attashe/Bernini-Wan2.2-fp8-scaled

Bernini (full) Wan2.2 renderer — fp8 e4m3 scaled (ComfyUI)

The two DiT renderer transformers of the full ByteDance Bernini pipeline (diff_dec = high-noise expert, diff_dec_low = low-noise expert), quantized to fp8 e4m3 scaled in the ComfyUI format.

The layout is byte-for-byte structurally identical to Comfy-Org/Bernini-R's wan2.2_bernini_r_*_fp8_scaled.safetensors (verified: same 1815 keys, shapes, dtypes, and __metadata__) — the difference is only the weights, which here are the full Bernini renderer (jointly trained with the MLLM planner) rather than the renderer-only Bernini-R.

Files

File	model_type	size
wan2.2_bernini_high_noise_fp8_scaled.safetensors	bernini_high	~15.5 GB
wan2.2_bernini_low_noise_fp8_scaled.safetensors	bernini_low	~15.5 GB

Drop them into ComfyUI/models/diffusion_models/ and use them anywhere the Bernini-R fp8_scaled files work (same model_type, same keys).

Quantization details

• Format marker per quantized weight: comfy_quant = {"format": "float8_e4m3fn"}.
• Quantized Linears: self_attn.{q,k,v,o}, cross_attn.{q,k,v} (cross-attn o kept in fp16), ffn.0, ffn.2 — 9 per block × 40 = 360 weights per expert.
• For each quantized weight W: scale = max(|W|)/448, W_fp8 = (W/scale).clamp(±448).to(float8_e4m3fn), stored alongside a scalar weight_scale (fp32). Dequant: W ≈ W_fp8.to(dtype) * weight_scale.
• Everything else (norms, modulation, patch_embedding, text/time_embedding, time_projection, head, all biases) is kept in fp16.
• Mean per-tensor reconstruction error ≈ 2.2%.
• Source: extracted from ByteDance/Bernini-Diffusers (bernini/ checkpoint, fp32), with diffusers WanTransformer3DModel keys remapped to the original Wan / ComfyUI naming.

License: Apache-2.0, inherited from the upstream Bernini release.