arXiv 2026

Decoding Without a Decoder: Direct Pixel Prediction from Diffusion Transformers

Latent diffusion needs no decoder. Pixel diffusion gets photorealistic texture.

Hansheng Chen1,2, Shengqu Cai1, Jan Ackermann1, Rayhan Zirvi1,3, Gordon Wetzstein1, Xinchen Yan*, Leonidas Guibas1
1Stanford University 2Luma AI 3California Institute of Technology
*Work done while at Luma AI

Removing the decoder keeps quality and text rendering. Wan2.1 14B with the Wan VAE decoder vs. DWD with no decoder, then FLUX.2 klein Base 9B vs. DWD, AsymFlow, and L2P on a LongText-Bench prompt. Same prompt and seed within each model family.

Abstract

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.

No decoder

Latent diffusion needs no decoder

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%.

1.55 → 1.43 FID

Pixel diffusion recovers photorealistic texture

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.

Two Heads, Two Roles

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.

Linear probes from denoised latents vs. last-layer tokens: relative RMSE and reconstructions

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.

Standard latent diffusion generator with a decoder vs. DWD with a parallel linear head

(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.

Latent Diffusion Needs No Decoder

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.

ImageNet 256×256

MethodDecoderGFLOPs (final step)FID↓IS↑
SiT-XL/2-REPASD-VAE2× 237.2 + 622.21.38306.2
 DWD–237.71.36 (−0.02)312.5 (+6.3)
SiT-XL/2-REPA-EE2E-VA-VAE2× 237.3 + 252.21.12302.9
 DWD–237.71.06 (−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.

Text-to-video: Wan2.1 14B 832×480 · 81 frames

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.

Text-to-Video 832×480, 81 Frames (VBench)

MethodParams (B)
(+ decoder)
PFLOPs
(final step)
Warping
error↓
Image metricsVBench dimensions
HPSv3↑HPS
v2.1↑
Subject
consist.↑
Backgr.
consist.↑
Temporal
flicker.↑
Motion
smooth.↑
Dynamic
degree↑
Aesthetic
quality↑
Imaging
quality↑
Wan2.1 14B14.29 + 0.072× 1.68 + 0.281.529.060.2660.9510.9810.9900.9830.7360.6820.677
 DWD14.491.681.189.040.2660.9520.9820.9930.9840.7110.6850.671
 finetuned decoder14.29 + 0.072× 1.68 + 0.281.399.030.2660.9510.9820.9920.9830.7310.6810.676
Wan2.1 14B (8-bit)14.29 + 0.072× 1.68 + 0.281.698.990.2650.9500.9790.9900.9820.7280.6750.676
 DWD14.491.681.209.010.2670.9510.9820.9930.9840.7030.6830.668

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.

Text-to-image: FLUX.2 klein 9B 1024×1024

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.

Text-to-Image 1024×1024

MethodTFLOPs (final step)COCO-10kOneIG↑LongText↑
FID↓pFID↓HPSv3↑HPSv2.1↑
FLUX.2 klein Base 9B2× 75.51 + 10.4715.1318.6910.540.2840.5560.909
 DWD75.5415.0518.3810.540.2840.5560.908
FLUX.2 klein 9B (4-step)75.51 + 10.4721.5025.5212.230.3070.5350.867
 DWD75.5421.9225.3312.150.3070.5370.871

Shaded rows are DWD, finetuned from the model above them. The 4-step model is guidance-distilled, so its DiT runs once per step.

Versus latent-to-pixel adaptation: AsymFlow and L2P LongText-Bench · 1024×1024

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.

DWD vs. Latent-to-Pixel Adaptation (1024×1024)

MethodFID↓pFID↓OneIG↑LongText↑Training cost (H100 GPU hours)
FLUX.2 klein Base 9B15.1318.690.5560.909–
 AsymFlow18.1022.620.5060.7771100 (15k iters)
 DWD15.0518.380.5560.908350 (26k iters)
Z-Image Turbo––0.5280.917–
 L2P––0.4990.787–

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.

pi-H3 DWD – the fastest MiniMax H3 2 steps · no video VAE decoder · preliminary

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 Recovers Photorealistic Texture

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).

ImageNet 256×256

MethodGFLOPs (final step)FID↓IS↑
JiT-H/162× 363.91.90300.9
 DWD363.91.85 (−0.05)305.7 (+4.8)
 w/o separate head363.91.91 (+0.01)302.4 (+1.5)
AnisoFlow-H/162× 364.81.55300.1
 DWD363.91.43 (−0.12)303.5 (+3.4)
 w/o separate head364.81.71 (+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.

Text-to-image: AnisoFlow-L/16-T2I pixel space · 1024×1024

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.

Text-to-Image 1024×1024

MethodTFLOPs (final step)COCO-10kOneIG↑LongText↑
FID↓pFID↓HPSv3↑HPSv2.1↑
AnisoFlow-L/16-T2I2× 5.9928.8636.5910.720.2970.5280.850
 DWD5.9827.5432.2610.630.3010.5330.852

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.

Citation

BibTeX
@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},
}