--- library_name: pytorch tags: - super-resolution - diffusion - pixel-diffusion-decoder - vae-decoder pipeline_tag: image-to-image base_model: - nvidia/PixelDiT-1300M-1024px - Tongyi-MAI/Z-Image - black-forest-labs/FLUX.1-dev - black-forest-labs/FLUX.2-dev - nyu-visionx/Scale-RAE-Qwen7B_DiT9.8B --- # PiD — Pixel Diffusion Decoder

PiD teaser

**[Paper](https://arxiv.org/abs/2605.23902), [Project Page](https://research.nvidia.com/labs/sil/projects/pid/)** [Yifan Lu](https://yifanlu0227.github.io/), [Qi Wu](https://wilsoncernwq.github.io/), [Jay Zhangjie Wu](https://zhangjiewu.github.io/), [Zian Wang](https://www.cs.toronto.edu/~zianwang/), [Huan Ling](https://www.cs.toronto.edu/~linghuan/), [Sanja Fidler](https://www.cs.utoronto.ca/~fidler/), [Xuanchi Ren](https://xuanchiren.com/)
PiD reformulates the latent-to-pixel decoder as a conditional pixel-space diffusion model, unifying decoding and upsampling into a single generative module. It denoises directly in high-resolution pixel space and produces a super-resolved image in one pass. This repository hosts the released decoder checkpoints, plus the encoder/decoder ("VAE") weights they depend on. All `PiD_*` checkpoints in this repo are **4-step distilled**. The non-`PiD_*` entries (`ae.safetensors`, `flux2_ae.safetensors`, `sdxl_vae.safetensors`, `QwenImage_VAE_2d.pth`, `sd3_vae/`, `rae/`, `scale_rae/`) are **the corresponding encoder/decoder VAE weights** that PiD plugs into — they're not PiD checkpoints themselves. ### License/Terms of Use This model is released under the [NSCLv1](https://huggingface.co/nvidia/PixelDiT-1300M-1024px/blob/main/LICENSE) License. The work and any derivative works may only be used for non-commercial (research or evaluation) purposes. ### Deployment Geography: Global ## PiD checkpoints Two variants are released for each diffusers-style backbone: - **`2k`** — trained at 2048px, used as a 4× decoder (512 LDM → 2048 px), or as an 8× decoder for the Scale-RAE backbone (256 → 2048). - **`2kto4k`** — trained with multi-resolution data bucketing 2048→4096 and an SD3-style dynamic shift; designed for 1024 LDM → 4K (4096 px) decoding. Both checkpoint variants support multiple aspect ratios. | Path | Latent space | SR factor | Variant | |-----------------------------------------------------------------|--------------|-----------|---------| | `checkpoints/PiD_res2k_sr4x_official_flux_distill_4step` | Flux1-dev | 4× | 2k | | `checkpoints/PiD_res2k_sr4x_official_flux2_distill_4step` | Flux2-dev | 4× | 2k | | `checkpoints/PiD_res2k_sr4x_official_sd3_distill_4step` | SD3 medium | 4× | 2k | | `checkpoints/PiD_res2k_sr4x_official_dinov2_distill_4step` | DINOv2-B | 4× | 2k | | `checkpoints/PiD_res2k_sr8x_official_siglip_distill_4step` | SigLIP-2 | 8× | 2k | | `checkpoints/PiD_res2kto4k_sr4x_official_flux_distill_4step` | Flux1-dev | 4× | 2kto4k | | `checkpoints/PiD_res2kto4k_sr4x_official_flux2_distill_4step_2606` | Flux2-dev | 4× | 2kto4k | | `checkpoints/PiD_res2kto4k_sr4x_official_sd3_distill_4step` | SD3 medium | 4× | 2kto4k | | `checkpoints/PiD_res2kto4k_sr4x_official_sdxl_distill_4step` | SDXL | 4× | 2kto4k | | `checkpoints/PiD_res2kto4k_sr4x_official_qwenimage_distill_4step` | Qwen-Image | 4× | 2kto4k | Each directory contains a single file, `model_ema_bf16.pth`, which is the EMA weights cast to bfloat16 — the format the inference scripts load by default. > **⚠️ Flux2-dev `2kto4k` — use the new `_2606` checkpoint.