The DOI matches its Crossref record; 61 of the 61 references it cites were found in a registry or an index; 1 of the 5 sources its claims cite were checked for support.
Scalable Diffusion Models with Transformers
Open accessAbstract
We explore a new class of diffusion models based on the transformer architecture. We train latent diffusion models of images, replacing the commonly-used U-Net backbone with a transformer that operates on latent patches. We analyze the scalability of our Diffusion Transformers (DiTs) through the lens of forward pass complexity as measured by Gflops. We find that DiTs with higher Gflops—through increased transformer depth/width or increased number of input tokens—consistently have lower FID. In addition to possessing good scalability properties, our largest DiT-XL/2 models outperform all prior diffusion models on the class-conditional ImageNet 512×512 and 256×256 benchmarks, achieving a state-of-the-art FID of 2.27 on the latter.
Overview
AI snapshot
Written from the paper’s full text.
TL;DR
Diffusion Transformers (DiTs) replace the standard U-Net backbone in latent diffusion models with a transformer operating on latent patches. DiT-XL/2 achieves state-of-the-art class-conditional ImageNet generation, reaching FID 2.27 at 256×256 and 3.04 at 512×512.
Key findings
- DiT block conditioning matters: adaLN-Zero outperforms cross-attention and in-context conditioning at all training stages while being the most compute-efficient; at 400K iterations its FID is nearly half that of the in-context model.
- Scaling DiT Gflops improves sample quality: across 12 models, transformer Gflops and FID-50K at 400K steps have a correlation of −0.93, with gains from increasing depth/width or number of input tokens.
- DiT-XL/2 sets state of the art: FID 2.27 on class-conditional ImageNet 256×256 with classifier-free guidance cfg=1.50, improving on LDM’s previous best 3.60; at 512×512 it reaches FID 3.04 versus ADM’s 3.85.
- Larger DiT models are more compute-efficient, and extra sampling compute cannot compensate for smaller models: DiT-L/2 with 1000 sampling steps uses 80.7 Tflops and gets FID-10K 25.9, while DiT-XL/2 with 128 steps uses 15.2 Tflops and gets FID-10K 23.7.
Methods
- Design: latent diffusion with a frozen off-the-shelf Stable Diffusion VAE (downsample factor 8); DiT patchifies latents, applies ViT-style positional embeddings, processes tokens through N transformer blocks, and linearly decodes to noise and covariance predictions. Design space includes patch sizes p=2,4,8; block variants in-context, cross-attention, adaLN, and adaLN-Zero; model sizes DiT-S/B/L/XL.
- Data and training: class-conditional ImageNet at 256×256 and 512×512; AdamW, constant learning rate 1×10⁻⁴, batch size 256, horizontal flips, EMA decay 0.9999, no learning-rate warmup or regularization; diffusion hyperparameters retained from ADM; implemented in JAX on TPU-v3 pods.
- Evaluation: FID-50K with 250 DDPM sampling steps using ADM’s TensorFlow evaluation suite; 12 DiT models swept over S/B/L/XL and p=8/4/2 at 400K steps; DiT-XL/2 trained for 7M steps at 256×256 and 3M steps at 512×512; additional comparisons of conditioning strategies and model versus sampling compute.
Limitations
- The paper does not state an explicit limitations section.
- Evident from the text: experiments are limited to class-conditional ImageNet at 256×256 and 512×512; text-to-image and other domains are left to future work.
- Evident from the text: DiTs use a frozen off-the-shelf VAE, so generation quality depends on that VAE; training hyperparameters were retained from ADM and not tuned; pixel-space DiT is mentioned as possible but not tested.
Model-written summary. Check it against the paper before relying on it.
Integrity signals
Registry
Cites
Statistics
Verification
Claims the paper makes
28 Own result · 5 Own interpretation · 4 No basis given · 3 Sources not found · 2 Own method · 1 Partly supported · 1 Not checked yet · accuracy not yet measured
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“As shown in Figure 2, LDMs achieve good performance while using a fraction of the Gflops of pixel space diffusion models like ADM.”
§ 3. Diffusion Transformers > 3.1. Preliminaries
What the sources say
- Partly supportedcites [48] sourceExperimental — accuracy not yet measured
Section 4.2 reports performance close to ADM on LSUN-Bedrooms while using four times fewer training resources, and Appendix D.4 describes competitive ImageNet performance with lower compute requirements. These passages support the performance-and-efficiency comparison, but our copy does not give a GFLOPs comparison with ADM; it reports training resources instead. (LLM, conf=0.85)
plaindirect- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Partly supported
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“Larger DiT models are more compute-efficient.”
plaindirect- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
The paper states this without a source or data behind it here, so there is nothing to check it against. The quote shows it is in the paper.
