JEDI: The Force of Jensen-Shannon Divergence in Disentangling Diffusion Models

Enis Simsar; Eric Tillmann Bill; Thomas Hofmann

REVIEW 4 major objections 5 minor 20 references

JEDI: The Force of Jensen-Shannon Divergence in Disentangling Diffusion Models

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read JEDI claims that minimizing a Jensen-Shannon divergence over internal attention maps at inference time separates subjects in diffusion model images better than the cosine-based CONFORM baseline, with fewer updates and no retraining.

desk verdict A plausible JSD-based test-time attention disentanglement method whose central quantitative claim currently rests on a circular metric and post hoc block selection. read the letter →

arxiv 2505.19166 v2 pith:FG4UPCGT submitted 2025-05-25 cs.CV cs.LG

Eric Tillmann Bill , Enis Simsar , Thomas Hofmann This is my paper

classification cs.CVcs.LG

keywords test-timeadaptationdiffusionmodelstext-to-imagegenerationattentionmapsJensen-ShannondivergencesubjectdisentanglementCLIP-freeevaluation

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

JEDI is a test-time adaptation method that tries to stop text-to-image diffusion models from blending two subjects into one. It reads the model's own attention maps as probability distributions and tweaks the latent image so that each prompt token's attention is concentrated, separated from other tokens, and not collapsed. The paper claims this Jensen-Shannon divergence objective improves subject separation and prompt alignment on Stable Diffusion 1.5 and 3.5 and on a multi-concept LoRA model, without retraining or external supervision. If that holds, JEDI gives a model-agnostic fix for attribute mixing that uses about 67 percent fewer latent updates than the cosine-based CONFORM baseline, and it doubles as a CLIP-free score for how disentangled an image is.

What carries the argument

The load-bearing object is the normalized Jensen-Shannon divergence over sets of attention maps, treated as discrete probability distributions: $D_{JS}(P) = \frac{1}{|P|} \sum_{p \in P} D_{KL}(p \| m)$ with $m$ the mean distribution, divided by $\log n$ so it lies in $[0,1]$. The paper combines three terms built from this object: intra-group coherence, which minimizes JSD within a subject's attention maps; inter-group separation, which maximizes JSD between subjects' mixture distributions; and Shannon entropy regularization, which keeps the maps spatially spread. Updates use the Fast Gradient Sign Method, $x_t \leftarrow x_t - \alpha \cdot \operatorname{sign}(\nabla_{x_t} \text{score}(A_t))$, keeping the latent close to the base model. The mechanism's work is to turn the qualitative failure of attribute mixing into a quantitative overlap penalty on the model's own spatial probability fields.

What would settle it

Run JEDI with the same objective on attention maps taken from a deliberately wrong source, such as blocks outside 5 to 15 or unsymmetrized attention rows, and compare the generated images: if the inter-group JSD still rises without visual disentanglement, the attention-to-subject link is not carrying the result. More directly, collect generated images judged entangled versus disentangled by human raters across many prompts and check that the paper's inter-group JSD score separates the two sets; if it does not, the proposed CLIP-free score is not measuring what the paper claims.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that semantic entanglement can be formulated as overlap between attention distributions and minimized directly at inference time. For each subject, JEDI collects its token attention maps, minimizes their within-group Jensen-Shannon divergence, maximizes the divergence between subjects' mixture distributions, and adds a small entropy regularizer to prevent degeneracy. The loss is backpropagated into the latent with sign gradients, so only the first 18 denoising steps are adjusted. In Stable Diffusion 3.5, where self-attention mixes image and text tokens in DiT blocks, the paper extracts a symmetrized attention map per prompt token, bypassing softmax during attention and applying it only at extraction time, and optimizes blocks 5 to 15; in SD1.5 and LoRACLR it uses cross-attention maps. The reported outcome is comparable or better subject separation than the CONFORM baseline with 18 updates versus 69 and a smaller learning rate, plus a disentanglement score from inter-group JSD that rises when attribute mixing drops.

Load-bearing premise

The paper's load-bearing premise is that the attention maps it optimizes, in particular the symmetrized self-attention expression over DiT blocks 5 to 15 in Stable Diffusion 3.5, faithfully show which image regions each prompt subject controls; if those maps do not track subject location, minimizing divergence on them will not transfer to real image disentanglement.

