Seedance 2.0 All Reference Video (8 Person + 7 BG)

First, the thing that will trip you up: the name makes it sound like a local model node, and it isn't. This is a wrapper around ByteDance's Seedance 2.0 API that runs through ComfyUI's official partner-node pipeline. No weights, no VRAM, no model download - you pay per generation, and the heavy lifting happens on ByteDance's servers. What the node actually solves is the input side: Seedance 2.0's reference endpoint accepts at most 9 reference images, and this node lets you feed it up to 8 people plus 7 backgrounds by packing the overflow into a contact sheet automatically.

That's a real niche. Multi-character consistency is the eternal nightmare of AI video - the community's answer to "how do I get two consistent people in one scene" has historically been masks, adapters, and prayer. Seedance's reference mode is one of the few things that just takes the images and keeps everyone straight. This node is the version of that for a whole cast.

How it works

Read the source and it's surprisingly tidy. Your person_1 through person_8 images go to the API as independent references. Your background_1 through background_7 do too - unless the total would blow past 9. In that case the backgrounds get stitched into a 3-column contact sheet of 512px tiles, sent as one image, and the prompt is rewritten to tell the model which images are people and which are backgrounds, plus "don't render the contact-sheet grid in the final video." The node also injects a header ("Use Seedance 2.0 all-reference mode…") and a consistency line about coherent motion and stable subjects.

Behind the scenes it uploads each reference to ByteDance's virtual library, fires a task at the official endpoint, polls every ~9 seconds, then downloads the finished video. There's a price badge on the node showing approximate USD cost, which is a nice touch - this thing can get spendy, and it tells you before you commit.

The inputs that matter

You set maybe four things as a beginner:

• model - Seedance 2.0 (quality, up to 4K), or Fast/Mini (cheaper, capped at 720p). This dropdown also carries the resolution, aspect ratio, duration (4–15s, default 7s) and audio toggle for the generation.
• person_1…8 / background_1…7 - your reference images, all optional. But at least one person or background is required; the node raises an error otherwise.
• seed - read the tooltip's fine print: it mostly controls whether the node re-runs. Seedance stays non-deterministic, so don't expect 42 to reproduce anything.
• watermark - default off. Leave it off.

Output is a single VIDEO, ready to wire into a preview, VHS, or upscale pass.

Installing it

ComfyUI Manager, search ComfyUI-Seedance2-AllRef, install, restart. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/caodidi804-bit/ComfyUI-Seedance2-AllRef

Then restart ComfyUI. There's no requirements.txt - the real dependency is comfy_api_nodes, which ships inside modern ComfyUI, so the pack is genuinely install-and-go. What it needs beyond that is a Comfy account with API credits; the node authenticates with your Comfy Org credentials, so if it sits there doing nothing, check you're signed in and funded, not your GPU.

Where people get burned

• It's an API, not a model. If you came for a local Seedance, this isn't that. The seed, the polling, the uploads all happen against ByteDance.
• The 9-image ceiling is real. Load more than 9 persons alone and you'll hit a "reduce person references to add backgrounds" error. The contact-sheet trick only stretches so far.
• Content moderation is aggressive. Seedance is ByteDance's moderated product, and people report images and prompts getting bounced as "sensitive content" through the Comfy pipeline. Don't expect the Wild West.
• Cost creeps. 1080p and 4K are dramatically pricier per second than 480p/720p. For test runs, Fast at 720p is the sensible habit.

If you're already paying for a Seedance subscription and want a proper reference workflow in the graph - whole casts included - this is the one. If you're allergic to API meters, keep walking; the open video stack is waiting.