in the revised version). What a prototype keeps depends on what the model was trained on. A shared prototype is also not the center of the images that share it. On DiT-XL/2 it lies 1.2 to 1.8 times as far from their mean as the images themselves. Mode ascent has no number of clusters and no step size, so the number of prototypes at each noise level is measured, not set. The prototypes are also reusable. Eight of them, taken only from the model's own samples, compose a ColorMNIST class the model never saw with 92.5% attribute correctness, against 37.7% for Comp. 1 (Table 11 and Figure 12 in the revised version).

“CIFAR and CelebA include almost the same size objects. Is the proposed method expected to work for multi-scale objects? It is questionable how the prototype represents difference scales of objects.”

Yes. A prototype is a full image at the resolution of the model, so it carries an object at whatever scale the model generates it. The ImageNet classes and the prompts in the experiment above cover objects of different sizes in the frame, from close-ups of fruit to dogs, birds and cars, and the behavior of Section 4 holds on them.

“Eqn. 5: "arg local max" is understandable but a bit strange. I prefer argmax_{x \in local}”

We changed Eq. (5) to xt∗∈arg⁡max⁡xt∈Upt(xt)x_t^{*} \in \arg\max_{x_t \in U} p_t(x_t) for an open neighborhood UU of xt∗x_t^{*}.

Since the new results only strengthen our arguments, we will update them in the final version.

Peebles & Xie. Scalable Diffusion Models with Transformers. ICCV 2023. Rombach et al. High-Resolution Image Synthesis with Latent Diffusion Models. CVPR 2022.

Reviewer_cc9wOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper focuses on compositional generalisation in the image domain. The proposed approach uses a pretained diffusion model to identify concepts, which are defined as fix points (modes) resulting from an iterative application of the denoising model to a noisy image but holding the noise level (time scale) of the denoiser fixed. The identified modes represent prototypes that are used by a product-of-experts model to compose a new image, where each prototype serves as an expert. This whole approach is tested by its ability to generalise out-of-distribution.

Critical strengths

Decomposing images into fundamental concepts that can be recombined to generalise out-of-distribution is an important problem. The idea to use of a diffusion model has the advantage that coarse grained views on images are strongly correlated with noise scale. The method seems to work well on simpler data sets.

Critical weaknesses
  1. The paper lacks an overview Fig. to illustrate the proposed pipeline and a Fig. that shows how prototype discovery evolves over iterations. That "a random draw from p_t is visibly noisy" should be explicitly shown, as well as the corresponding result after gradient ascent.
  2. Denoising a noisy query image without updating the noise scale argument of denoiser means that denoising is done more coarse grained for higher noise levels and more fine grained for lower noise levels. The result for fashion MNIST can be seen in the supplement, and shows that the effect is simply getting clean images at different coarse graining levels but not the indented decomposition. As such I doubt that this approach is a strong, broadly applicable approach for compositional generalisation.
  3. Some introduced hyperparameters lack how their value is determined and the quantification of the effect when they are tuned. For example, how are the weights determined for the product-experts-model? What happens to out-of-distribution generalisation if we increase or decrease the number of selected modes for the investigated data sets?
Questions for authors
  1. How does the approach work on larger/more complex datasets?
  2. Can orthogonality of concepts/prototypes be enforced?
  3. Why should diffusion models be suitable for extracting compositional information besides coarse grained information?
2 comments
AuthorsOpenReview ↗

We thank the reviewer for finding the problem important and the use of the diffusion noise scale well motivated.

“The paper lacks an overview Fig. to illustrate the proposed pipeline and a Fig. that shows how prototype discovery evolves over iterations. That "a random draw from p_t is visibly noisy" should be explicitly shown, as well as the corresponding result after gradient ascent.”

We added both. Figure 10 (Appendix C) in the revised version shows the pipeline on one query, from the noised copies to the selected prototypes, their weights, the product of experts (PoE) and the samples. Figure 8 (Appendix B.3) shows the noisy draw from ptp_t next to the ascent steps and the prototype they reach, at four noise levels.