Editorial extensions

If this is right

Any diffusion model that exposes attention maps can be retrofitted with JEDI at test time, with no retraining, no external models, and no hand-made layouts.
JEDI uses only 18 latent updates in the first 18 timesteps, roughly 67 percent fewer than CONFORM's 69, and its smaller learning rate keeps generated images closer to the base model's distribution.
The inter-group JSD value doubles as a lightweight, CLIP-free disentanglement score, with higher values tracking visually disentangled subjects during sampling.
The same method works on U-Net style SD1.5, DiT-style SD3.5, and the multi-concept LoRACLR model, indicating the objective is architecture-agnostic.
The objective can in principle be moved into training or fine-tuning, which the paper identifies as a route to removing the inference-time overhead.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

Beyond the paper: shifting the JEDI objective into training or fine-tuning, rather than test-time optimization, would directly penalize entangled attention maps during learning and remove the doubling of inference time, though the paper only sketches this direction.
Beyond the paper: because the method needs only internal attention maps, it should transfer to other transformer-based generators and even joint-attention video models; this is a prediction, not a paper claim.
Beyond the paper: the inter-group JSD score could serve as a cheap diagnostic for attribute mixing across any text-to-image pipeline, testable by comparing it against human perceptual judgments on models the paper never ran.
Beyond the paper: the blocks 5 to 15 choice for Stable Diffusion 3.5 suggests a principled block-selection criterion could be learned or derived, since the paper selects these blocks by visual inspection and computational efficiency.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Short version: JEDI is a plausible test-time fix for attribute mixing in text-to-image models, built on a simple and defensible idea: treat attention maps as probability distributions and minimize a normalized Jensen-Shannon divergence between subject groups instead of maximizing cosine similarity. The paper is cleanly written, the math is correct, the code is promised, and the efficiency claim (18 updates vs CONFORM's 69) is concrete. But the evidence for the headline claim is mostly qualitative, and the one quantitative-looking figure (Figure 10) measures the very quantity the method optimizes. Right now the paper earns 'promising', not 'consistently improves'.

What is actually new: applying JSD to attention-map disentanglement at inference, the specific DiT self-attention extraction trick for SD3.5, the sign-gradient update with 18 steps, and the CLIP-free disentanglement score. The bounds in Lemmas B.1 and B.2 are textbook, so the novelty is in the application and the efficiency story, not the mathematics.

Credit where it is earned: the method description is unambiguous, the synthetic toy comparison to NT-Xent is informative, the learning-rate sweep is honest, and the ablations (though visual) show the inter-group term is the active ingredient. The comparison to CONFORM is conducted under matched conditions and the runtime difference is concrete. The authors also note that when subjects are already well separated JEDI leaves the image unchanged, which is a plausible sign the loss is not just distorting everything.

Soft spots, in order of severity. First, the disentanglement score in Section 5 is inter-group JSD computed over blocks 7-15, essentially the objective being minimized; Figure 10 is expected by construction. That doesn't mean the visual results are fake, but it means the paper has not yet shown image-level disentanglement with an independent metric. Second, the SD3.5 block selection (5-15) and the bypass-softmax extraction are chosen post hoc, based on visual analysis. That makes the SD3.5 results harder to trust as a method evaluation; they are closer to a demonstration with selected hyperparameters. Third, the SD1.5 'quantitative experiments' announced in Section 4 never appear as numbers; there is no metric table, no seed variance, no user study. These are fixable problems, not fatal ones.

Bottom line: the central idea is reasonable and the qualitative evidence is suggestive. A serious referee should ask for an external metric (CLIP or human rating), seed-level variance, and a block-selection sensitivity analysis. I would send it out rather than desk-reject, but I would not yet cite it as evidence that JSD beats cosine similarity in general.

