Skip to content
Other

Anima Pixel Decoder

Anima DiT VAE decoder replacement

2 min readupdated

The model is a 25M-parameter decoder that reads features from the final layer of the Anima diffusion transformer (DiT) on the last sampling step and writes RGB pixels directly. This replaces the standard VAE decoder, reducing VRAM usage and speeding up the workflow. The architecture is based on the FastFrames approach using ConvNeXt blocks. The development is useful for ComfyUI and anima_lora users for faster image generation.

Key features

  • Decoding acceleration: decoding time is reduced from 141–151 ms to 7 ms.
  • VRAM savings: peak decoder memory drops from 1.22–1.38 GB to 0.21–0.24 GB.
  • Compatibility: works with the base DiT (Euler, 20–50 steps) and the 4-step turbo mode (anima-turbo-42step).
  • High quality: the PSNR (peak signal-to-noise ratio) compared to a standard VAE is 36.9 dB.

What you can do

  • Speed up generation in ComfyUI: install the Anima Pixel Decoder node, connect the pixel tap to the model before the KSampler, and use pixel decoding instead of VAE.
  • Run via anima_lora: use the --decoder pixel_head argument in the inference.py script.
  • Generate artwork faster: comfortably create 1024×1024 resolution images on graphics cards like the RTX 5070 Ti with lower memory load.

Example

Example of usage in ComfyUI: add the Anima Pixel Head Tap node to the model before the KSampler, and then decode the latent using the Anima Pixel Decode node.

Read moreDetails
The file weighs 95 MB (fp32, safetensors format).
Input: 2048-dim features per token.
Output: 16×16 RGB patch per token in the [-1, 1] range.
Limitations: screentone dots and dense hatching may be smoothed into soft gray shades (1–2.5 levels darker than VAE).
License: CircleStone Labs Non-Commercial License (commercial use of generated images is permitted).
Examples
Discuss Telegram 1

Will it run?

Runs in BF16

Fits without quantisation.

  • BF16weights 0.1 GB · ready files0.6 GB
  • FP8 / INT8recommendedweights 0.0 GB0.6 GB
  • 4-bitweights 0.0 GB0.5 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.

    # sorryhyun/anima-pixel-decoder
    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 sorryhyun/anima-pixel-decoder --local-dir ~/models/anima-pixel-decoder
    # The model is in ~/models. How to run it is in the model card (link above).
    Add more LoRAs to anima-pixel-decoder
  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

  1. Cosmos-Predict2-2B-Text2Imagefamily root
  2. Anima
  3. anima-pixel-decoder

New posts about versions of Cosmos-Predict2-2B-Text2Image (quants, finetunes, LoRAs) will show up under the bell.

Quantised versions 3All on Hugging Face
Fine-tunes 12All on Hugging Face

About the model

sorryhyun/anima-pixel-decoder
Author
sorryhyun
Category
Other
Released
Posted

On NeyroLearn

Hugging Face

01

Ask AI about this model

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

Comments