Full search page →

JiT — clean-image prediction in pixel space

Optional · paper guide 5 of 6 for ML 08 · Diffusion Models and Flow Matching. Citation numbers refer to the reference list at the bottom of this page.

Navigate: this page and this track: Navigate: this page and this track

Question and thesis

Why can large-patch pixel Transformers work when asked to predict clean images yet fail when directly predicting noise or velocity? Li and He argue that clean natural images occupy a much lower-dimensional structure than the surrounding noisy space [42]. Their Just image Transformers, or JiT, study the interaction between prediction target and network capacity. Read ViT patch dimensions and the chapter’s distinction between direct output and loss space first.

Method and figure

Pixel patches pass through a plain Transformer and a linear output projection to form a clean-image prediction. This architecture needs no latent tokenizer, but still uses a sampling trajectory.
Source: Li and He, CVPR 2026 proceedings. Figure 3, PDF p. 5. Pixel patches pass through a plain Transformer and a linear output projection to form a clean-image prediction. This architecture needs no latent tokenizer, but still uses a sampling trajectory. [42]

Figure 3 shows pixel patches, a linear embedding, Transformer blocks and a linear prediction of image patches. Time and class condition the network. There is no separate learned latent encoder/decoder in this generator. However, predicting a clean image at each time is not the same as learning a complete one-step generative map.

To make the time-direction difference explicit, call JiT’s time ss: noise is at zero and data at one. Then

ys=(1−s)ϵ+sx,v^θ=x^θ(ys,s)−ys1−s,∥v^θ−(x−ϵ)∥2=∥x^θ−x∥2(1−s)2.y_s=(1-s)\epsilon+s x,\qquad \widehat v_\theta=\frac{\widehat x_\theta(y_s,s)-y_s}{1-s},\qquad \|\widehat v_\theta-(x-\epsilon)\|^2=\frac{\|\widehat x_\theta-x\|^2}{(1-s)^2}.

Here xx is clean data, ϵ\epsilon is prior noise, ysy_s is the interpolated input, x^θ\widehat x_\theta is the direct network output, and v^θ\widehat v_\theta is the derived instantaneous velocity. The equation holds for s<1s\lt1 before numerical safeguards. In MeanFlow’s convention, substitute t=1−st=1-s and reverse the velocity sign. The solver uses velocity even though the network predicts an image; the velocity loss is a time-weighted image error, not an unweighted image MSE.

Evidence

Table 2 crosses three direct outputs with three loss spaces. For JiT-B/16 on ImageNet 256×256, 200 epochs and CFG, the velocity-loss row gives FID 8.62 with image output, 372.38 with noise output and 96.53 with velocity output. The smaller 64×64, patch-4 setting does not show the same decisive separation. This controls the loss while revealing a capacity-sensitive prediction-space effect. Figure 4 separately shows that a bottleneck patch embedding can remain effective with image prediction.

Design analysis

A patch’s scalar dimension is different from the number of patches. Increasing image resolution and patch size together can hold token count fixed while increasing the information per token. The paper’s synthetic embedding experiment and image ablations support its manifold explanation under finite network capacity. They do not prove that all clean-image predictors universally outperform every noise predictor. Adjusting loss weights or noise sampling alone also does not reproduce the architectural comparison.

Reading checkpoint

Name three distinct quantities: the direct output, the loss target and the quantity consumed by the ODE solver. Then explain why a single clean prediction from highly noisy input may be an average over possibilities rather than a sharp sample from the image distribution.

Limits and next step

JiT targets class-conditioned generation and still uses numerical sampling; it is not inherently a one-step or freeform text-to-image model. Its no-tokenizer recipe also should not be conflated with every later pixel generator’s use of pretrained losses. Continue to Pixel MeanFlow, which adds a two-time denoised-image field and an interval objective to support one-call pixel generation.