“Denoising a noisy query image without updating the noise scale argument of denoiser means that denoising is done more coarse grained for higher noise levels and more fine grained for lower noise levels. … the effect is simply getting clean images at different coarse graining levels but not the indented decomposition...”

A prototype is not a coarse-grained copy of the image. A coarse-grained copy of a green digit on a deep red background is still green on deep red. Our ColorMNIST model is trained on 12 of the 16 combinations of digit color and background color. We took digits in a combination that was never in training (such as green on deep red) and recolored each one to combinations that were in training, one with the same background (yellow on deep red) and one with the same digit color, so the images differ only in color. We ran mode ascent from all of them at the same noise level, and report how often the prototype still has the colors of the image it started from (%, 40 digits, 32 runs each, Table 3 and Figure 7 in the revised version):

Colors of the image the ascent starts fromt=100t=100t=200t=200t=300t=300t=400t=400
Never together in training60.022.39.79.8
In training, background color kept100.092.589.885.7
In training, digit color kept94.890.092.596.8

At the same noise level, the prototype keeps colors that the model has seen together and replaces colors that it has not. What a prototype keeps is set by what the model has learned, not by the amount of smoothing. Fashion-MNIST images are in-distribution for their model, so mode ascent keeps what they show, as it keeps the color combinations seen in training here.

The decomposition can be measured. We removed every ascent that starts at the query, so that all prototypes come from the model's own unconditional samples. We report attribute correctness against the number of selected prototypes KK, and the share of queries for which every factor of the query appears in at least one of its KK prototypes. The factors are the digit, its color and the background color on ColorMNIST and the 11 attributes on CelebA (Table 11 and Figure 12 in the revised version):

KK12358Real
ColorMNIST, attribute correctness (%)51.288.791.492.392.593.3
ColorMNIST, all three factors appear among the KK prototypes (%)2.748.269.284.290.5
CelebA, attribute correctness (%)59.669.574.276.377.178.3
CelebA, all 11 attributes appear among the KK prototypes (%)2.522.745.073.391.7

A single prototype has only a part of the query's factors, and eight together have all of them for more than 90% of the queries. At least 95% of the ColorMNIST prototypes show a class seen in training, so the unseen class is composed from prototypes of seen classes.

“Some introduced hyperparameters lack how their value is determined and the quantification of the effect when they are tuned. For example, how are the weights determined for the product-experts-model? What happens to out-of-distribution generalisation if we increase or decrease the number of selected modes for the investigated data sets?”

The weights are not tuned. Eq. (12) sets them from the likelihood that each selected prototype assigns to the query at each dimension, with one temperature τ\tau. On ColorMNIST with K=5K=5, τ=1.0/0.5/0.2/0.05\tau = 1.0 / 0.5 / 0.2 / 0.05 gives 87.7 / 91.2 / 92.3 / 92.4% attribute correctness. For the number of selected prototypes, we re-ran the experiments for K=1,2,3,5,8K = 1, 2, 3, 5, 8 and 1212 on both datasets (Appendix I in the revised version). The table above shows that attribute correctness rises with KK and saturates between K=5K=5 and K=8K=8 on both datasets, so Table 1 in the revised version uses K=8K=8 and τ=0.05\tau=0.05. Mode ascent has no step size to tune. The step of Eq. (6) is η=1−αˉt\eta=1-\bar\alpha_t, which moves the iterate to the rescaled posterior mean of Eq. (7), and it is halved whenever the iterate jumps across the mode (Section 3.2, Appendix F.1). The sampler needs no guidance schedule, and the effect of the scale SS of the PoE variance is shown in Figure 11.

“How does the approach work on larger/more complex datasets?”

