ComfyUI v0.46.0: Native support for PAN 2

ComfyUI released v0.46.0 on 11 18. Regarding open-source models, MBZUAI's PAN 2 world model (the Comfy-Org packaged version was introduced in previous issues) has been integrated into the core and can be loaded directly using built-in loading nodes. A new Video Deflicker node contributed by Kijai has been added, which uses RAFT optical flow to align flickering frames processed frame-by-frame with adjacent frames, while also switching to a faster native RAFT implementation. Additionally, there is a Split Image Channels node that separates images by red, green, blue, and alpha channels. Other improvements include SenseNova support for thinking and interleaved generation, Generate Text support for Qwen3-1.7B, and Save Image Advanced capability to save single-channel PNGs. DirectML support has been removed, and the --directml parameter is no longer available.

ComfyUI released v0.46.0 on 11 18. Regarding open-source models, MBZUAI's PAN 2 world model (the Comfy-Org packaged version was introduced in previous issues) has been integrated into the core and can be loaded directly using built-in loading nodes. A new Video Deflicker node contributed by Kijai has been added, which uses RAFT optical flow to align flickering frames processed frame-by-frame with adjacent frames, while also switching to a faster native RAFT implementation. Additionally, there is a Split Image Channels node that separates images by red, green, blue, and alpha channels. Other improvements include SenseNova support for thinking and interleaved generation, Generate Text support for Qwen3-1.7B, and Save Image Advanced capability to save single-channel PNGs. DirectML support has been removed, and the --directml parameter is no longer available.

Why it matters Videos that have been upscaled or modified frame-by-frame often suffer from flickering; the new node uses optical flow to align each frame with its neighbors. Note that this version also removes DirectML, so be cautious before upgrading.

License: Conditional commercial use · ComfyUI: Native support · VRAM: VRAM requirements not specified

Reasons

  • ComfyUI official changelog; the deflicker node and PAN 2 support come from core PRs #17754 and #17755 respectively
  • Both features use only core nodes and do not require installing custom nodes

Usage

  • Deflicker: Update to v0.46.0, download torchvision's raft_large.pth and place it in ComfyUI/models/optical_flow, then connect Load Optical Flow Model and Video Deflicker (video category); connect flickering frames to images and clean pre-processing frames to guide
  • PAN 2: Download the diffusion model, Qwen3.5-9B text encoder, and Wan 2.2 VAE from Comfy-Org/PAN2, and load them using the built-in UNETLoader and CLIPLoader (select type pan2)

Risks

  • DirectML has been removed, and the changelog recommends switching to CUDA or CPU; Windows users running on DirectML should not upgrade yet
  • PAN 2 requires prompts written as structured JSON descriptions; according to the ComfyUI Wiki, single-sentence prompts may produce noise; license terms for each model are subject to their respective repositories

Open

ComfyUI 官方更新日志

Explore the latest AI Radar updates