In the paper
“Adaptive layer norm works best.”
plaindirect- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: unclear — It is unclear why the paper says it holds.
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“Transformer Gflops are strongly correlated with FID.”
boostedcorrelational- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“Performance steadily improves in FID as model flops increase.”
plaincorrelational- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“Increasing transformer forward pass Gflops increases sample quality.”
plaindirect- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“The results demonstrate that different DiT configs obtain similar FID values when their total Gflops are similar (e.g., DiT-S/2 and DiT-B/4).”
boostedcorrelational- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“Indeed, scaling both model size and the number of tokens yields notable improvements in visual quality.”
plaindirect- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“Across all four configs, significant improvements in FID are obtained over all stages of training by making the transformer deeper and wider.”
plaindirect- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“Scaling the transformer backbone yields better generative models across all model sizes and patch sizes.”
plaindirect- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- The quote only partly states this claim. 4
- Probably not a substantial claim. 6
- It may state two separate claims as one. 18
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“Our best model, DiT-XL/2, is compute-efficient and outperforms all prior U-Net-based diffusion models, like ADM and LDM.”
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
The quote only partly states this claim.
The paper states this without a source or data behind it here, so there is nothing to check it against. The quote shows it is in the paper.
In the paper
“Cross-attention adds the most Gflops to the model, roughly a 15% overhead.”
§ 3. Diffusion Transformers > 3.2. Diffusion Transformer Design Space
plaindirect- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: fails — The paper’s conclusions may not depend on it.
- Based: passes — The paper gives a reason why it holds.
Probably not a substantial claim. The paper’s conclusions may not depend on it. It may state two separate claims as one.
None of the sources cited for this claim have been checked yet.
In the paper
“As with prior work using guidance, we observe that larger scales increase visual fidelity and decrease sample diversity.”
plaindirect- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
It may state two separate claims as one.
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“For both the 256 × 256 and 512 × 512 DiT-XL/2 models, we never observed FID saturate and continued training them as long as possible.”
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: fails — The paper’s conclusions may not depend on it.
- Based: passes — The paper gives a reason why it holds.
Probably not a substantial claim. The paper’s conclusions may not depend on it. It may state two separate claims as one.
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“Of the three block designs we explore, adaLN adds the least Gflops and is thus the most compute-efficient.”
§ 3. Diffusion Transformers > 3.2. Diffusion Transformer Design Space
plaindirect- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
It may state two separate claims as one.
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“Our best model, DiT-XL/2, is compute-efficient and outperforms all prior U-Net-based diffusion models, like ADM and LDM.”
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
It may state two separate claims as one.
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“Table 2. Benchmarking class-conditional image generation on ImageNet 256×256. DiT-XL/2 achieves state-of-the-art FID.”
§ 107 108 109 1010 1011 1012 > 5.1. State-of-the-Art Diffusion Models
plaindirect- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
It may state two separate claims as one. The quote only partly states this claim.
- Rests on an assumption: FID-50K values reported by prior work are directly comparable to the FID-50K values computed in this paper.
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“When trained for only 2.35M steps (similar to ADM), XL/2 still outperforms all prior diffusion models with an FID of 2.55.”
§ 107 108 109 1010 1011 1012 > 5.1. State-of-the-Art Diffusion Models
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
It may state two separate claims as one.
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“In all cases, we find that increasing model size and decreasing patch size yields considerably improved diffusion models.”
plaindirect- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
It may state two separate claims as one.
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“XL/2 again outperforms all prior diffusion models at this resolution, improving the previous best FID of 3.85 achieved by ADM to 3.04.”
§ 107 108 109 1010 1011 1012 > 5.1. State-of-the-Art Diffusion Models
plaindirect- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
It may state two separate claims as one.
- Rests on an assumption: FID-50K values reported by prior work are directly comparable to the FID-50K values computed in this paper.
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“we did not find learning rate warmup nor regularization necessary to train DiTs to high performance.”
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
It may state two separate claims as one.
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“Increasing the Gflops in the model—either by increasing transformer depth/width or increasing the number of input tokens—yields significant improvements in visual fidelity.”
plaindirect- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
It may state two separate claims as one.
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“In all cases, we find that increasing model size and decreasing patch size yields considerably improved diffusion models.”
plaindirect- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
The quote only partly states this claim.
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“We find a strong negative correlation between model Gflops and FID-50K, suggesting that additional model compute is the critical ingredient for improved DiT models.”
- Also stated in: § 100 101 102 · boosted
“Transformer Gflops are strongly correlated with FID.”
boosteddirect- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
It may state two separate claims as one.
- Also stated in: § 100 101 102 · boosted
The paper states this without a source or data behind it here, so there is nothing to check it against. The quote shows it is in the paper.
