Wan22_Bernini_FP8_Scaled

This repository contains optimized FP8 Scaled (e4m3fn) checkpoints for Bernini-R, a unified framework for video generation and editing that combines an MLLM-based semantic planner with a DiT-based renderer.

These weights are derived from ByteDance's raw transformer checkpoints (ByteDance/Bernini-R) and have been fully calculated with accurate inverse scale tensors using the optimization pipeline popularized by Kijai. This allows high-quality inference in ComfyUI with dramatically reduced VRAM overhead while fully preserving the dynamic range of the original model layers.

📦 Model Information

• Architecture: Diffusion Transformer (DiT) based on Wan 2.2
• Precision: FP8 (e4m3fn) with Custom Scaled Coefficients
• File Types Included:
  • Wan22_Bernini_HIGH_fp8_e4m3fn_scaled.safetensors (~15.6 GB) — Optimized high-noise diffusion layers.
  • Wan22_Bernini_LOW_fp8_e4m3fn_scaled.safetensors (~15.6 GB) — Optimized low-noise refining layers.

---

🛠️ ComfyUI Integration & Usage

Unlike naive FP8 conversions that truncate model data and cause color saturation artifacts, these weights include per-tensor scale matrices that native ComfyUI nodes can interpret directly.

Prerequisites

Make sure your ComfyUI architecture is fully up to date to support native Wan 2.2 scaling structures. You will need the core tracking node wrappers or advanced custom nodes (such as Kijai/ComfyUI-WanVideoWrapper) to load the High/Low handoff pipeline.

Directory Setup

Place both downloaded .safetensors files into your default ComfyUI checkpoint or diffusion model directory:

ComfyUI/models/checkpoints/
└── Wan22_Bernini_HIGH_fp8_e4m3fn_scaled.safetensors
└── Wan22_Bernini_LOW_fp8_e4m3fn_scaled.safetensors