<pre><code>An MXFP8 copy of LyloGummy's Anima 3.8B v1.1 (https://civarchive.com/models/2880480) for ComfyUI.
5.6 GB instead of 8.8 GB, and about 1.25 times faster per image on an RTX 5080. Same bundle as the
original (the 52-block DiT plus the Semantic Connector v2); only the large DiT projections are
8-bit. Not an official CircleStone Labs or LyloGummy release.

**What was changed.** The self-attention, cross-attention and MLP projection weights of all 52 DiT
blocks (520 tensors) were converted to MXFP8: FP8 E4M3 values with one shared scale per 32
values. Everything else is unchanged bf16, including llm_adapter and the Semantic Connector v2.

**Requirements**

- An RTX 50-series GPU for the speedup. Other GPUs load the file but run it dequantized, according
  to ComfyUI's documentation (only an RTX 5080 was tested).
- ComfyUI (tested with 0.42.1).
- The GumGum10/comfyui-anima-3-8B custom node (https://github.com/GumGum10/comfyui-anima-3-8B)
  with a small patch: without it, the anima.3-8B-v2 loader stops with
  "AttributeError: 'Linear' object has no attribute 'weight'" on quantized checkpoints. The patch
  and instructions are in the GitHub repository
  (https://github.com/lepetitprince716-prog/anima-3.8b-mxfp8).
- The same text encoders and VAE as the original v1.1: qwen_3_06b_base, qwen35_4b and
  qwen_image_vae.

**Usage.** Put the file in ComfyUI/models/diffusion_models, install and patch the node, restart
ComfyUI, and load the workflow from the GitHub repository. Samples: 832×1216, 28 steps, CFG 7,
res_multistep with the beta scheduler, no LoRA.

**Speed.** RTX 5080, ComfyUI 0.42.1 with --use-sage-attention --fast fp16_accumulation, 832×1216,
28 steps, CFG 7, warm, including VAE decode: 14.00 s per image for the original, 11.19 s for this
file. With --fast fp16_accumulation ComfyUI runs the original as fp16 with fp16 accumulation, the
fastest baseline on GeForce cards; the bf16 case was not measured.

**Quality.** The same seed does not reproduce the original image; pose and framing can change.
A same-seed comparison is on the Hugging Face page
(https://huggingface.co/kefallrain/Anima-3.8B-v1.1-MXFP8).

**License.** CircleStone Labs Non-Commercial License v1.2. The model may be used for
non-commercial purposes only; generated images may be used for any purpose, including commercial
ones, except where the license says otherwise. Rights are granted directly by CircleStone Labs
LLC. The full license, attribution notice and modification statement are in the Hugging Face and
GitHub repositories. Built on NVIDIA Cosmos.

Credits: CircleStone Labs (Anima), LyloGummy (Anima 3.8B and the Semantic Connector v2), GumGum10
(ComfyUI nodes), ComfyUI and comfy-kitchen (MXFP8 support).</code></pre><p></p>

Anima 3.8B v1.1 MXFP8

<p>Recommended settings: CFG 7, steps 28, sampler res_multistep, beta scheduler, 832x1216, no LoRA.</p>

Files with SHA256: 8ada9bc806fc65ebe81519a5696bc18329eb841d0bdf66fb5cc234114f02b940