In the paper
“Finally, we also observe that DiT-XL/2 achieves higher recall values at all tested classifier-free guidance scales compared to LDM-4 and LDM-8.”
§ 107 108 109 1010 1011 1012 > 5.1. State-of-the-Art Diffusion Models
plaincorrelational- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
The quote only partly states this claim.
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“Increasing DiT model Gflops (via transformer size or number of input tokens) causes the training loss to decrease more rapidly and saturate at a lower value.”
- Also stated in: § 0 10K 20K 30K 40K 50K 60K 70K 80K 90K 100K · plain
“Note that scaled-up DiT models exhibit lower training losses.”
plaindirect- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
It may state two separate claims as one.
- Also stated in: § 0 10K 20K 30K 40K 50K 60K 70K 80K 90K 100K · plain
What the paper did — it cites no source for this, so there is none to check. The quote shows it is in the paper.
In the paper
“We did not tune learning rates, decay/warm-up schedules, Adam β1/β2 or weight decays.”
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: fails — The paper’s conclusions may not depend on it.
- Based: passes — The paper gives a reason why it holds.
Probably not a substantial claim. The paper’s conclusions may not depend on it. It may state two separate claims as one.
In the paper
“which have been shown to scale more effectively for visual recognition than traditional convolutional networks (e.g., ResNet [15]).”
What the sources say
- Not matchedcites [10] Dosovitskiy, An image is worth 16x16 words: Transformers for image recognition at scale
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
boosteddirect- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
Probably not a substantial claim.
- Not matched
What the paper did — it cites no source for this, so there is none to check. The quote shows it is in the paper.
In the paper
“Even without these techniques, training was highly stable across all model configs and we did not observe any loss spikes commonly seen when training transformers.”
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
It may state two separate claims as one.
None of the sources cited for this claim could be matched to a paper in Crossref, so there is nothing to check it against.
In the paper
“diffusion models have been at the forefront of recent advances in image-level generative models [9,46]; yet, they all adopt a convolutional U-Net architecture as the de-facto choice of backbone.”
What the sources say
- Not matchedcites [9] Dhariwal, Diffusion models beat gans on image synthesis
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- Not matchedcites [46] Ramesh, 2022, Hierarchical text-conditional image generation with clip latents
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
It may state two separate claims as one.
- Not matched
In the paper
“Note that prior work [9] measures Precision and Recall using 1000 real samples for 512 × 512 resolution; for consistency, we do the same.”
§ 107 108 109 1010 1011 1012 > 5.1. State-of-the-Art Diffusion Models
What the sources say
- Not matchedcites [9] Dhariwal, Diffusion models beat gans on image synthesis
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
plainn = 1,000- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
Probably not a substantial claim.
- Not matched
A conclusion the paper draws from its own results — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“halving pwill quadruple T , and thus at least quadruple total transformer Gflops.”
§ 3. Diffusion Transformers > 3.2. Diffusion Transformer Design Space
plaindirect- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
It may state two separate claims as one.
A conclusion the paper draws from its own results — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“A smaller patch size p results in a longer sequence length and thus more Gflops.”
§ 3. Diffusion Transformers > 3.2. Diffusion Transformer Design Space
plaindirect- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: fails — The paper’s conclusions may not depend on it.
- Based: passes — The paper gives a reason why it holds.
Probably not a substantial claim. The paper’s conclusions may not depend on it. It may state two separate claims as one.
The paper relies on this without support.
FID-50K values reported by prior work are directly comparable to the FID-50K values computed in this paper.
“We show samples from the model in Figures 1, and we compare against state-of-the-art class-conditional generative models. We report results in Table 2.”
Claims other papers make about this one
Each is checked against this paper’s own text: does it say what the citing paper says it does?
The cited source says what the claim attributes to it.
In the citing paper
“Unlike prior works that employ text-to-image (T2I) U-Net [10, 33, 35, 40] for talking head generation, MoCha is built on a diffusion transformer (DiT) [26]. By incorporating text and speech conditions sequentially via cross-attention, it effectively captures both semantics and temporal dynamics.”
From MoCha: Towards Movie-Grade Talking Character Synthesis (2025) · cites this paper as [26]
In the citing paper
“Supervised learning, a foundational approach in AI, uses labeled datasets to guide models towards accurate predictions, and it has been integral to various applications, including image recognition and NLP [163], [164], [165].”
