ComfyUI 0.38 adds Seedance 2.5 drafts and ControlNet for MiniMax H3

ComfyUI 0.38.0 came out on 29 September. For anyone running video models locally, the headline is control over MiniMax's open H3 model; for anyone using hosted ones, it is three new partner nodes.

MiniMax H3: ControlNet

H3, MiniMax's 33-billion-parameter model for video with sound, gets Fun-ControlNet-Union 2.0: generation guided by a structure such as pose or depth taken from another clip. The release also adds tile blending to H3's VAE, so tiled decodes no longer leave seams, and fixes a crash in it.

Partner nodes: Seedance, Hunyuan, Seedream

Three hosted models arrive as partner nodes: a Draft mode for ByteDance's Seedance 2.5 video model, for quick previews before a full render; Hunyuan Image 3.5 for text-to-image and editing; and Seedream 5.0 Flash.

Colour for grading

The Convert Color Space node learns HDR LogC3 and ACEScct, and stops mishandling negative values — the plumbing that lets footage leave ComfyUI in a space a colourist works in.

Lighter and faster

ComfyUI no longer depends on torchaudio, one fewer package to match against your PyTorch build. The release adds ID-V2V support on Wan 2.1 / VACE, optimises the SeedVR2 upscaler, and adds w6a8 quantisation.

Sora's nodes are gone

The deprecated Sora nodes have been removed. OpenAI shut the Sora 2 API down on 24 September, so the nodes had nothing left to call; a workflow that still contains them will report them as missing.