并发提交｜Kling 视频生成

Kling is one of the models people actually seek out rather than settle for - it's a closed model with no local door, so the API wrapper is the only way to get it inside a ComfyUI graph at all (the KB's external-api-nodes doc is explicit that "want a model with no open weights" is the first reason a workflow routes out). This node is the concurrent-submit version of the pack's Kling generator, and it's a big one: nineteen models spanning t2v, image-to-video/first-last-frame, and O3 reference-to-video, with standard, pro, turbo, and 4k tiers.

The pattern is the pack standard: set params, the node validates them synchronously, then the job runs on a background video thread pool and returns a SEEDANCE_VIDEO_FUTURE. Feed it into SeedanceConcurrent_Video_Await, which waits on up to ten futures at once and returns VIDEO outputs in slot order. Comparing a turbo t2v against a pro t2v? That's two of these nodes and one receiver.

The inputs that matter

• model - 19 choices; the t2v set takes prompt-only, the i2v set takes image1 (+ optional image2 as the last frame), and the kling-o3-*-r2v set takes up to four reference images.
• prompt - required for t2v and r2v; optional for i2v.
• seconds - 5 or 10. No -1, no 15.
• ratio - forwarded as metadata.ratio when it isn't adaptive.
• negative_prompt - optional, forwarded to metadata. Kling is one of the models in the pack that actually supports a negative prompt, so use it.
• image1…image4 - i2v uses 1–2 (first/last), r2v uses however many you connect, compacted to order.

Plus api_config, skip_error (placeholder on failure), and seed (cache seed; fixed reuses cached results so re-runs don't double-bill).

The honest warning about r2v

Read the README before you build a workflow around the O3 r2v models. Only kling-o3-4k-r2v has been verified working end-to-end; kling-o3-std-r2v and kling-o3-pro-r2v were still returning an upstream 502 at submit time when the pack was tested. That's an upstream API problem, not a bug you can fix on the node - pick the 4k r2v model or use i2v if you hit it. The motion, kling-elements-advanced, and lip-sync paths aren't exposed as nodes at all.

How it runs

Standard wrapper machinery: upload your reference images, submit, poll, download, and only pass the MP4-header check counts as a finished VIDEO. Failures isolate per slot in a batch - raise stops and names the slot, placeholder keeps the rest going.

Installing

# ComfyUI Manager: search "ComfyUI Seedance" → install → restart.
# or:
comfy node install 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

Just requests; no weights. API key via Seedance API Config, SEEDANCE_API_KEY, or config/.env.

Where people trip

• r2v 502s - use kling-o3-4k-r2v or fall back to i2v; the std/pro r2v channels were still broken at last check.
• 5/10 only - don't try 8 seconds or -1; validation will bounce it.
• Batch killers - with failure_mode=raise on the receiver, one bad prompt cancels the other pending futures. If you're iterating on prompts, switch the receiver to placeholder.

Per-clip metering and data-leaves-the-machine caveats apply, as with every node in this pack. Save Kling results you care about - those links expire.