arXiv 2026
Latent diffusion needs no decoder. Pixel diffusion gets photorealistic texture.
Latent diffusion models rely on a decoder to map denoised latents to pixels, adding complexity and inference cost. We introduce Decoding Without a Decoder (DWD), which lets a finetuned diffusion transformer produce final pixels directly through a linear head. Unlike previous methods, where the denoising output leads to final pixels, DWD separates the two roles: the original denoising head continues to drive sampling, then the new DWD head predicts the pixels at the final step. We extend pretrained latent models with DWD efficiently by finetuning only the last few transformer blocks, training the DWD head with pixel quality losses while anchoring the denoising head to the pretrained model to preserve the sample trajectory. Across latent image and video generators, DWD removes the decoder and reduces inference cost while maintaining or even improving pixel quality. On Wan video generation, it visibly reduces detail distortion and flickering, lowering warping error by 23%. Remarkably, the same separation also improves pixel diffusion models that do not even use a decoder: a final-step DWD head produces substantially better texture detail in text-to-image generation, and improves ImageNet FID from 1.55 to 1.43, a new state of the art among pixel flow models.
A linear head on a finetuned DiT matches or beats the decoder on every model we tested: SiT on ImageNet, FLUX.2 klein (32-step and 4-step), and Wan2.1 14B, where it lowers warping error by 23%.
A separate output head with its own pixel losses restores fine texture in pixel-space text-to-image generation, and sets a new state of the art among pixel flow models on ImageNet.
A latent diffusion transformer (DiT) denoises latents, and a separate decoder network turns the final latent into pixels. Yet the DiT's last-layer tokens already encode the pixels: a linear probe on them reconstructs images more accurately than one on the denoised latents.
Last-layer tokens support more accurate linear pixel reconstruction. We fit patch-wise linear probes from the denoised latents and from the last-layer tokens to ground-truth pixels, with the pretrained models frozen. (a) Held-out relative RMSE; percentages show the error reduction over the latent probe. (b, c) Probes of last-layer tokens keep finer image detail and reduce video ghosting, before any finetuning.
DWD therefore adds a linear DWD head next to the original denoising head, both reading the same last-layer tokens. The denoising head still drives the sampler. At the final step, the DWD head maps each token directly to a pixel patch, and the decoder is no longer needed.
We finetune only the last few transformer blocks. The two heads get decoupled losses: the DWD head is trained with pixel losses (L1, LPIPS, and adversarial), while the denoising head is anchored to the frozen pretrained model (L2 loss), so the same seed still gives nearly the same content.
(a) A standard latent generator maps the denoised latent to pixels with a decoder. (b) DWD keeps the denoising head for sampling and adds a parallel linear head that predicts the final pixels at the final step. Only the last blocks and the heads are finetuned.
We remove the decoder from every latent model we tested, from ImageNet SiT models to the Wan2.1 14B video model and FLUX.2 klein. Quality is the same or better in every case. On ImageNet, DWD lowers FID for both REPA (1.38 → 1.36) and REPA-E (1.12 → 1.06) while removing the decoder.
| Method | Decoder | GFLOPs (final step) | FID↓ | IS↑ |
|---|---|---|---|---|
| SiT-XL/2-REPA | SD-VAE | 2× 237.2 + 622.2 | 1.38 | 306.2 |
| DWD | – | 237.7 | 1.36 (−0.02) | 312.5 (+6.3) |
| SiT-XL/2-REPA-E | E2E-VA-VAE | 2× 237.3 + 252.2 | 1.12 | 302.9 |
| DWD | – | 237.7 | 1.06 (−0.06) | 303.1 (+0.2) |
Shaded rows are DWD, finetuned from the model above them. GFLOPs are for the final sampling step; with guidance the DiT runs twice, while the decoder runs once.
On Wan2.1, the decoder distorts fine detail such as lettering, mesh, and dense lights, and the distortion changes from frame to frame, which shows up as flicker. DWD keeps these details stable and lowers warping error by 23%, and by 29% under 8-bit inference.
| Method | Params (B) (+ decoder) | PFLOPs (final step) | Warping error↓ | Image metrics | VBench dimensions | |||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| HPSv3↑ | HPS v2.1↑ | Subject consist.↑ | Backgr. consist.↑ | Temporal flicker.↑ | Motion smooth.↑ | Dynamic degree↑ | Aesthetic quality↑ | Imaging quality↑ | ||||
| Wan2.1 14B | 14.29 + 0.07 | 2× 1.68 + 0.28 | 1.52 | 9.06 | 0.266 | 0.951 | 0.981 | 0.990 | 0.983 | 0.736 | 0.682 | 0.677 |
| DWD | 14.49 | 1.68 | 1.18 | 9.04 | 0.266 | 0.952 | 0.982 | 0.993 | 0.984 | 0.711 | 0.685 | 0.671 |
| finetuned decoder | 14.29 + 0.07 | 2× 1.68 + 0.28 | 1.39 | 9.03 | 0.266 | 0.951 | 0.982 | 0.992 | 0.983 | 0.731 | 0.681 | 0.676 |
| Wan2.1 14B (8-bit) | 14.29 + 0.07 | 2× 1.68 + 0.28 | 1.69 | 8.99 | 0.265 | 0.950 | 0.979 | 0.990 | 0.982 | 0.728 | 0.675 | 0.676 |
| DWD | 14.49 | 1.68 | 1.20 | 9.01 | 0.267 | 0.951 | 0.982 | 0.993 | 0.984 | 0.703 | 0.683 | 0.668 |
Shaded rows are DWD, finetuned from the model above them. Warping error is ×10−3. Finetuned decoder keeps the original DiT and finetunes its decoder on the same data with the same pixel losses as DWD. The 8-bit rows use FP8 W8A8 quantization with 8-bit SageAttention2.
