Pre-trained model weights for ComfyUI-SD-CN-Animation, a ComfyUI custom node pack for temporally coherent video generation and stylization using optical flow.
Models
File
Size
Used By
Description
FloweR_0.1.2.pth
~1.2 GB
Txt2Vid
FloweR optical flow reconstruction model. Predicts motion, occlusion, and next frame from 4 previous frames.
raft-things.pth
~20 MB
Vid2Vid
RAFT optical flow estimation model. Computes dense bidirectional optical flow between video frame pairs.
Auto-Download
If you install the ComfyUI-SD-CN-Animation node pack, these weights are downloaded automatically on first use. No manual setup required.
Manual Installation
If you prefer to download manually, place the files in your ComfyUI models directory:
Generate video sequences from a text prompt. FloweR drives the motion by predicting optical flow and the next frame, while Stable Diffusion adds detail and style through a two-pass sampling pipeline (inpaint + refine). Histogram matching prevents color drift.
Vid2Vid (Video Stylization)
Stylize existing videos with any SD checkpoint. RAFT computes real optical flow between input frames, warps the previous stylized frame forward, and uses occlusion detection to determine what needs regenerating. Maintains temporal coherence across the full sequence.
Both modes support ControlNet and IPAdapter for additional control over style and structure.
Model Origins
FloweR — Optical Flow Reconstruction model by volotat, from the SD-CN-Animation project.
RAFT — Recurrent All-Pairs Field Transforms for Optical Flow by Zachary Teed and Jia Deng (Princeton Vision Lab). BSD 3-Clause License. Paper | Code
License
FloweR model weights: MIT License (from SD-CN-Animation)
RAFT model weights: BSD 3-Clause License (from Princeton)
