Kling 视频生成

Kling is one of the marquee closed video models (Kuaishou's), and this node front-ends the biggest slice of it: text-to-video across the standard/pro/turbo/4k/O3 tiers, image-to-video with an optional end frame, and O3 reference-to-video using up to four images. It's the "make the video from scratch" sibling to the pack's Kling_Edit_Video node, and for most Kling work it's the one you'll reach for first.

It's an API wrapper from T8mars/ComfyUI_Seedance, so: key from api.seedance.nz, no GPU, requests only.

How it works

Pick a model and the node adapts its input slots. T2V models want just the prompt. I2V models grow image1 (required) and image2 (optional end frame). O3 R2V models use image1 through image4, submitted in the order you connected them. The ratio you choose is passed through as metadata.ratio unless it's adaptive, and negative_prompt goes to metadata too. From there it's the pack's usual submit → poll → download → decode loop, with the resilient MP4 download chain everyone leans on.

The inputs that matter

• model - the long dropdown: t2v and i2v variants of kling-v3.0 std/pro, v3-turbo std/pro, v3-4k, and o3 std/pro/4k, plus the three o3-r2v models.
• prompt - up to 20480 characters.
• seconds - 5 or 10.
• ratio - adaptive or a fixed aspect.
• negative_prompt - optional, forwarded to metadata.
• image1 (i2v, required), image2 (optional end frame), image1..4 (o3-r2v, in connected order).

Outputs: video, video_url, task_id, response.

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, key from api.seedance.nz/console, Seedance API Config into api_config.

Common issues - read this before picking an O3 r2v model

The README is refreshingly honest about what doesn't work yet. As of the last real test pass:

• kling-o3-std-r2v and kling-o3-pro-r2v fail at submit with an upstream 502 - only kling-o3-4k-r2v has actually run end-to-end. If you pick std/pro r2v and it dies on submission, you're not doing anything wrong; it's the gateway's channel.
• Motion, kling-elements-advanced and lip-sync are not exposed as nodes at all - they need special structures or multi-step flows the pack deliberately skips.

Otherwise: i2v with no image1 fails validation (it's required), and if a r2v's references aren't holding character, the O3-4k path is the one with a known-good result to compare against. Kling is also one of the models where the "cheap drafts first" advice matters - 4k jobs are the priciest in the pack, so validate your prompt on 5-second std before spending on the big one.