Bedovyy/Anima-FP8

FP8 Quantized model of ANIMA

!! Currently FP8, MXFP8 and NVFP4 doesn't work properly with torch.compile, so it is better to use original BF16 model. !!

There are two models - FP8 and NVFP4Mixed.

• FP8 (2.4GB) : (recommend) maximize generation speed while preserving quality as much as possible.
• NVFP4Mixed (2.0GB): (marginal quality) Mixture of FP8 and NVFP4.

To use torch.compile, use the TorchCompileModelAdvanced node from KJNodes, set the mode to max-autotune-no-cudagraphs, and make sure dynamic is set to false.

Generation speed

Tested on

• RTX5090 (400W), ComfyUI with --fast option, torch2.10.0+cu130
• Generates 832x1216, 30steps, cfg 4.0, er sde, simple

quant	none	sage+torch.compile
bf16	7.13s/4.21it/s	5.16s/5.81it/s (+38%)
fp8	6.66s/4.50it/s (+11%)	4.52s/6.64it/s (+58%)
nvfp4mix	6.37s/4.71it/s (+12%)	4.99s/6.01it/s (+43%)

Sample

anima-preview3-base

anima-preview2

anima-preview

quant	sample
bf16
fp8
nvfp4mixed

Quantized layers

fp8

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

nvfp4mixed

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