FLUX 3 视频生成与草稿增强（8 合 1）

FLUX is Black Forest Labs' closed video line - no open weights, no local install, ever. This node is how you get FLUX 3 clips into a ComfyUI graph anyway: eight model variants behind one dropdown, covering domestic and global T2V (text-to-video), I2V (keyframe image-to-video with up to ten frames), V2V (edit an existing video), and the one that makes this node interesting - Draft Enhance. That last mode turns a cheap first-pass render into a refined final clip using a cached "draft" from the same line, which is the classic two-pass trick, just hosted.

Like the rest of T8mars/ComfyUI_Seedance, it's an API wrapper: a key from api.seedance.nz instead of a GPU, requests as the only dependency, and every frame rendered on someone else's hardware.

How it works

You pick a model, and the node's front end reshapes itself: T2V shows just prompt and settings, I2V expands the image1..image10 keyframe slots (submitted in slot order), V2V switches to a single video source (input_video or video_url, public URL wins if both are set), and Draft Enhance replaces everything with a draft_cache input. Set draft on a normal generation and the node returns a draft_cache string you can feed straight into a Draft Enhance call - the pack's example workflows wire that connection directly.

Underneath it's the standard async video flow: upload any local media, submit, poll, stream the MP4 down with header verification.

The inputs that matter

• model - the eight-way switch: flux-3-video-{t2v,i2v,v2v,draft-enhance} plus the -global- overseas variants.
• prompt - required for T2V/I2V/V2V; Draft Enhance runs off draft_cache instead.
• seconds - 5 to 20.
• resolution - hd or fhd.
• ratio - auto or one of eight documented aspect ratios (21:9, 2:1, 16:9, 4:3, 1:1, 3:4, 9:16).
• audio_mode - api_default, enabled or disabled.
• safety_tolerance - api_default or a 0 (strictest) to 4 dial.

Outputs are the standard video quartet: video, video_url, task_id, response - and draft_cache rides along when you had draft on.

Installing it

ComfyUI Manager → search "ComfyUI Seedance", or:

cd ComfyUI/custom_nodes
git clone https://github.com/T8mars/ComfyUI_Seedance.git
cd ../..
python -m pip install -r custom_nodes/ComfyUI_Seedance/requirements.txt

Restart ComfyUI, get a key from api.seedance.nz/console, connect a Seedance API Config node.

Common issues

• Draft Enhance complains about draft_cache - it must come from a FLUX 3 generation on the same line (domestic draft → domestic enhance). Mixed lines won't link up.
• Keyframes ignored - you're on T2V; I2V is the only mode that reads image1..image10.
• V2V source confusion - both input_video and video_url set? The URL wins, per the tooltip. Clear one if that's not what you meant.
• Slow start on multi-keyframe jobs - ten uploaded keyframes means ten uploads, and the gateway rate-limits uploads per token; the node waits and retries, so the graph will sit on "uploading" for a while before generation begins.

A note on expectations: this is the closed frontier via a reseller gateway, so content filtering follows the model, and your prompt and keyframes leave the machine. That's the price of a model you were never allowed to download.