From From Google Gemini to OpenAI Q* (Q-Star): A Survey on Reshaping the Generative Artificial Intelligence (AI) Research Landscape (2025) · cites this paper as [164]
Citations
Cites
Deep Residual Learning for Image Recognition
He, K., Zhang, X., Ren, S., & Sun, J. (2016) · 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) · doi:10.1109/cvpr.2016.90 · Cited by 235,474
Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., Kaiser, L., & Polosukhin, I. (2017) · arXiv · doi:10.48550/arxiv.1706.03762 · Cited by 120,404
ImageNet classification with deep convolutional neural networks
Krizhevsky, A., Sutskever, I., & Hinton, G. E. (2017) · Communications of the ACM · doi:10.1145/3065386 · Cited by 112,155
U-Net: Convolutional Networks for Biomedical Image Segmentation
Ronneberger, O., Fischer, P., & Brox, T. (2015) · Lecture notes in computer science · doi:10.1007/978-3-319-24574-4_28 · Cited by 99,404
Adam: A Method for Stochastic Optimization
Kingma, D. P., & Ba, J. (2014) · arXiv · doi:10.48550/arxiv.1412.6980 · Cited by 98,946
BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding
Devlin, J., Chang, M.-W., Lee, K., & Toutanova, K. (2019) · Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers) · doi:10.18653/v1/n19-1423 · Cited by 54,325
Language Models are Few-Shot Learners
Brown, T. B., Mann, B., Ryder, N., Subbiah, M., Kaplan, J., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al. (2020) · arXiv · doi:10.48550/arxiv.2005.14165 · Cited by 23,372
Image-to-Image Translation with Conditional Adversarial Networks
Isola, P., Zhu, J.-Y., Zhou, T., & Efros, A. A. (2017) · 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) · doi:10.1109/cvpr.2017.632 · Cited by 22,930
High-Resolution Image Synthesis with Latent Diffusion Models
Rombach, R., Blattmann, A., Lorenz, D., Esser, P., & Ommer, B. (2022) · 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) · doi:10.1109/cvpr52688.2022.01042 · Cited by 20,302
Auto-Encoding Variational Bayes
Kingma, D. P., & Welling, M. (2013) · arXiv · doi:10.48550/arxiv.1312.6114 · Cited by 20,067
Research works printed but not linked: 3
- Chen, Generative pre-training from pixels
- Dosovitskiy, An image is worth 16x16 words: Transformers for image recognition at scale
- Goodfellow, Generative adversarial nets
Cited by
Sorted by citations
Items 3,601–3,684 of 3,684
Rectified Language FlowPreprint
Su, B., Liao, R., Chen, L., & Liu, Q. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2610.33811
Beyond Distribution Matching: Self-Supervised Representation Forcing for Few-Step Video GenerationPreprint
Chen, J., Zhang, H., Fu, Z., Wang, T., & Guo, C. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2610.35594
PED in 2027: integrating experimental and predicted protein conformational ensembles at scale
Monzón, A. M., Ghafouri, H., Castro, X. A., Khezrian, S., Aspromonte, M. C., Singla, K., Barethiya, S., Bernadó, P., Bonomi, M., Chen, J., et al. (2026) · Nucleic Acids Research · doi:10.1093/nar/gkag1048
Video World Models Need Predictable LatentsPreprint
Scardecchia, M., & Yang, S. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2610.34663
Kang, G., Liu, P., Qian, B., & Cheng, X. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2610.34834
TriTile: Three-Tier Attention for Video Diffusion TransformersPreprint
Kim, Y., Kwon, O., Seo, S. H., & Lee, J. W. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2610.35706
安向京, Jiao, Z., & Chen, A. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2610.33451
When Diffusion Recommenders Stop Sampling: Preserving Generative DiversityPreprint
Martin, L. C. S., Rakotomamonjy, A., & Gallinari, P. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2610.36954
Video frame interpolation in the era of deep learning and generative AI: a systematic survey
Wang, S., Dong, Z., Kang, Z., & Pang, M. (2026) · The Visual Computer · doi:10.1007/s00371-026-04772-9
Read, Allocate, Route: Role-Coupled Tokenization for Pixel-Space Generative TransformersPreprint
Yang, J., Fan, W., Diao, H., Cai, Z., & Liu, Z. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2610.40647
SaiVLA-JEPA: Evaluating Future Latent Prediction in JEPA-Conditioned VLA ControlPreprint
Xiao, R., Xiao, R., Chen, W., & Sun, X. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2610.38508
Evaluation Pitfalls in Post-Training Quantization of Diffusion Transformers: An Empirical Audit
Jin, Z. (2026) · Frontiers in Computing and Intelligent Systems · doi:10.54097/v6881m05
PTE: A Novel Model for Generating Electrocardiogram Through Photoplethysmography
Dong, X., Xin, W., & Zhang, J. (2026) · doi:10.1109/icsip70099.2026.11744369
WeatherBridge and the limits of global weather interpolation at held-out query timesPreprint
Sukhorukov, D., Zakharov, A., Maximov, Y., Makarov, I., & Oseledets, I. (2026) · Research Square · doi:10.21203/rs.3.rs-11183024/v1