Referee Report

4 major / 5 minor

Summary. JEDI is a test-time adaptation method for diffusion models. At each early denoising step it extracts attention maps, groups them by prompt subject, and updates the latent by signed gradient descent on a loss of three terms: normalized Jensen-Shannon divergence within subject groups, separation between subject-mixture distributions, and a Shannon-entropy diversity regularizer. The paper proves upper bounds on JSD and entropy, demonstrates on toy Gaussians that JEDI preserves support structure better than NT-Xent, and reports qualitative comparisons on Stable Diffusion 1.5, Stable Diffusion 3.5, and LoRACLR. It also proposes an internal attention-based disentanglement score and claims roughly 67% faster updates than CONFORM.

Significance. If the empirical claims held, JEDI would be a simple, model-agnostic test-time intervention with a principled probabilistic objective and meaningful efficiency gains. The mathematical core is sound: Lemmas B.1 and B.2 are correct, and the synthetic comparison in Figure 2 is an instructive illustration of the difference between JSD and cosine/NT-Xent objectives. The paper also ships code and a project page. However, the empirical support for the central claim of consistent disentanglement is not yet established: the main score is the optimized objective itself, and the SD3.5 attention extraction protocol is selected post hoc on the evaluated examples. External, independent metrics and a predefined extraction protocol are needed before the central claims can be accepted.

major comments (4)

[Section 5, Figure 10] The claimed disentanglement score is not independent of the optimized objective. The inter-group loss term in Section 3 is exactly 1 − DJS(M), where M contains the per-subject mixture attention maps, and JEDI minimizes this term during optimization. Figure 10 then reports that JEDI produces higher inter-group JSD than the base model. This improvement is guaranteed by construction, because JEDI directly maximizes the reported quantity. To support the claim of image-level disentanglement, the paper needs an external metric (e.g., instance segmentation, object detection, CLIP-based alignment, or a blinded human study) that is not a function of the optimized attention maps.
[Appendix E] The SD3.5 attention extraction protocol is selected post hoc on the very images used for evaluation. The paper states that DiT blocks 5 to 15 are chosen 'based on visual analysis and computational efficiency' and that bypassing softmax during attention 'empirically yields more stable and consistent results.' Because the block range and softmax handling are tuned on the same prompts and images shown in Figures 4, 8, and 9, the reported improvements may reflect selection effects. A predefined protocol, or validation on a separate set of prompts and seeds, is necessary to rule out this confound.
[Section 4] The section header promises 'quantitative experiments' on Stable Diffusion 1.5, including a comparison to CONFORM, but the results are exclusively qualitative: no metric table, no error bars over seeds, and no user study are reported anywhere in the paper or appendices. The only quantitative claim in the CONFORM comparison is the number of update steps (18 versus 69), which is an efficiency measure, not a quality measure. The abstract's statement that JEDI 'consistently improv[es] prompt alignment and disentanglement' is therefore not supported by the experiments as presented. Please add quantitative evaluation or soften the claim.
[Section 5, 'Unbiased Disentanglement Score'] The label 'unbiased' is not justified for a score computed from the same internal attention distributions that JEDI directly optimizes. Even if the score is not identical to the full JEDI loss, it shares the same mechanism and is therefore circular for evaluating disentanglement. This is not merely a naming issue: Section 5 uses this score as the primary evidence that the method works. The paper should either supply an external evaluation or explicitly characterize the score as a diagnostic of the optimized objective, not as an independent benchmark.

minor comments (5)

[Figure 10] The caption says 'Intra-Group JSD' while the surrounding text and the discussion in Section 5 refer to inter-group JSD; these labels should be reconciled.
[Appendix B] In the proof of Lemma B.2, 'Langrage multiplier' should be 'Lagrange multiplier'.
[Appendix C, Algorithm 1] The attention maps A_t appear in the update step but are never returned by Model in the pseudocode; the extraction step should be made explicit.
[Section 4 and Appendix E] The number of prompts in the 'custom dataset' for SD3.5 is not stated; without a sample size it is difficult to interpret the claim that improvements are consistent.
[Section 3 and Appendix A] The method relies on several free hyperparameters (learning rate alpha, diversity weight lambda, optimization window K, and the DiT block range), but only alpha is analyzed; please report how the other fixed choices were made or provide a robustness check.

