Sign up

rzgar/wan2.2_i2v_fp8_e4m3fn_scaled_hq

🤗 On Hugging Faceimage-to-videoapache-2.046 GBother✓ Checksum-verifiedupdated 0d ago
Magnet

Wan2.2 I2V — HQ Hybrid Scaled FP8

_Left: This model | Right: BF16 | Fixed seed, Same LightX2V and prompt_

Your browser does not support the video tag.

Quantized from: Wan2.2-I2V-A14B-FP32-ComfyUI. High-fidelity hybrid scaled FP8.

  • FFN -> float8_e4m3fn + per-tensor scale_weight (MSE-tuned)
  • All attention -> motion / identity preserved
  • patch_embedding -> FP32

attention is not quantized. Better I2V fidelity.