** The previous > `PiD_res2kto4k_sr4x_official_flux2_distill_4step` (without the `_2606` suffix) > suffered from a color-drifting issue. The new > `PiD_res2kto4k_sr4x_official_flux2_distill_4step_2606` fixes it — please use it > and do **not** use the old one. See the > [comparison](https://github.com/nv-tlabs/pid/blob/main/docs/FLUX2_2kto4k_new_ckpt_compare.md) > for details. ### Latent space → compatible LDMs A PiD decoder is tied to a *latent space*, not to a single generative model. Any LDM that produces latents in that space can reuse the same checkpoint. The `--backbone` aliases below pick the right LDM pipeline; they all decode through the latent space's checkpoint above. | Latent space | VAE / vision encoder weights | compatible `--backbone` | Corresponding LDM Links | |--------------|------------------------------------|-------------------------------------------|-----------------| | Flux1-dev | `checkpoints/ae.safetensors` | `flux`, `zimage`, `zimage-turbo` | [FLUX.1-dev](https://huggingface.co/black-forest-labs/FLUX.1-dev), [Z-Image](https://huggingface.co/Tongyi-MAI/Z-Image), [Z-Image-Turbo](https://huggingface.co/Tongyi-MAI/Z-Image-Turbo) | | Flux2-dev | `checkpoints/flux2_ae.safetensors` | `flux2`, `flux2-klein-4b`, `flux2-klein-9b` | [FLUX.2-dev](https://huggingface.co/black-forest-labs/FLUX.2-dev), [FLUX.2-klein-4B](https://huggingface.co/black-forest-labs/FLUX.2-klein-4B), [FLUX.2-klein-9B](https://huggingface.co/black-forest-labs/FLUX.2-klein-9B) | | SD3 medium | `checkpoints/sd3_vae/` | `sd3` | [SD3-medium](https://huggingface.co/stabilityai/stable-diffusion-3-medium-diffusers) | | SDXL | `checkpoints/sdxl_vae.safetensors` | `sdxl` | [SDXL-base-1.0](https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0) | | Qwen-Image | `checkpoints/QwenImage_VAE_2d.pth` | `qwenimage`, `qwenimage-2512` | [Qwen-Image](https://huggingface.co/Qwen/Qwen-Image), [Qwen-Image-2512](https://huggingface.co/Qwen/Qwen-Image-2512) | | DINOv2-B | `checkpoints/rae/` | `dinov2` | [RAE](https://github.com/bytetriper/RAE) (class-conditional; DINOv2-B) | | SigLIP-2 | `checkpoints/scale_rae/` | `siglip` | [Scale-RAE](https://github.com/ZitengWangNYU/Scale-RAE) (text-conditional; nyu-visionx/Scale-RAE-Qwen1.5B_DiT2.4B) | For example, Z-Image and Z-Image-Turbo share Flux1-dev's VAE, so they reuse the `flux` checkpoints (both `2k` and `2kto4k`) — no separate `zimage` checkpoint is shipped. Likewise `qwenimage-2512` reuses the `qwenimage` decoder (same VAE, different transformer). ## Usage The decoder checkpoints are loaded by the inference scripts in the [PiD codebase](https://github.com/nv-tlabs/pid). The exact `(backbone, ckpt_type) → path` mapping is the single source of truth in [`pid/_src/inference/checkpoint_registry.py`](https://github.com/nv-tlabs/PiD/blob/main/pid/_src/inference/checkpoint_registry.py) — clone the repo, point it at this snapshot, and the demos pick the right file automatically: ```bash # Pull just the checkpoints/ tree into the repo root (skips this README and # the teaser figure so they don't clobber the files in the source repo). hf download nvidia/PiD --local-dir . --include "checkpoints/*" # Then run any of the demos, e.g.: PYTHONPATH=. python -m pid._src.inference.from_ldm --backbone flux \ --prompt "A photorealistic half-body portrait of a brown tabby cat with bold stripes sitting attentively on a rustic wooden kitchen table, soft morning light streaming sideways through a large window, fine fur detail and stripe patterns sharply visible, intense amber-green eyes in razor-sharp focus, warm farmhouse kitchen softly out of focus, cinematic shallow depth of field, ultra-detailed fur texture, photorealistic" \ --ldm_inference_steps 28 --save_xt_steps 24 \ --output_dir ./results/official_demo/flux \ --pid_inference_steps 4 ``` Pick the `2kto4k` variant via `--pid_ckpt_type 2kto4k` when decoding at 4K. ## Citation ``` @article{lu2026pid, title={PiD: Fast and High-Resolution Latent Decoding with Pixel Diffusion}, author={Lu, Yifan and Wu, Qi and Wu, Jay Zhangjie and Wang, Zian and Ling, Huan and Fidler, Sanja and Ren, Xuanchi}, journal={arXiv preprint arXiv:2605.23902}, year={2026} } ```