Circularity Check

1 steps flagged · score 6.0 of 10

The SD3.5 disentanglement score is the JEDI objective itself, so the reported quantitative improvement is guaranteed by construction rather than independently measured.

fitted input called prediction [Section 5 (Discussion and Future Work) and Appendix F, Figure 10]
"We find that the inter-group loss term in the JEDI objective naturally serves as an effective metric for measuring subject disentanglement during generation. For the images in Figure 1, the disentangled image achieves a mean JSD of 0.40 ± 0.15, compared to 0.17 ± 0.10 for the entangled counterpart, with full progression over time shown in Figure 10."

The JEDI objective's Inter-group Separation term minimizes 1 - D_JS(M) over subject-mixture attention distributions (Section 3). Appendix F defines Figure 10 as 'inter-group JSD between the two subjects ... computed across DiT blocks 7 to 15', which is exactly D_JS(M) over the same attention extraction used for optimization (blocks 5-15 optimized; 7-15 evaluated). Therefore JEDI directly maximizes the reported metric, so the higher Figure 10 value for JEDI relative to the base model is expected by construction. This is not an independent test of image-level disentanglement; the quantity reported as an evaluation metric is the same quantity being minimized as a loss.

full rationale

The strongest circularity is in the SD3.5 quantitative evaluation: the paper proposes the inter-group JSD as an unbiased disentanglement score, but this score is precisely the inter-group separation term of the JEDI loss, computed over the same attention blocks used for optimization. Thus the headline quantitative claim that JEDI 'consistently improves' disentanglement on SD3.5 is, in its only quantitative form, a restatement of the optimization objective rather than an independent outcome. The synthetic experiment and the SD1.5/CONFORM visual comparisons provide some independent qualitative content, and the mathematical lemmas are self-contained, which prevents a maximal circularity score. However, the central SD3.5 claim lacks an independent metric, so the paper's central quantitative evidence reduces by construction to its own loss. The Appendix E block selection ('Based on visual analysis and computational efficiency, we select blocks 5 to 15') is a post hoc choice that compounds the concern, but it is a selection risk rather than a separate definitional circularity.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The central method rests on the assumption that attention maps encode subject position as probability distributions. For SD3.5 this requires a hand-designed extraction over selected DiT blocks. Hyperparameters alpha, lambda, update window, and block range are chosen from visual inspection, and the evaluation score is the optimized objective. No new physical entities are introduced.

free parameters (4)

alpha (learning rate) = 0.003
Controls the signed gradient step size; chosen via the sweep in Figure 6 and stated in Section 3 and Appendix A.
lambda (diversity regularizer weight) = 0.01
Weights the entropy regularization term; stated in Methodology, with ablations showing a minimal effect in image generation.
optimization window K = 18 (SD1.5 and SD3.5), 30 (LoRACLR)
Number of initial denoising timesteps during which latent updates are applied; chosen per architecture in Appendix E.
DiT block range = blocks 5 to 15 of 24
Selected for SD3.5 'based on visual analysis and computational efficiency' in Appendix E; a post-hoc choice affecting the extracted attention distributions.

assumptions (4)

domain assumption Attention maps can be treated as discrete probability distributions over image positions.
Section 2 states that softmax normalization allows interpreting attention maps as discrete probability distributions, which is central to defining the JSD objective.
domain assumption The symmetrized self-attention expression in SD3.5 approximates the spatial influence of a prompt token.
Appendix E defines the approximation using A = softmax(QK^T) and a symmetrized term, then bypasses softmax because renormalization was less stable. This approximation is load-bearing and unproven.
domain assumption Cross-attention maps for a subject token group localize that subject in the image.
Used to define intra-group coherence and inter-group separation in Methodology; if token grouping is wrong, the objective optimizes the wrong alignment.
standard math JSD and Shannon entropy upper bounds hold as stated.
Lemmas B.1 and B.2 are correct and are used to normalize the loss components.

how reviews work

0 comments

Cite this review

Pith. "Pith review of JEDI: The Force of Jensen-Shannon Divergence in Disentangling Diffusion Models." pith.science (2026). https://pith.science/paper/FG4UPCGT