Indoor Scene Synthesis via Conditional Diffusion Model with Residual Enhanced U-Net
Li, Y. (2026) · International Conference on Computer Engineering and Application (ICCEA) · doi:10.1109/iccea69962.2026.11743518
Reuse, Reduce, Compute: A Similarity-Aware Accelerator for Diffusion Model Inference
Qian, J., Wei, Y., Lu, S., Qin, D., Wang, Q., Shi, X., Wu, C., He, L., & Shi, L. (2026) · International Symposium of Electronics Design Automation (ISEDA) · doi:10.1109/iseda70083.2026.11751525
Fan, L., Zheng, Y., Mao, D., Ran, H., Hu, S., Mao, W., Zhao, R., Wu, W., Fang, G., Chen, Y., et al. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2610.40191
Fast simulation in ATLAS for LHC Run 3 and beyond
Corchia, F. A. G. (2026) · CNR Solar (Scientific Open-access Literature Archive and Repository) (Consiglio Nazionale delle Ricerche)
Small molecule structure elucidation from NMR spectra with generative and reasoning modelsPreprint
Xiong, Z., Ying, V. Y., Huang, J. Z., Koch, Z., Clark, J. D., Heeter, R. C., An, J. S., Matanachai, K., Rodriques, S. G., White, A. D., et al. (2026) · bioRxiv (Cold Spring Harbor Laboratory) · doi:10.64898/2026.10.29.762711
DIFFNAS: Predictor-Free Neural Architecture Search via Diffusion Model and Zero-Cost EvaluationPreprint
Sun, Y., 马翰章, Wu, Z., Li, X., Volker Tresp, C., & Ma, Y. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2610.41818
T ext ZGCA: Text-Guided Zero-Gated Cross-Attention for 3D Human Motion Prediction
Qiao, G., Zhou, L., Jiang, D., & Wang, J. (2026) · ACM International Conference on Information and Knowledge Management (CIKM) · doi:10.1145/3799682.3839952
A Better Critic Matters for Distribution Matching DistillationPreprint
Wu, Z., Wang, X., Geng, Z., Cao, M., Zhang, Z., Zhang, R., Shechtman, E., & Cao, R. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2610.42135
Semantic diffusion and large language model collaboration for cross-modal sign language translation
Wei, S., & Yuan, T. (2026) · Multimedia Systems · doi:10.1007/s00530-026-02768-y
Closing Generalization Gaps in Continual Face Forgery Detection
Moon, B., Le, B. M., & Woo, S. S. (2026) · ACM International Conference on Information and Knowledge Management (CIKM) · doi:10.1145/3799682.3841128
Hwang, J., & Jo, J. (2026) · IEEE Access · doi:10.1109/access.2026.3748904
Zhang, Y., Duan, J., Junfu, Z., Zou, Y., & Feng, K. (2026) · ACM International Conference on Information and Knowledge Management (CIKM) · doi:10.1145/3799682.3841011
Diffusion and Backward Markovian Models
Sayed, A. H. (2026) · Emerald Group Publishing Limited eBooks · doi:10.1108/978-1-83742-670-620261001
Z-Cache: Accelerating Diffusion Transformers via Self-ReflectionPreprint
Cheng, Z., Wang, Z., Liu, J., Tu, X., Zou, C., Chen, J., Ma, Y., Ma, Z., Ren, J., Zou, L., et al. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2611.01846
Learned Warm Starts for Steady Transonic RANS Simulations of Wing and Aircraft ConfigurationsPreprint
Zafar, M. I., Abuhanieh, S., & Yiğit, Ş. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2611.01240
FlowBatt: Flow Matching for Probabilistic Battery Degradation PredictionPreprint
Eivazi, H., Sharaawy, E., Wittek, S., Hebenbrock, A., Ginster, R., & Blömeke, S. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2611.03150
Sonny-ENS: demystifying machine-learning ensemble weather forecastingPreprint
Cheon, M. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2611.01614
W2A: Distilling World Priors into Action Models for Real-Time Robotic ControlPreprint
Zhang, X., Mu, Y., Weng, Y., Zhao, C., & Li, Y. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2611.02600
Learned Surrogates of Three-Dimensional Magnetohydrodynamic TurbulencePreprint
Kacmaz, S., & Huerta, E. A. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2611.01356
Huang, Z., You, T., & Chen, G. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2611.00027
Makwana, D., & Mantravadi, A. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2611.01913
MultiBinder: Preventing Multi-Subject Identity Leakage in Personalized Video Generation
Li, W., Zhang, Y., Qiu, Z., Yao, T., Sun, X., & Mei, T. (2026) · ACM International Conference on Multimedia (ACM MM) · doi:10.1145/3767308.3836548
Liu, S., Ren, S., Zhu, X., Liang, Q., Li, Z., Li, Q., Hu, X., & Huang, K. (2026) · ACM International Conference on Multimedia (ACM MM) · doi:10.1145/3767308.3835997
Adversarial Attacks on Deepfake Detectors: A Challenge in the Era of AI-Generated Media
Guarnera, F., Guarnera, L., Montibeller, A., Sambataro, S. A., Pishvai, D. S., Beritelli, L., Puglisi, G., Akhtar, Z., Boato, G., & Battiato, S. (2026) · ACM International Conference on Multimedia (ACM MM) · doi:10.1145/3767308.3837702
Vu, T.-T. D., Vu, A.-K. N., & Nguyen, V.-T. (2026) · ACM International Conference on Multimedia (ACM MM) · doi:10.1145/3767308.3836403
IMPEL: Can Pretrained Image Diffusion Enable Realistic LiDAR Point Cloud Generation?
