Seedance 2.5 · Reference to Video

Reference-to-video is the node that solves the problem every local workflow has been hacking around for three years: keeping a character or a product consistent across shots. Instead of training a LoRA, wiring IP-Adapter and praying, you hand the model up to 30 images of what should stay the same and up to 10 clips of how it should move, and it renders a clip that carries those references forward. No weights, no per-character training run - and no illusion of control, either. You get what the model decides to keep.

The mechanism

This is PopNow's node for ByteDance's bytedance/seedance-2.5/reference-to-video model. The three reference inputs are autogrow socket groups: each image or clip you connect adds a new socket, so a 12-image character sheet is twelve connections and no extra nodes. On queue, everything is uploaded, the prompt plus references go to the model, and the finished MP4 comes back decoded as a ComfyUI VIDEO. Uploads are content-hash cached for 12 hours, so iteration doesn't re-upload the batch.

Rates start at 11.07 credits per second - cheaper than 2.5's text-to-video, because references do part of the work the model would otherwise invent.

What to set, and the limits that bite

• reference_image_urls - up to 30 images. Characters, style, scene composition. The tooltip adds a rule people miss: it can't be combined with a first-frame image. This node has no first-frame socket, so you're fine here; just don't expect to also pin your opening frame.
• reference_video_urls - up to 10 clips, each 2–30 seconds, 30 seconds total. Motion and pacing. Reference footage is added to the billed duration, so ten clips costs you on both ends: once as reference seconds, once as output.
• reference_audio_urls - up to 10 clips, same 2–30s / 30s-total limits. Generated soundtrack follows them - the tooltip names beat matching and lip sync, which is the most useful thing on this node you probably haven't tried.
• duration_seconds - 4 to 30, integer. The billed quantity.
• resolution - 480p, 720p, 1080p. No 4K tier on 2.5.
• aspect_ratio - 21:9 through 9:16, or adaptive.
• output_format - mp4 or mov. MOV if it's going into an editor that prefers ProRes-ish pipelines; otherwise mp4.
• watermark - off by default. Leave it off unless you want the visible "AI generated" mark.
• generate_audio, enable_web_search, seed - audio on by default, web search off, seed fixed.

Outputs

Three: video, last frame (an IMAGE, only when return_last_frame is on - otherwise the socket blocks downstream with an error that reads like a bug and isn't), and task id for the API console.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/popnow-ai/popnow-comfyui

Or search PopNow in Manager. Restart, then add your key from popnow.ai/api/console/keys via Settings > PopNow or POPNOW_API_KEY, and restart again so the node list rebuilds from your catalog. Zero pip dependencies - the pack declares none.

The honest problems

• Reference capacity is not reference fidelity. Thirty images is not thirty times better. Two clean, well-lit, consistent-angle images usually beat twenty near-duplicates, and conflicting references (different faces, different lighting) give you a mushy average.
• Audio references cost seconds twice. Easy to forget, expensive to discover.
• PopNow 40004 - an unsupported combination; the message names the field. Most often it's a total reference length over the 30-second budget.
• PopNow 40201 - credits gone. At 11+ credits per second with references on top, a 30-second test run isn't a test run.
• Cancel ≠ stop. The job continues on PopNow and bills if it succeeds. Use Settings > PopNow > Max credits per run as the real guard.

Where this beats the local route is narrow and real: a character you need in three shots this afternoon, and no time to train anything. Where it loses is everything reproducible - no LoRA, no seed stability across prompts, no "same character forever". If you're building a series, this is a fast previs tool, not a pipeline.