@misc{pith2026250519166,
  author       = {Pith},
  title        = {Pith review of: JEDI: The Force of Jensen-Shannon Divergence in Disentangling Diffusion Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FG4UPCGT}},
  note         = {Machine review of arXiv:2505.19166}
}

read the original abstract

We introduce JEDI, a test-time adaptation method that enhances subject separation and compositional alignment in diffusion models without requiring retraining or external supervision. JEDI operates by minimizing semantic entanglement in attention maps using a novel Jensen-Shannon divergence based objective. To improve efficiency, we leverage adversarial optimization, reducing the number of updating steps required. JEDI is model-agnostic and applicable to architectures such as Stable Diffusion 1.5 and 3.5, consistently improving prompt alignment and disentanglement in complex scenes. Additionally, JEDI provides a lightweight, CLIP-free disentanglement score derived from internal attention distributions, offering a principled benchmark for compositional alignment under test-time conditions. Code and results are available at https://ericbill21.github.io/JEDI/.

Figures

Figures reproduced from arXiv: 2505.19166 by the authors.

**Figure 1.** JEDI enables test-time subject disentanglement. For the prompt “A horse and a bear in a forest”, JEDI reduces attribute mixing and improves subject separation in Stable Diffusion 3.5. a single entity (attribute mixing), or the spatial arrangement may appear incoherent, refer to [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗

**Figure 2.** Optimization Evolution of JEDI and NT-Xent. Synthetic example with four overlapping distributions: blue/green correspond to one subject, red/purple to another. Overlaps of blue and green form teal, while red and purple form pink. JEDI preserves coherent group structure, while NT-Xent collapses modes. 3. Methodology We propose JEDI (Jensen-Shannon Divergence for Disentanglement at Inference), a test-time adaptation… view at source ↗

**Figure 3.** Comparison of JEDI and CONFORM on Stable Diffusion 1.5. Each image triplet was generated under identical conditions. For more details and examples, refer to Appendix G. Entropy ranges from 0 (single peak) to log d (uniform). We normalize it by log d, yielding Hˆ (p) ∈ [0, 1], which allows scale-independent balancing, where high entropy indicates spatial spread; low entropy implies tight localization. For a proof of… view at source ↗

Figures from the paper (11 more)

**Figure 4.** Figure 4: Side-by-side comparison of Stable Diffusion 3.5 (left) and Stable Diffusion 3.5 + JEDI (right). The base model often mixes attributes or omits subjects, while JEDI corrects these issues. See Figures 11 and 12 in Appendix G for full prompts and more examples. “<Messi> a…

**Figure 5.** Figure 5: Comparison between LoRACLR (left) and LoRACLR + JEDI (right). The base model without a control shows attribute mixing, while JEDI produces clearer subject separation. visually similar object pairs (e.g., “apple” and “pear”). For each prompt, we generate two images: one…

**Figure 6.** Figure 6: Effect of learning rate α on image generation. Outputs generated for the same prompt, “A horse and a bear in a forest,” using Stable Diffusion 3.5 under identical settings, varying only the learning rate. Higher α values lead to excessive changes in the latent space, d…

**Figure 7.** Figure 7: Effect of individual components in the JEDI objective. Outputs generated from the same prompt using Stable Diffusion 3.5 under identical sampling settings. The left column shows results with all components enabled. The middle columns each omit one component of the JEDI…

**Figure 8.** Figure 8: Attention maps for the subject “bear”. Extracted from diffusion timestep 13 of 28 across all 24 DiT blocks for the prompt “A horse and a bear in a forest”, using Stable Diffusion 3.5 with JEDI. The final generated image is shown in [PITH_FULL_IMAGE:figures/full_fig_p0…

**Figure 9.** Figure 9: Attention maps for the subject “horse”. Extracted from diffusion timestep 13 of 28 across all 24 DiT blocks for the prompt “A horse and a bear in a forest”, using Stable Diffusion 3.5 with JEDI. The final generated image is shown in [PITH_FULL_IMAGE:figures/full_fig_p…

**Figure 10.** Figure 10: presents the inter-group JSD between the two subjects from [PITH_FULL_IMAGE:figures/full_fig_p009_10.png]