Li, H., Shen, D., Zhou, Q., Cui, Y., & Gu, J. (2026) · ACM International Conference on Multimedia (ACM MM) · doi:10.1145/3767308.3835965
Ren, B., Zhao, B., & Yuan, Y. (2026) · ACM International Conference on Multimedia (ACM MM) · doi:10.1145/3767308.3836129
Diff-TrajUViT: A Transformer-Based U-Net Model for Trajectory Generation
Huang, H., Zhang, Q., Wang, J., Zhu, T., Han, L., Zhang, R., & Sun, L. (2026) · Lecture notes in computer science · doi:10.1007/978-981-92-7741-4_5
ACM Multimedia 2026 Grand Challenge Report: Single-Image Guided Multi-Angle Image Synthesis
Chen, J., Xie, H., Qi, H., Yang, J., Song, S., Wang, C., & Lao, Y. (2026) · ACM International Conference on Multimedia (ACM MM) · doi:10.1145/3767308.3838687
GenTS: Unified Generative Modeling for Multimodal Time Series Analysis
Gu, S., Zhao, Y., Jing, B., & Ren, K. (2026) · ACM International Conference on Multimedia (ACM MM) · doi:10.1145/3767308.3836409
AgriVision: adversarial image detection using Deep Learning in agricultural contexts
Yardımcı, M. O., Kulkarni, A., Ward, T., & Batarseh, F. A. (2026) · Scientific Reports · doi:10.1038/s41598-026-77167-4
Tuning-Free Continuous Visual Synthesis via Semantic-Aware Neural Flow
Li, S., Tan, S., & Pan, Y. (2026) · ACM International Conference on Multimedia (ACM MM) · doi:10.1145/3767308.3836061
Hitchcock-300K: A Master-Level Camera Control Dataset for Video Generation and Editing
Peng, B., Wang, Y., YaWen, Z., Yuan, J., Wang, Y., Sun, H., Zhou, Q., Jiang, S., Jiang, W., Song, J., et al. (2026) · ACM International Conference on Multimedia (ACM MM) · doi:10.1145/3767308.3838647
Next-Rank Prediction: Autoregressive Image Generation via Rank-Ordered Visual Tokenization
Gao, W., Xue, J., Lang, C., Li, Y., & Cao, Y. (2026) · ACM International Conference on Multimedia (ACM MM) · doi:10.1145/3767308.3836487
PanelLayout: Story-Aware Hierarchical Panel Layout Generation via Synchronized Diffusion
Tian, Y., 任, 海., Deng, X., Zhang, Y., Lu, H., & Zhang, L. (2026) · ACM International Conference on Multimedia (ACM MM) · doi:10.1145/3767308.3835438
CacheParallel: Comprehensively Exploring Multi-level Acceleration for Diffusion Models
Lyu, Q., Niu, D., Xu, J., Jiang, J., Yang, X., Ma, S., & Li, D. (2026) · ACM/IEEE Design Automation Conference (DAC) · doi:10.1145/3770743.3804252
Back to Music: Enhancing Piano Performance Motion Generation via Acoustic-Consistency Supervision
Cai, M., Li, J., Fan, Z., Yu, M., & Zhang, J. (2026) · ACM International Conference on Multimedia (ACM MM) · doi:10.1145/3767308.3835629
RadAR: Accelerating Autoregressive Visual Generation with Radial Decoding
Wang, S., Li, H., Li, W., Hu, J., Chen, X., Huang, J., & Zhao, F. (2026) · ACM International Conference on Multimedia (ACM MM) · doi:10.1145/3767308.3835138
StraightEdit: Efficient Low-Step Inversion-Free Text-Guided Image Editing
Wang, X., Zheng, M., Liu, D., & Xu, C. (2026) · ACM International Conference on Multimedia (ACM MM) · doi:10.1145/3767308.3835799
He, X., Fan, Z., Li, H., Zhuo, F., Xu, H., cheng, S., Weng, D., Liu, H., Ye, C., & Wu, B. (2026) · ACM International Conference on Multimedia (ACM MM) · doi:10.1145/3767308.3835382
DynaCartoon: 3D Dynamics-Guided Cartoon Video Generation from a Single Illustration and Forces
Chen, Z., Quan, Y., Mao, H., Zhao, Y., & Yang, X. (2026) · ACM International Conference on Multimedia (ACM MM) · doi:10.1145/3767308.3835019
Discrete Coding and Masked Modeling for Text-to-Stereo Audio Generation
Li, Y., Sun, P., Ye, Z., Pan, J., Cheng, S., Kang, B., Dai, J., Li, K., Lu, L., & Xue, W. (2026) · ACM International Conference on Multimedia (ACM MM) · doi:10.1145/3767308.3835212