Continue reading: Pixel MeanFlow — one-step pixels through a denoised field

References

  1. L. Yang, Z. Zhang, Y. Song, S. Hong, R. Xu, Y. Zhao, W. Zhang, B. Cui and M.-H. Yang, “Diffusion Models: A Comprehensive Survey of Methods and Applications,” ACM Computing Surveys, 2023. arXiv
  2. Y. Lipman, M. Havasi, P. Holderrieth, N. Shaul, M. Le, B. Karrer, R. T. Q. Chen, D. Lopez-Paz, H. Ben-Hamu and I. Gat, “Flow Matching Guide and Code,” arXiv:2412.06264, 2024. arXiv
  3. J. Ho, A. Jain and P. Abbeel, “Denoising Diffusion Probabilistic Models,” NeurIPS, 2020. arXiv
  4. Y. Song, J. Sohl-Dickstein, D. P. Kingma, A. Kumar, S. Ermon and B. Poole, “Score-Based Generative Modeling through Stochastic Differential Equations,” ICLR, 2021. arXiv
  5. J. Song, C. Meng and S. Ermon, “Denoising Diffusion Implicit Models,” ICLR, 2021. arXiv
  6. J. Ho and T. Salimans, “Classifier-Free Diffusion Guidance,” NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications; arXiv:2207.12598, 2022. arXiv
  7. R. Rombach, A. Blattmann, D. Lorenz, P. Esser and B. Ommer, “High-Resolution Image Synthesis with Latent Diffusion Models,” CVPR, 2022. arXiv
  8. W. Peebles and S. Xie, “Scalable Diffusion Models with Transformers,” ICCV, 2023. arXiv
  9. P. Esser, S. Kulal, A. Blattmann et al., “Scaling Rectified Flow Transformers for High-Resolution Image Synthesis,” ICML, PMLR 235, 2024. PMLR
  10. Y. Lipman, R. T. Q. Chen, H. Ben-Hamu, M. Nickel and M. Le, “Flow Matching for Generative Modeling,” ICLR, 2023. arXiv
  11. X. Liu, C. Gong and Q. Liu, “Flow Straight and Fast: Learning to Generate and Transfer Data with Rectified Flow,” ICLR, 2023. arXiv
  12. Y. Song, P. Dhariwal, M. Chen and I. Sutskever, “Consistency Models,” ICML, 2023. arXiv
  13. Z. Geng, M. Deng, X. Bai, J. Z. Kolter and K. He, “Mean Flows for One-step Generative Modeling,” NeurIPS, 2025. arXiv
  14. S. J. D. Prince, Understanding Deep Learning, MIT Press, 2023. book site
  15. J. Sohl-Dickstein, E. Weiss, N. Maheswaranathan and S. Ganguli, “Deep Unsupervised Learning using Nonequilibrium Thermodynamics,” ICML, PMLR 37, 2015. PMLR
  16. P. Vincent, “A Connection Between Score Matching and Denoising Autoencoders,” Neural Computation, vol. 23, no. 7, pp. 1661–1674, 2011. doi
  17. Y. Song and S. Ermon, “Generative Modeling by Estimating Gradients of the Data Distribution,” NeurIPS, 2019. arXiv
  18. P. Dhariwal and A. Nichol, “Diffusion Models Beat GANs on Image Synthesis,” NeurIPS, 2021. NeurIPS
  19. A. Q. Nichol and P. Dhariwal, “Improved Denoising Diffusion Probabilistic Models,” ICML, PMLR 139, 2021. PMLR
  20. T. Karras, M. Aittala, T. Aila and S. Laine, “Elucidating the Design Space of Diffusion-Based Generative Models,” NeurIPS, 2022. arXiv
  21. C. Lu, Y. Zhou, F. Bao, J. Chen, C. Li and J. Zhu, “DPM-Solver: A Fast ODE Solver for Diffusion Probabilistic Model Sampling in Around 10 Steps,” NeurIPS, 2022. arXiv
  22. T. Salimans and J. Ho, “Progressive Distillation for Fast Sampling of Diffusion Models,” ICLR, 2022. arXiv
  23. M. Heusel, H. Ramsauer, T. Unterthiner, B. Nessler and S. Hochreiter, “GANs Trained by a Two Time-Scale Update Rule Converge to a Local Nash Equilibrium,” NeurIPS (NIPS), 2017. arXiv
  24. T. Kynkäänniemi, T. Karras, S. Laine, J. Lehtinen and T. Aila, “Improved Precision and Recall Metric for Assessing Generative Models,” NeurIPS, 2019. arXiv
  25. G. Stein, J. C. Cresswell, R. Hosseinzadeh et al., “Exposing Flaws of Generative Model Evaluation Metrics and Their Unfair Treatment of Diffusion Models,” NeurIPS, 2023. arXiv
  26. S. Jayasumana, S. Ramalingam, A. Veit, D. Glasner, A. Chakrabarti and S. Kumar, “Rethinking FID: Towards a Better Evaluation Metric for Image Generation,” CVPR, 2024. CVF