On FLUX.2 klein 9B, both the 32-step Base model and the 4-step distilled model, DWD's output quality is visually indistinguishable from the decoder's, text rendering included. FID, patch FID, OneIG, and LongText stay on par or slightly better.
| Method | TFLOPs (final step) | COCO-10k | OneIG↑ | LongText↑ | |||
|---|---|---|---|---|---|---|---|
| FID↓ | pFID↓ | HPSv3↑ | HPSv2.1↑ | ||||
| FLUX.2 klein Base 9B | 2× 75.51 + 10.47 | 15.13 | 18.69 | 10.54 | 0.284 | 0.556 | 0.909 |
| DWD | 75.54 | 15.05 | 18.38 | 10.54 | 0.284 | 0.556 | 0.908 |
| FLUX.2 klein 9B (4-step) | 75.51 + 10.47 | 21.50 | 25.52 | 12.23 | 0.307 | 0.535 | 0.867 |
| DWD | 75.54 | 21.92 | 25.33 | 12.15 | 0.307 | 0.537 | 0.871 |
Shaded rows are DWD, finetuned from the model above them. The 4-step model is guidance-distilled, so its DiT runs once per step.
AsymFlow and L2P also remove the decoder, but they turn a latent model into a pixel model, which changes its sampling trajectory and its samples. Both fall short of their base models in text rendering. DWD keeps the latent trajectory nearly intact, so its samples stay close to the base model's for the same seed, at about a third of AsymFlow's training cost.
| Method | FID↓ | pFID↓ | OneIG↑ | LongText↑ | Training cost (H100 GPU hours) |
|---|---|---|---|---|---|
| FLUX.2 klein Base 9B | 15.13 | 18.69 | 0.556 | 0.909 | – |
| AsymFlow | 18.10 | 22.62 | 0.506 | 0.777 | 1100 (15k iters) |
| DWD | 15.05 | 18.38 | 0.556 | 0.908 | 350 (26k iters) |
| Z-Image Turbo | – | – | 0.528 | 0.917 | – |
| L2P | – | – | 0.499 | 0.787 | – |
Indented rows are finetuned from the model above them. DWD and AsymFlow start from FLUX.2 klein Base 9B; L2P starts from Z-Image Turbo, so its images above come from a different base model. All images share the same seed. Z-Image Turbo scores are from its official report.
DWD also powers pi-H3, our few-step MiniMax H3 audio-video model, where removing the video VAE decoder makes generation much faster. This is still preliminary: the video below comes from an early experimental checkpoint, and improved checkpoints will be released soon.
Read the blog post →
Pixel diffusion models have no decoder to remove, yet they gain the most from DWD. Their denoising head is trained only with the L2 diffusion loss, which oversmooths fine texture. A separate DWD head trained with pixel losses fixes this: on ImageNet, it improves JiT-H/16 from 1.90 to 1.85 FID and AnisoFlow-H/16 from 1.55 to 1.43, a new state of the art among pixel flow models. Separation is the key: putting the same pixel losses on a single shared head makes FID worse (1.55 → 1.71).
| Method | GFLOPs (final step) | FID↓ | IS↑ |
|---|---|---|---|
| JiT-H/16 | 2× 363.9 | 1.90 | 300.9 |
| DWD | 363.9 | 1.85 (−0.05) | 305.7 (+4.8) |
| w/o separate head | 363.9 | 1.91 (+0.01) | 302.4 (+1.5) |
| AnisoFlow-H/16 | 2× 364.8 | 1.55 | 300.1 |
| DWD | 363.9 | 1.43 (−0.12) | 303.5 (+3.4) |
| w/o separate head | 364.8 | 1.71 (+0.16) | 302.9 (+2.8) |
Shaded rows are DWD, finetuned from the model above them. W/o separate head applies the same pixel losses and trajectory anchoring to a single shared head, which hurts FID. 1.43 FID is a new state of the art among pixel flow models.
On AnisoFlow-L/16-T2I, the DWD head restores skin, hair, foliage, and fabric texture while keeping the composition, lowering patch FID from 36.59 to 32.26.
| Method | TFLOPs (final step) | COCO-10k | OneIG↑ | LongText↑ | |||
|---|---|---|---|---|---|---|---|
| FID↓ | pFID↓ | HPSv3↑ | HPSv2.1↑ | ||||
| AnisoFlow-L/16-T2I | 2× 5.99 | 28.86 | 36.59 | 10.72 | 0.297 | 0.528 | 0.850 |
| DWD | 5.98 | 27.54 | 32.26 | 10.63 | 0.301 | 0.533 | 0.852 |
Shaded rows are DWD, finetuned from the model above them. Patch FID (pFID) is computed on image crops and is most sensitive to fine detail.
@article{chen2026decoding,
title={Decoding Without a Decoder: Direct Pixel Prediction from Diffusion Transformers},
author={Hansheng Chen and Shengqu Cai and Jan Ackermann and Rayhan Zirvi and Gordon Wetzstein and Xinchen Yan and Leonidas Guibas},
journal={arXiv preprint arXiv:2611.22287},
url={https://arxiv.org/abs/2611.22287},
year={2026},
}