ZeroCut: Seamless Transition Generation between Visually Distinct Video Scenes
He, R., Fu, S., Liu, Y., She, D., Li, Y., Fu, Z., & Huang, J. (2026) · ACM International Conference on Multimedia (ACM MM) · doi:10.1145/3767308.3835696
Fast-Rotating Panoramic HDR Reconstruction from Single-Second Spikes
Tao, Y., Chang, Y., Wei, S., Nie, L., Li, Q., Li, B., Shi, B., & Zhao, Y. (2026) · ACM International Conference on Multimedia (ACM MM) · doi:10.1145/3767308.3835660
TC-Flow: Trajectory-Coupled Multimodal Flow Matching for Incomplete Modality Learning
Meng, H., Mo, L., Fang, Y., Qin, K., Duan, G., & Dai, R. (2026) · ACM International Conference on Multimedia (ACM MM) · doi:10.1145/3767308.3835736
SegMoDiff: Semantic-Faithful Text-to-Motion Generation via Segmented Raw-Sequence Diffusion
Wang, J., Wei, S., Li, F., Cong, R., Zhao, Y., & Bai, H. (2026) · ACM International Conference on Multimedia (ACM MM) · doi:10.1145/3767308.3835713
Rectifying and Consolidating Latent Diffusion for Remote Sensing Object Detection
Li, F., Zhang, X., & Zhao, W. (2026) · ACM International Conference on Multimedia (ACM MM) · doi:10.1145/3767308.3835006
StyleCom: Taming Style Transfer and Collage Prior for Training-Based Cross-Domain Image Composition
Lyu, Q., Yang, T., Wu, K., Li, Y., Yuanjun, L., Qiu, Z., Shi, H., Gong, Y., Si, J., Chen, F., et al. (2026) · ACM International Conference on Multimedia (ACM MM) · doi:10.1145/3767308.3835612
Cai, Y. (2026) · ACM International Conference on Multimedia (ACM MM) · doi:10.1145/3767308.3835405
Real-Time Co-Speech Gesture Video Generation
Zhu, J., Xu, S., Tan, S., Yang, J., & Pan, Y. (2026) · ACM International Conference on Multimedia (ACM MM) · doi:10.1145/3767308.3834918
Excavating Consistency Across Editing Steps for Effective Multi-Step Image Editing
Qi, C., Song, Z., Gu, C., Fu, Z. F., Chang, J., Wang, J., Liang, X., & Guan, H. (2026) · ACM/IEEE Design Automation Conference (DAC) · doi:10.1145/3770743.3804383
Wu, X., Yang, D., Sun, H., Liang, X., Feng, C., & Ran, J. (2026) · ACM International Conference on Multimedia (ACM MM) · doi:10.1145/3767308.3835522
Lee, D., Lee, Y., & Yang, J. (2026) · ACM International Conference on Multimedia (ACM MM) · doi:10.1145/3767308.3835651
World Models for Embodied Scene Anomaly Detection in Robotic Cleaning
Wang, Y., Li, Y., Zheng, S., Zhao, Q., Zhao, Y., liu, R., Sun, B., Zhang, R., Liu, K., & Yang, X. (2026) · ACM International Conference on Multimedia (ACM MM) · doi:10.1145/3767308.3835000
Zhang, B., Liu, Z., Ding, S., Chen, W., Li, X. R., Liu, J., Ma, L., & Liu, R. (2026) · ACM International Conference on Multimedia (ACM MM) · doi:10.1145/3767308.3834923
Tan, X., Yu, G. W., Luo, Y., Liu, C., Wang, S., Liu, F., & Cai, Z. (2026) · ACM International Conference on Multimedia (ACM MM) · doi:10.1145/3767308.3835493
Reflection Proxy Purification: A Generative Data Engine for Real-World Reflection Separation
Ren, B., Zhao, B., & Yuan, Y. (2026) · ACM International Conference on Multimedia (ACM MM) · doi:10.1145/3767308.3835049
SpRDA: An Efficient Sparse Accelerator for Robotics Diffusion Model
Jian, R., Chen, Y., Zhao, P., Chen, Y., Wang, Z., 汪雪玲, Zhang, C., & He, G. (2026) · ACM/IEEE Design Automation Conference (DAC) · doi:10.1145/3770743.3804269
DSR-Tryon: Dual-Scale Semantic Reasoning for Versatile and Coherent Virtual Try-On
Wang, P., Wen, R., Xu, X., Guo, Q., & Ge, T. (2026) · ACM International Conference on Multimedia (ACM MM) · doi:10.1145/3767308.3835735
