ComfyUI 0.38 adds a 480p draft render for Seedance 2.5

The release also adds identity-preserving video-to-video on Wan 2.1 and VACE models, a new quantization format, and removes the retired Sora nodes.

ComfyUI v0.38.0 shipped on September 29, and for anyone making video in it the release is mostly about one thing: getting to a usable result without paying for a full-resolution render you are going to throw away.

The changelog lists a new Draft mode for ByteDance's Seedance 2.5 that renders a fast 480p preview whose draft_task_id then drives the 1080p final render. That is a two-pass workflow inside one node: iterate cheaply at low resolution, lock the shot, then spend the money once.

What else changes for video

The bigger capability addition is ID-V2V, short for identity-preserving video-to-video, supported on Wan 2.1 and VACE-based models. Video-to-video workflows restyle or re-render existing footage; the failure mode is that a face or a character drifts away from the source across the clip. ID-V2V is ComfyUI's answer to that drift, aimed at the same problem character consistency has been on this beat all month.

Two partner-node changes land with it. Seedream 5.0 Flash arrives on the Seedream node as a faster, lower-cost option, and the Pro tier now reaches 4.62MP. The OpenAI Chat node picks up two cheaper GPT-6 tiers, Sol and Luna.

And one line matters more than its length: the Sora video nodes are gone. OpenAI shut the Sora 2 video API down on September 24, so the nodes that called it were dead weight, and ComfyUI removed them rather than leave them failing in old workflows.

The model support that is not video

For image work, v0.38.0 adds Ming Image 0.1 Design, a 6-billion-parameter text-to-image model aimed at text-heavy design layouts with transparent RGBA output, plus a Union Fun ControlNet for Qwen-Image 2.1 editing and a tiny VAE for Qwen-Image 2.1 that speeds up previews and cuts memory use. MiniMax-H3 gets a Fun ControlNet Union 2.0 that combines several control conditions in one H3 generation.

Two node-level additions are worth knowing about even if you never touch the API. w6a8 is a new quantization format ComfyUI now supports, and model files can declare which attention implementation each block uses.

What got faster and what got fixed

The performance list is longer than usual. Async weight offload now runs on Ascend NPU devices, ACE-Step 1.5 and Yue2 pick up faster autoregressive generation, SeedVR2's video upscaling is optimized, fast disk detection extends to every model loader, and asset scans no longer block the interface on large libraries. The torchaudio dependency is gone, which makes for a lighter install.

Four bugs are called out: ImageUpscaleWithModel no longer crashes on RGBA images and keeps transparency through the upscale, Qwen VL preprocessing stops crashing on four-channel RGBA reference images, a MiniMax-H3 VAE rms_rope crash on offloaded qk_norm_scale is fixed, and Convert Image Color Space now handles negative values in HDR output correctly.

How to get it

ComfyUI updates the usual way: pull the repository or use the desktop build's updater, then restart. Bundled workflow templates moved to v0.11.70 in the same release, so a template that looked wrong before this update is worth reloading before you rebuild it by hand.

The Seedance 2.5 Draft mode is the part to test first if you are working on shot coverage, since it changes the cost of being wrong about a take. The Sora node removal is the part to check if you have a workflow from earlier in September that has quietly kept failing.

ComfyUI v0.38.0 is out now; the changelog is at docs.comfy.org/changelog.