The prototypes behave the same way on models trained on ImageNet and on LAION-400M. Section 4 of the revised version runs its analysis on four pretrained models. Next to the two DDPMs it has DiT-XL/2 (Peebles & Xie, 2023), a 0.68B-parameter model trained on ImageNet, and the text-to-image latent diffusion model of Rombach et al. (2022), a 1.45B-parameter model trained on LAION-400M. We report the number of distinct prototypes that a set of images reaches at each noise level tt, starting from the number of images (Figure 2 and Table 2 in the revised version):

ModelParametersDistinct prototypes as tt grows
CIFAR-10 DDPM36M512 → 139 (t=300t=300) → 14 (t=380t=380) → 1 (t=460t=460)
CelebA-HQ DDPM114M48 → 21 (t=260t=260) → 4 (t=380t=380) → 2 (t=460t=460)
DiT-XL/20.68B256 → 131 (t=400t=400) → 37 (t=460t=460) → 1 (t=580t=580)
Text-to-image LDM1.45B256 → 125 (t=500t=500) → 21 (t=660t=660) → 6 (t=740t=740) → 1 (t=860t=860)

On all four models, every image has its own prototype at low noise, and more and more images end at the same prototype as the noise grows. Images with the same content share a prototype first (see our response to Reviewer uMCM). For composition, on CelebA at 128×128128\times128 our method reaches 77.0% attribute correctness on the 23 held-out classes, against 63.4% for Comp. 1 and 78.3% for real members of the class (Tables 1 and 9 in the revised version).

“Can orthogonality of concepts/prototypes be enforced?”

The selection already prefers prototypes that complement each other. The greedy step of Eqs. (10) and (11) adds the prototype that best explains the dimensions of the query that the selected ones explain poorly. On ColorMNIST, one prototype has 1.5 of the query's three factors on average, three together have 2.7 and eight have 2.9.

“Why should diffusion models be suitable for extracting compositional information besides coarse grained information?”

Because the score is learned from the training data, mode ascent moves an image toward what the model has learned. It keeps a color combination that was in training in 89.8% of the runs and one that was not in 9.7% (table above, t=300t=300). The same holds on the 1.45B text-to-image model. For six objects generated in their usual color and in an unusual one (for example a yellow and a blue banana), 10 ascent steps at t=400t=400 keep the usual color in 95% of the images and the unusual one in 81% (Appendix B.3 in the revised version).

Since the new results only strengthen our arguments, we will update them in the final version.

Peebles & Xie. Scalable Diffusion Models with Transformers. ICCV 2023. Rombach et al. High-Resolution Image Synthesis with Latent Diffusion Models. CVPR 2022.

Reviewer_GbTVOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper shows that you can discover prototypes of the data by using a pretrained diffusion model and a single datum by score ascent towards the nearest data mode.

The noise level t applied to the datum initially naturally sets the level of abstraction which is partially justified by data.

Furthermore the authors show that this can be used for compositional generalization, by identifying prototypes from a single ood data.

Critical strengths
  • The presentation of the paper is relatively clear.

  • The prototype discovery method and varying the abstraction level with noise is interesting.

  • The compositional generalization results are sound.

Critical weaknesses

CW1: The general applicability or even scientific insight is unclear. i.e. It isn't clear whether the paper is trying to sell a practical method or establish a scientific technique. CW2: The compositional generalization is one-shot instead of pure generalization CW3: The SDEdit result seems under tuned in terms of hyper parameters.

Questions for authors

Q1: Are the SDEdit results hyperparam tuned?

1 comment
AuthorsOpenReview ↗

We thank the reviewer for finding the prototype discovery method interesting and the compositional generalization results sound.

“CW1: The general applicability or even scientific insight is unclear. i.e. It isn't clear whether the paper is trying to sell a practical method or establish a scientific technique.”

The paper is a scientific inquiry. It asks one question, “whether a pretrained diffusion model can instead discover its own reusable prototypes” (abstract), and answers it in three steps:

  1. Section 3 defines a prototype as a mode of a noisy marginal and recovers it by mode ascent on the learned score, with no labels, no concept library and no training.
  2. Section 4 analyzes the recovered prototypes: they become less separated as the noise grows, and one datum yields distinct prototypes at different noise levels.
  3. Section 5 tests what this analysis implies: “We next test whether the prototypes recovered from one out-of-distribution query support generalization to other instances of its unseen class” (end of Section 4). Compositional generalization is how we test the finding.