Zhuoxu, J., Liu, J., Yan, W., Wang, X., Jiang, S., Yang, Q., & Li, Z. (2026) · ACM International Conference on Multimedia (ACM MM) · doi:10.1145/3767308.3835762
SyncDiT: Parallel DiT with Spatiotemporal Synchronization for Joint Audio-Video Generation
李恒森, & Fang, T. (2026) · ACM International Conference on Multimedia (ACM MM) · doi:10.1145/3767308.3835200
TextAE: A Benchmark and Flow-Matching Baseline for Text-Controlled Layered Animation
Rong, J., Yang, H., Shan, X., Yuan, D., Zhang, Q., & Deng, X. (2026) · ACM International Conference on Multimedia (ACM MM) · doi:10.1145/3767308.3835784
STRIDE: Semantically Stratified Historical Retrieval for LiDAR World Modeling
Liu, Y., Yang, W., Liang, A., Xu, X., Kong, L., Chen, R., & Fei, B. (2026) · ACM International Conference on Multimedia (ACM MM) · doi:10.1145/3767308.3835050
HeatOL: Heat Dissipation Inspired Operator Learning for Point Cloud Generation
Yan, T., Dai, W., Zou, J., Zheng, Z., Kan, N., Li, C., & Xiong, H. (2026) · ACM International Conference on Multimedia (ACM MM) · doi:10.1145/3767308.3834993
ColorFlow: Robust Image Colorization with Pixel Flow Matching
Xue, D., Dai, L., Wang, H., & Tang, J. (2026) · ACM International Conference on Multimedia (ACM MM) · doi:10.1145/3767308.3835845
ClickSteer: One-click residual routing for precise instruction-based image editing
Yin, K., Gao, S., Xu, Y., He, X., Fu, J., & Zhang, W. (2026) · Pattern Recognition · doi:10.1016/j.patcog.2026.115259
Liming, 魏. W., & Zhang, Y. (2026) · Communications in computer and information science · doi:10.1007/978-981-92-7537-3_36
T2I-Eval-R1: Reinforcement Learning-Driven Reasoning for Interpretable Text-to-Image Evaluation
Ma, Z.-A., Lan, T., Tu, R.-C., Liu, S.-H., Huang, H., Wu, Z., Xu, C., & Mao, X.-L. (2025) · arXiv · doi:10.48550/arxiv.2505.17897
Metadata
- Authors
- Peebles, W., & Xie, S.
- Affiliations
- University of California, Berkeley (US)
- New York University (US)
- ORCID
- Xie, Saining
- Venue
- 2023 IEEE/CVF International Conference on Computer Vision (ICCV)
- Publisher
- IEEE
- Published
- 2023-10-01
- Pages
- 4172-4182
- OpenAlex
- W4390872297
- Semantic Scholar
- 736973165f98105fec3729b7db414ae4d80fcbeb
- Cited by
- 2,284
- References
- 59
- Citations by year
- 2026: 1007 · 2025: 925 · 2024: 315 · 2023: 37
- FWCI
- 244.26
- Type
- conference paper
- Language
- en
- Open access
- green
- arxiv.org
- Locations
- IEEE/CVF International Conference on Computer Vision (ICCV) (publishedVersion)
- arXiv (Cornell University) (submittedVersion, cc-by)
- arXiv (Cornell University) (cc-by)
- Topics
- Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Generative Adversarial Networks and Image Synthesis
- Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Advanced Neural Network Applications
- Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Image and Signal Denoising Methods
- Keywords
- diffusion models, latent diffusion models, diffusion transformer, scalability, image generation
About these checks
Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.
A language model drew the claims from the paper on 2026-11-08, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).
A second model labelled each claim on 2026-11-08 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).
Cited sources are checked with a language model’s help (gpt-6-sol).
Statistics and arithmetic are recomputed, not judged.
Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.
What counts as a claim ›How we measure confidence ›How we check this ›
Are you an author, or do you see a problem with this page? Request a review