**Figure 11.** Figure 11: Side-by-side comparison of Stable Diffusion 3.5 (left) and Stable Diffusion 3.5 + JEDI (right). Each image pair was generated under identical conditions with a guidance scale of 4.5 and 28 inference steps. 10 [PITH_FULL_IMAGE:figures/full_fig_p010_11.png]

**Figure 12.** Figure 12: Side-by-side comparison of Stable Diffusion 3.5 (left) and Stable Diffusion 3.5 + JEDI (right). Each image pair was generated under identical conditions with a guidance scale of 4.5 and 28 inference steps. 11 [PITH_FULL_IMAGE:figures/full_fig_p011_12.png]

**Figure 13.** Figure 13: Comparison of JEDI and CONFORM on Stable Diffusion 1.5. Each image triplet was generated under identical conditions with 50 inference steps and a guidance scale of 7.5. For details of each method, refer to Appendix E. 12 [PITH_FULL_IMAGE:figures/full_fig_p012_13.png]

**Figure 14.** Figure 14: Comparison between LoRACLR (left) and LoRACLR + JEDI (right). The baseline model exhibits attribute mixing between subjects (e.g., “Taylor” appears in football attire), whereas LoRACLR + JEDI achieves clearer subject disentanglement and preserves subject-specific feat…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

20 extracted references · 4 canonical work pages

[1]

write newline

" write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...
[2]

Make it count: Text-to-image generation with an accurate number of objects

Binyamin, L., Tewel, Y., Segev, H., Hirsch, E., Rassin, R., and Chechik, G. Make it count: Text-to-image generation with an accurate number of objects. arXiv preprint arXiv:2406.10210, 2024

arXiv 2024
[3]

Attend-and-excite: Attention-based semantic guidance for text-to-image diffusion models

Chefer, H., Alaluf, Y., Vinker, Y., Wolf, L., and Cohen-Or, D. Attend-and-excite: Attention-based semantic guidance for text-to-image diffusion models. ACM transactions on Graphics (TOG), 42 0 (4): 0 1--10, 2023

work page 2023
[4]

A simple framework for contrastive learning of visual representations

Chen, T., Kornblith, S., Norouzi, M., and Hinton, G. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pp.\ 1597--1607. PmLR, 2020

2020
[5]

Scaling rectified flow transformers for high-resolution image synthesis

Esser, P., Kulal, S., Blattmann, A., Entezari, R., M \"u ller, J., Saini, H., Levi, Y., Lorenz, D., Sauer, A., Boesel, F., et al. Scaling rectified flow transformers for high-resolution image synthesis. In Forty-first international conference on machine learning, 2024

work page 2024
[6]

J., Shlens, J., and Szegedy, C

Goodfellow, I. J., Shlens, J., and Szegedy, C. Explaining and harnessing adversarial examples. arXiv preprint arXiv:1412.6572, 2014

arXiv 2014
[7]

Z., Shi, Y., Chen, Y., Fan, Z., Xiao, W., Zhao, R., Chang, S., Wu, W., et al

Gu, Y., Wang, X., Wu, J. Z., Shi, Y., Chen, Y., Fan, Z., Xiao, W., Zhao, R., Chang, S., Wu, W., et al. Mix-of-show: Decentralized low-rank adaptation for multi-concept customization of diffusion models. Advances in Neural Information Processing Systems, 36: 0 15890--15902, 2023

work page 2023
[8]

Momentum contrast for unsupervised visual representation learning

He, K., Fan, H., Wu, Y., Xie, S., and Girshick, R. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp.\ 9729--9738, 2020

2020

Show all 20 references

[9]

Denoising diffusion probabilistic models

Ho, J., Jain, A., and Abbeel, P. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33: 0 6840--6851, 2020

2020
[10]

C., and Heilbron, F

Kwon, G., Jenni, S., Li, D., Lee, J.-Y., Ye, J. C., and Heilbron, F. C. Concept weaver: Enabling multi-concept fusion in text-to-image models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 8880--8889, 2024

2024
[11]

T., Ben-Hamu, H., Nickel, M., and Le, M

Lipman, Y., Chen, R. T., Ben-Hamu, H., Nickel, M., and Le, M. Flow matching for generative modeling. arXiv preprint arXiv:2210.02747, 2022