  27. F.-A. Croitoru, V. Hondru, R. T. Ionescu and M. Shah, “Diffusion Models in Vision: A Survey,” IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. doi
  28. H. Cao, C. Tan, Z. Gao, Y. Xu, G. Chen, P.-A. Heng and S. Z. Li, “A Survey on Generative Diffusion Models,” IEEE Transactions on Knowledge and Data Engineering, vol. 36, no. 7, pp. 2814–2830, 2024. doi
  29. C.-H. Lai, Y. Song, D. Kim, Y. Mitsufuji and S. Ermon, “The Principles of Diffusion Models,” monograph, arXiv:2510.21890, 2025. arXiv
  30. H. Shen, J. Zhang, B. Xiong et al., “Efficient Diffusion Models: A Survey,” Transactions on Machine Learning Research, 2025. arXiv
  31. A. Melnik, M. Ljubljanac, C. Lu, Q. Yan, W. Ren and H. Ritter, “Video Diffusion Models: A Survey,” Transactions on Machine Learning Research, 2024. arXiv
  32. C. Lu and Y. Song, “Simplifying, Stabilizing and Scaling Continuous-Time Consistency Models,” ICLR, 2025. arXiv
  33. A. Sauer, D. Lorenz, A. Blattmann and R. Rombach, “Adversarial Diffusion Distillation,” ECCV, 2024. arXiv
  34. M. S. Albergo and E. Vanden-Eijnden, “Building Normalizing Flows with Stochastic Interpolants,” ICLR, 2023. arXiv
  35. J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li and L. Fei-Fei, “ImageNet: A Large-Scale Hierarchical Image Database,” CVPR, 2009, pp. 248–255. doi
  36. C. Szegedy, V. Vanhoucke, S. Ioffe, J. Shlens and Z. Wojna, “Rethinking the Inception Architecture for Computer Vision,” CVPR, 2016, pp. 2818–2826. arXiv
  37. O. Ronneberger, P. Fischer and T. Brox, “U-Net: Convolutional Networks for Biomedical Image Segmentation,” MICCAI, LNCS 9351, pp. 234–241, 2015. arXiv
  38. A. Radford, J. W. Kim, C. Hallacy et al., “Learning Transferable Visual Models From Natural Language Supervision,” ICML, 2021. arXiv
  39. M. Oquab, T. Darcet, T. Moutakanni et al., “DINOv2: Learning Robust Visual Features without Supervision,” Transactions on Machine Learning Research, 2024. arXiv
  40. B. Efron, “Tweedie’s Formula and Selection Bias,” Journal of the American Statistical Association, vol. 106, no. 496, pp. 1602–1614, 2011. doi
  41. A. Dosovitskiy, L. Beyer, A. Kolesnikov et al., “An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale,” ICLR, 2021. arXiv
  42. T. Li and K. He, “Back to Basics: Let Denoising Generative Models Denoise,” CVPR, 2026 (preprint 2025). paper, publication record, official JiT code.
  43. B. Zheng, N. Ma, S. Tong and S. Xie, “Diffusion Transformers with Representation Autoencoders,” ICLR, 2026 (preprint 2025). proceedings, manuscript.
  44. S. Tong, B. Zheng, Z. Wang, B. Tang, N. Ma, E. Brown, J. Yang, R. Fergus, Y. LeCun and S. Xie, “Scaling Text-to-Image Diffusion Transformers with Representation Autoencoders,” CVPR Findings Track, 2026. manuscript, author publication record.
  45. D. Podell, Z. English, K. Lacey, A. Blattmann, T. Dockhorn, J. Müller, J. Penna and R. Rombach, “SDXL: Improving Latent Diffusion Models for High-Resolution Image Synthesis,” ICLR, 2024. proceedings, official model card.
  46. Black Forest Labs, “FLUX.1 [dev],” official pretrained-model card, 2024. model card.
  47. W. Kong, Q. Tian, Z. Zhang et al., “HunyuanVideo: A Systematic Framework For Large Video Generative Models,” technical report, arXiv:2412.03603, 2024. report, official release.
  48. Team Wan, A. Wang, B. Ai, B. Wen et al., “Wan: Open and Advanced Large-Scale Video Generative Models,” technical report, arXiv:2503.20314, 2025. report, official Wan2.1 release.
  49. Z. Geng, Y. Lu, Z. Wu, E. Shechtman, J. Z. Kolter and K. He, “Improved Mean Flows: On the Challenges of Fastforward Generative Models,” CVPR, pp. 30467–30476, 2026. CVF
  50. Y. Lu, S. Lu, Q. Sun, H. Zhao, Z. Jiang, X. Wang, T. Li, Z. Geng and K. He, “One-step Latent-free Image Generation with Pixel Mean Flows,” ICML, PMLR 306, pp. 81265–81277, 2026. PMLR