Skip to content

← The Lab

ComfyUI v0.46.0: a Video Deflicker node, 2.3x faster RAFT, mixed-frame-rate concatenation, and new Seedance, Kling and HeyGen partner nodes

ComfyUI tagged v0.46.0 at 3:55 PM MST on Nov 18. Several pieces were already merged and covered here one by one (PAN 2, HeyGen Avatar 6, the LTX VFX nodes). The tagged release is the moment they land for people on stable builds. The two local-video additions worth knowing are a Video Deflicker node and a faster native RAFT optical-flow model.

Video Deflicker (kijai, PR #17754). It's a new node for the shimmer you get after per-frame processing, like upscaling every frame separately. It uses a known blind temporal-consistency method (Bonneel et al., 2015): for each frame it computes optical flow, warps the previous cleaned frame forward, then blends toward it while staying faithful to the input. You can set strength, direction (bidirectional or forward) and a threshold. You can also compute flow from a separate guide video, for example the clean original, while deflickering the processed version.

Faster RAFT. The same PR replaces torchvision's RAFT-large in the Optical Flow Loader with a native version that loads the same weights. Per the PR it's about 2.3x faster, which also helps the VOID nodes that already use it.

Concatenate Video stops breaking on mixed clips (jtydhr88, PR #17621). Joining a 25 fps clip and a 30 fps clip used to crash Save Video or produce variable frame rate. Now every clip conforms to the first clip's frame rate by time, dropping or repeating frames as needed so durations stay right. Get Video Components also converts audio to the first clip's sample rate and channels. That matters if you're stitching outputs from different video models.

Partner (API) nodes in this cycle. Seedance 2.5 Flash added to the Seedance 2.5 nodes. HeyGen Avatar 6 nodes, plus 1080p, 30 s clips and auto duration on HeyGen Video 1.0. Kling 4.0 Image and Long Image nodes, then 4:3 and 3:4 for Kling 4.0. LTX Alpha Gen, SDR to HDR, Reframe and Upscale. MiniMax H3 Image, Bria RMBG 3.0 and Claude Fable 5.2 in the Claude node. FishAudio speech-to-text now uses transcribe-1-pro. Wan, Krea, Luma and more partner nodes moved to Comfy Router. These are paid and billed to Comfy credits.

Housekeeping that might bite. DirectML support is gone (the last PyTorch with a DirectML plugin was 2.5), so AMD-on-Windows setups that relied on it need another path. There's a new yellow warning if your GPU supports CUDA 13 and you're not on it. And there's an opt-in --force-output-rerun flag that makes output nodes write their files again on a cache hit, aimed at servers that delete outputs after each job.

Signals: official release notes only. Not HOT by our rule, but a significant stable release for anyone running ComfyUI video locally.

Related tools on ReelStack

Skills & repos in our directory

More from the Lab

Topics: ComfyUI · Seedance · Kling · Open-weight video & audio models