FP8/MXFP8 Quantized model of ANIMA

To use torch.compile on FP8/MXFP8 models, set to max-autotune-no-cudagraphs mode and dynamic to false.

Generation speed

Tested on

• RTX5090 (400W), ComfyUI(commit id c96fcdd) with --fast option, torch2.12.0+cu130
• Generates 832x1216, 30steps, cfg 5.0, er_sde, simple

Sample

anima-base-v1.0

anima-preview3-base

anima-preview2

anima-preview

quant	sample
bf16
fp8
nvfp4mixed

Quantized layers

quantized by comfy-dit-quantizer

fp8

{
  "format": "comfy_quant",
  "block_names": ["net.blocks."],
  "rules": [
    { "policy": "keep", "match": ["blocks.0.", "blocks.1.", "blocks.27.", "adaln_modulation"] },
    { "policy": "float8_e4m3fn", "match": ["q_proj", "k_proj", "v_proj", "output_proj", ".mlp"] },
    { "policy": "nvfp4", "match": [] }
  ]
}

nvfp4mixed

{
  "format": "comfy_quant",
  "block_names": ["net.blocks."],
  "rules": [
    { "policy": "keep", "match": ["blocks.0.", "blocks.1.", "blocks.27.", "adaln_modulation"] },
    { "policy": "mxfp8", "match": ["q_proj", "k_proj", "v_proj", "output_proj", ".mlp"] },
    { "policy": "nvfp4", "match": [] }
  ]
}