The revised version adds analyses that strengthen each step. Section 4 now runs on four pretrained models from 36M to 1.45B parameters, in pixel space and in latent space, with U-Net and transformer backbones, with class and with text conditioning. Appendix B adds analyses of modes as prototypes, and Appendix I adds analyses of the composition.

General applicability: the same behavior holds on every model, from 36M to 1.45B parameters.

Scientific insight: what a prototype keeps is what the model has learned, and the learned prototypes compose a class the model never saw.

“CW2: The compositional generalization is one-shot instead of pure generalization”

The samples are not built from the query. The query only tells the method which unseen class to generate, as a prompt would, since the backbone has no token for a held-out class. We removed every ascent that starts at the query, so that all prototypes come from the model's own unconditional samples and the query only selects and weights them. We report attribute correctness (%) on the held-out classes (Appendix I and Table 11 in the revised version):

MethodColorMNISTCelebA
Comp. 137.763.4
Comp. 339.362.2
Ours92.577.0
Ours, no ascent from the query92.577.1
Real93.378.3

The result is unchanged without any ascent from the query. On ColorMNIST 96% of these prototypes show a class seen in training and fewer than 1% show the query's own class, so the unseen class is composed from prototypes of seen classes. SDEdit is the one-shot method in Table 1, since it starts from the query's pixels.

“CW3: The SDEdit result seems under tuned in terms of hyper parameters.” “Q1: Are the SDEdit results hyperparam tuned?”

We swept SDEdit over t0t_0 on both datasets, on the same queries as our method. We report attribute correctness (Attr., %), KID per query batch (lower is better) and diversity (Div., higher is better), averaged over the held-out classes (Table 1 and Appendix H in the revised version):

MethodColorMNIST Attr.KIDDiv.CelebA Attr.KIDDiv.
SDEdit, t0=50t_0=5092.90.107−3.4078.30.218−4.21
SDEdit, t0=100t_0=10092.80.094−2.6278.10.197−3.57
SDEdit, t0=150t_0=15092.70.089−2.3777.70.189−3.39
SDEdit, t0=200t_0=20092.30.082−2.1777.10.179−3.21
SDEdit, t0=250t_0=25091.80.078−2.1276.70.174−3.10
SDEdit, t0=300t_0=30089.90.070−2.2876.00.165−2.96
Ours92.50.079−2.1277.00.166−3.13

At equal attribute correctness, our method outperforms SDEdit in KID and in diversity on both datasets. On CelebA, SDEdit at t0=200t_0=200 has our attribute correctness (77.1% vs. 77.0%), and our KID is lower in all 23 held-out classes (ours minus SDEdit: −0.013±0.002-0.013_{\pm 0.002}, mean and standard error over the classes) and our diversity higher in 21 of them. On ColorMNIST, our attribute correctness lies between SDEdit at t0=150t_0=150 and t0=200t_0=200, and our KID and diversity are better than both (ours minus SDEdit in KID: −0.010±0.002-0.010_{\pm 0.002} and −0.004±0.002-0.004_{\pm 0.002} over the 40 held-out classes). A larger t0t_0 lowers SDEdit's KID only together with its attribute correctness.

Since the new results only strengthen our arguments, we will update them in the final version.

Peebles & Xie. Scalable Diffusion Models with Transformers. ICCV 2023. Rombach et al. High-Resolution Image Synthesis with Latent Diffusion Models. CVPR 2022.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General response

We thank all reviewers. They found the problem important (cc9w), prototype discovery across noise levels interesting and well motivated (GbTV, cc9w), prototype composition natural and the experiments well designed (uMCM), and the compositional generalization results sound (GbTV).

