并发提交｜Seedance 2.5 Standard 视频生成（6 合 1）

Seedance 2.5 is the pack's namesake model line - the current Standard generation - and this is its concurrent-submit wrapper. Six models: domestic and global- overseas text-to-video, image-to-video, and the multimodal one. This is also the heavyweight of the pack: the Multi model accepts up to 30 images, 10 videos, and 10 audios as references in a single job. When you need the kitchen sink, this is where it lives.

Wrapper mechanics are pack-standard: validate parameters up front, push the job to a background video thread pool, return a SEEDANCE_VIDEO_FUTURE. Feed it into SeedanceConcurrent_Video_Await (up to ten video futures), and it returns finished VIDEO values in slot order. Running three different durations of the same Multi job side by side? Three of these, one receiver.

The inputs that matter

• model - seedance-2.5-standard-{t2v,i2v,multi} plus the three -global- variants.
• prompt - required for T2V and Multi; optional for I2V. Multi references material in the prompt with @Image 1, @Video 1, @Audio 1.
• seconds - 4 to 30, or -1 to let the model pick. That -1 gets submitted as metadata.duration=-1.
• resolution - 480p through 4k, plus the native1080p preset (2.5 Standard's native output). Note it's native1080p only - no native4k on this line.
• ratio - the seven usual options, adaptive default.
• generate_audio - voice-over/sound effects on or off.
• image1…image30, video1…video10, audio1…audio10 - the Multi reference bank. I2V uses image1 as the required first frame and image2 as an optional last frame; Multi takes the full spread, with the caveat that single reference clips must be 2–30s and combined reference media can't exceed 30s total.
• seed - -1 = random (that's what gets forwarded when native), non-negative = fixed.

Plus api_config and skip_error (placeholder video on failure). The cache seed applies too: fixed reuses cached results so an unchanged re-run doesn't spend another paid call.

How it runs

Standard wrapper machinery: reference media gets uploaded and folded into metadata.content, then submit → poll → download with the MP4 integrity check. The Multi path with big reference sets is the slow one to start - the README warns the API rate-limits uploads per token and the pack retries automatically, so a 30-image job can sit in upload for a while before generation even begins.

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

• Multi reference limits - it's not just 30 images; total reference media across all types is capped at 50 items and combined audio/video length at 30s. Exceed it and validation bounces the job.
• native4k doesn't exist here - 2.5 Standard supports native1080p only; pick 4k (an upscaled output tier) if you want more.
• @Image numbering - tags follow connected slot order and gaps get compacted; the console prints a renumber hint when it happens.
• Slow batch starts on big Multi jobs - upload throttling, not a hang. Give it a minute before you suspect a bug.

Per-clip metering and data-leaves-the-machine caveats as always - and with 50 reference assets per job, that data leaving is a lot of data.