2022 arXiv
[12]

Liu, N., Li, S., Du, Y., Torralba, A., and Tenenbaum, J. B. Compositional visual generation with composable diffusion models. In European Conference on Computer Vision, pp.\ 423--439. Springer, 2022

2022
[13]

Meral, T. H. S., Simsar, E., Tombari, F., and Yanardag, P. Conform: Contrast is all you need for high-fidelity text-to-image diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 9005--9014, 2024

2024
[14]

Oord, A. v. d., Li, Y., and Vinyals, O. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018

2018 arXiv
[15]

and Xie, S

Peebles, W. and Xie, S. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pp.\ 4195--4205, 2023

2023
[16]

Sdxl: Improving latent diffusion models for high-resolution image synthesis

Podell, D., English, Z., Lacey, K., Blattmann, A., Dockhorn, T., M \"u ller, J., Penna, J., and Rombach, R. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023

2023 arXiv
[17]

Hierarchical text-conditional image generation with clip latents

Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., and Chen, M. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 0 (2): 0 3, 2022

2022 arXiv
[18]

High-resolution image synthesis with latent diffusion models

Rombach, R., Blattmann, A., Lorenz, D., Esser, P., and Ommer, B. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp.\ 10684--10695, 2022

2022
[19]

Loraclr: Contrastive adaptation for customization of diffusion models

Simsar, E., Hofmann, T., Tombari, F., and Yanardag, P. Loraclr: Contrastive adaptation for customization of diffusion models. arXiv preprint arXiv:2412.09622, 2024

2024
[20]

Enhancing mmdit-based text-to-image models for similar subject generation

Wei, T., Chen, D., Zhou, Y., and Pan, X. Enhancing mmdit-based text-to-image models for similar subject generation. arXiv preprint arXiv:2411.18301, 2024

2024 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

write newline

" write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...

[2] [2]

Make it count: Text-to-image generation with an accurate number of objects

Binyamin, L., Tewel, Y., Segev, H., Hirsch, E., Rassin, R., and Chechik, G. Make it count: Text-to-image generation with an accurate number of objects. arXiv preprint arXiv:2406.10210, 2024

arXiv 2024

[3] [3]

Attend-and-excite: Attention-based semantic guidance for text-to-image diffusion models

Chefer, H., Alaluf, Y., Vinker, Y., Wolf, L., and Cohen-Or, D. Attend-and-excite: Attention-based semantic guidance for text-to-image diffusion models. ACM transactions on Graphics (TOG), 42 0 (4): 0 1--10, 2023

work page 2023

[4] [4]

A simple framework for contrastive learning of visual representations

Chen, T., Kornblith, S., Norouzi, M., and Hinton, G. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pp.\ 1597--1607. PmLR, 2020

2020

[5] [5]

Scaling rectified flow transformers for high-resolution image synthesis

Esser, P., Kulal, S., Blattmann, A., Entezari, R., M \"u ller, J., Saini, H., Levi, Y., Lorenz, D., Sauer, A., Boesel, F., et al. Scaling rectified flow transformers for high-resolution image synthesis. In Forty-first international conference on machine learning, 2024

work page 2024

[6] [6]

J., Shlens, J., and Szegedy, C

Goodfellow, I. J., Shlens, J., and Szegedy, C. Explaining and harnessing adversarial examples. arXiv preprint arXiv:1412.6572, 2014

arXiv 2014

[7] [7]

Z., Shi, Y., Chen, Y., Fan, Z., Xiao, W., Zhao, R., Chang, S., Wu, W., et al

Gu, Y., Wang, X., Wu, J. Z., Shi, Y., Chen, Y., Fan, Z., Xiao, W., Zhao, R., Chang, S., Wu, W., et al. Mix-of-show: Decentralized low-rank adaptation for multi-concept customization of diffusion models. Advances in Neural Information Processing Systems, 36: 0 15890--15902, 2023

work page 2023

[8] [8]

Momentum contrast for unsupervised visual representation learning

He, K., Fan, H., Wu, Y., Xie, S., and Girshick, R. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp.\ 9729--9738, 2020

2020