The paper is a scientific inquiry. It asks “whether a pretrained diffusion model can instead discover its own reusable prototypes” (abstract), and answers in three steps:

  1. Section 3 defines a prototype as a mode of a noisy marginal pt(xt)p_t(x_t) and recovers it by mode ascent on the learned score, with no labels, no concept library and no training.
  2. Section 4 analyzes the recovered prototypes. They become less separated as the noise grows, and one datum yields distinct prototypes at different noise levels.
  3. Section 5 tests what this analysis implies: “We next test whether the prototypes recovered from one out-of-distribution query support generalization to other instances of its unseen class” (end of Section 4). Compositional generalization is how we test the finding.

During the rebuttal, we added the following analyses to the revised version, which further strengthen each step.

1. The finding holds across diffusion models (uMCM, cc9w, GbTV. Section 4, Figures 1 and 2, Appendix B.1). Section 4 now runs on four pretrained models from 36M to 1.45B parameters: the CIFAR-10 and CelebA-HQ DDPMs in pixel space, DiT-XL/2 (0.68B), a class-conditional transformer trained on ImageNet in latent space, and the text-to-image latent diffusion model of Rombach et al. (2022) (1.45B), trained on LAION-400M. On all four, every image has its own prototype at low noise, and more and more images end at the same prototype as the noise grows, until one or two remain. The finding depends only on the mathematics of diffusion, in which ptp_t is the data distribution smoothed by a Gaussian kernel that widens with tt (Section 3.1), so it applies across diffusion model architectures, scales, data and conditionings.

2. The modes are prototypes of the model (uMCM, cc9w, GbTV. Appendices B.2, B.3 and I). Reviewer uMCM asks “why the discovered mode peaks can be called "prototypes"” and whether they are “simply cluster centers of natural images”, and Reviewer cc9w asks whether the effect is “simply getting clean images at different coarse graining levels”. We call a mode a prototype because what it keeps depends on what the model has learned, because images with the same content share it, and because it can be reused to compose a class the model never saw. The revised version tests each of the three:

  • A prototype keeps what the model has learned. Our ColorMNIST model is trained on 12 of the 16 combinations of digit color and background color. We ran mode ascent from digits in a combination that was never in training and from the same digits recolored to a combination that was in training, so the two images differ only in color. At the same noise level (t=300t=300), where both are smoothed by the same amount, the prototype still has the colors of its image in 89.8% of the runs for the combination seen in training and in 9.7% for the unseen one. The 1.45B text-to-image model behaves in the same way (Appendix B.3, Table 3, Figures 7 to 9).
  • Images with the same content share a prototype first. On DiT-XL/2 at t=440t=440, 44% of the image pairs from the same class share a prototype, against 21% of the pairs from different groups of classes. The shared prototype is not the center of these images, since it lies 1.2 to 1.8 times as far from their mean as the images themselves (Appendix B.2, Figure 6).
  • The prototypes are reusable. We took every prototype from images that the model generates itself, so that none is found from the query. Eight of them compose the query's unseen class with 92.5% attribute correctness on ColorMNIST and 77.1% on CelebA, against at most 39.3% and 63.4% for composable diffusion (Appendix I, Table 11, Figure 12).

3. Further analyses and what each shows for the inquiry:

AnalysisWhat it showsAsked byIn the revised version
SDEdit over its noise level t0t_0Composing discovered prototypes outperforms editing the query. At equal attribute correctness, our KID is lower and our diversity higher on both datasets.GbTVSection 5.4, Appendix H, Table 10, Figure 11
Number of selected prototypes KK and temperature τ\tauComposition needs several prototypes that complement each other. Attribute correctness saturates between K=5K=5 and K=8K=8. One prototype has 1.5 of the query's three ColorMNIST factors, and eight together have 2.9.cc9wAppendix I, Table 11, Figure 12
Overview figure and mode ascent step by stepThey show the noisy draw from ptp_t, the ascent steps and the prototype they reach.cc9wFigures 8 and 10

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.