Skip to content
Other

Sample-Aligned Decoders (SAD)

drop-in decoders for image generators

2 min readupdated

A collection of pretrained decoders from researchers at the Chinese Academy of Sciences and Tsinghua University that replace the default autoencoder decoder (VAE — the module that turns a compressed image representation into pixels) in latent generative models. Each file contains only the decoder weights (decoder.* and post_quant_conv.*), fine-tuned with Sample-Aligned Decoders on the latents of a specific generator. The encoder, latent space, and generator itself remain unchanged, keeping inference speed and cost identical. Built for researchers and developers working with DiT, SiT, and PixArt-α.

Key features

  • Act as a drop-in replacement — swap the decoder in place without retraining the base generator.
  • Add zero inference overhead — the architecture and compute cost stay identical to the original.
  • Support popular generator families: DiT-XL/2, SiT-XL/2 + REPA, LightningDiT-XL, REPA-E, and PixArt-α (at 256px resolution).
  • Include repository variants trained via GAN discrepancy in DINOv2 feature space (sad-gan) and via MMD (sad-mmd).

What you can do

  • Swap the decoder in diffusers pipelines for DiT-XL/2 or PixArt-α 256px models using the load_sad_decoder helper or standard load_state_dict.
  • Load the lightningdit-xl-vavae-256 and repa-e-e2evae-256 weights in the LightningDiT and REPA-E codebases for VA-VAE and E2E-VAE autoencoders.
  • Train your own SAD decoder tailored to your specific generator and sampler following the instructions in the project repository.

Example

from diffusers import DiTPipeline
from sad import load_sad_decoder

pipe = DiTPipeline.from_pretrained("facebook/DiT-XL-2-256").to("cuda")
load_sad_decoder(pipe.vae, "dit-xl2-256/sad-gan")
images = pipe(class_labels=[279, 388], num_inference_steps=250, guidance_scale=1.5).images
Read moreDetails
Training: each decoder was fine-tuned for 2,000 steps (~1 GPU-hour) using a GAN loss in DINOv2 space or MMD, combined with an L1 encoder anchor.
Limitations: each checkpoint is tied to its specific generator, tokenizer (SD-VAE ft-EMA, VA-VAE, E2E-VAE), and sampler.
Compatibility: weights for VA-VAE and E2E-VAE are not directly compatible with diffusers and require classes from LightningDiT and REPA-E.
Discuss Telegram 0

Will it run?

Runs in BF16

Fits without quantisation.

  • BF16weights 0.2 GB · ready files0.7 GB
  • FP8 / INT8recommendedweights 0.1 GB0.6 GB
  • 4-bitweights 0.1 GB0.6 GB

Tip: go 8-bit.FP8 / INT8 is within about a percent of full precision in quality, takes half the memory and runs faster on RTX 40/50 and H100. Full precision only pays off for fine-tuning.

What to get for this model

Massed ComputeOpen Massed Compute

Your own GPU: FP8 / INT8 — from RTX 2060 · BF16 — from RTX 2060

Massed Compute hourly prices, November 2026. The discounted price applies with the coupon at checkout.

Estimate: weights plus working memory. Long videos, high resolutions and long contexts need more.

Run it on Massed Compute

The script downloads the model to the rented machine. How to run it is in the model card.

  1. 1. Rent a machine with a fitting GPU

    Open Massed Compute
  2. 2. Paste the script into the machine's terminal

    Connect over SSH (the Massed Compute panel shows the command) or open a terminal on the machine's desktop.

    HF_TOKEN / CIVITAI_TOKENThe script starts with token lines: paste yours between the quotes. The Hugging Face token is for gated models, the Civitai one for Civitai models.

    # torpedo2003/SAD
    mkdir -p ~/models
    curl -LsSf https://astral.sh/uv/install.sh | sh
    source "$HOME/.local/bin/env"
    uv tool install "huggingface_hub[cli]"
    hf download torpedo2003/SAD --local-dir ~/models/SAD
    # The model is in ~/models. How to run it is in the model card (link above).
    Add more LoRAs to SAD
  3. 3. Use it

    The model is in ~/models. How to run it is in the model card (link above).

Stop the machine in the panel when you are done: it bills by the minute while it runs.

How to speed it up

Picked for this model type. Choose your hardware above to see what fits you exactly.

  • FlashAttention 44.0.0b40from November 18, 2026new release

    fastest exact attention on B200

    Kernels for Blackwell (beta). Exact attention without quantisation — fine for everything, LLMs included.

    How to enable: pip install flash-attn-4; vLLM and SGLang pick it up themselves.

    Earlier generations (2)
    • FlashAttention 3 — The Hopper (H100/H200) version with FP8.
    • FlashAttention 2 — The standard for RTX 30/40/50 and A100: exact attention without the memory cost of long contexts.
  • torch.compile2.15.0from October 28, 2026new release

    10–30% for free

    Compiles the model for your GPU. Slower first run, faster afterwards.

    How to enable: model = torch.compile(model); in ComfyUI the TorchCompileModel node.

Model family

New posts about versions of SAD (quants, finetunes, LoRAs) will show up under the bell.

No known relatives of this model yet. All derivatives on Hugging Face

About the model

torpedo2003/SAD
Author
torpedo2003
Category
Other
Released
Posted

On NeyroLearn

Hugging Face

00

Ask AI about this model

Answered by Gemini 4 · the chat is kept only in this tab

Comments

Other

mars-dfd-hr

mars-dfd-hr