Seedance 2.0 Mini · Reference to Video

This carries the lowest per-second badge in the Seedance 2.0 family - at the time of writing it started around 1.65 credits ($0.0165) per second. That's genuinely cheap for reference-conditioned video with sound. It's also the number people misread, so let's deal with the asterisk first.

Reference-to-video takes footage as input, and reference footage is billed on a linear basis against clip length. The rate is lower because you're supplying a lot of the information; the total is not automatically lower, because the references carry their own cost. Short, pointed references keep both the quality and the invoice sane.

What "reference" means here

Three autogrow sockets, one per modality. Connect an image and you get reference_image_0; connect a second and the node grows another. Batch input fans out into multiple references. The same goes for reference_video_urls and reference_audio_urls.

What each one buys you: images carry character, style and composition; video carries motion and pacing; audio makes the generated soundtrack follow it. That's a wider control surface than most hosted video nodes give you, and it's the reason this node is worth learning rather than just using image-to-video.

Mechanically it's the same client as every other node in the pack - upload (cached 12 hours per key), quote, submit under an idempotency key, poll, download, decode. Video and audio are re-encoded on the way up; the ceiling is 90 MiB.

The inputs that matter

prompt - subject, action, camera, light, sound, constraints. With references in play, write the relationships: what comes from which reference, and what must not change.

reference_image_urls / reference_video_urls / reference_audio_urls - the source material, as described.

duration_seconds - 4 to 15, or -1 for smart duration, where the model chooses. Billed quantity, as always.

resolution - 480p or 720p. Mini doesn't go higher; if the reference-driven shot is the final, plan to promote it to Fast.

aspect_ratio - adaptive by default, following your references, plus the usual set.

enable_web_search - web lookup before render, billed at the web-search rate. Pointless on most reference work; leave it off.

generate_audio - on by default. bitrate_mode, watermark - delivery cosmetics. seed - fixed by default, so the same graph re-queued gives the same clip.

return_last_frame - turn it on if you'll chain; the last frame output blocks downstream until you do.

Outputs

video, last frame (IMAGE, gated by the switch), task id.

Installing it

ComfyUI Manager → PopNow → Install → restart. Or:

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

Create a key at popnow.ai/api/console/keys, restart, paste it under Settings > PopNow (or export POPNOW_API_KEY), restart once more for the catalog. No downloads, no dependency wrangling - the pack leans entirely on what ComfyUI ships.

Where people get burned

Dumping everything in and hoping. Three images, a fifteen-second clip and a music track with a one-line prompt produces an average of your references, not a synthesis of them. Be specific about which element comes from where, and cut references that aren't earning their keep - they're the expensive part.

Expecting Mini's references to behave like full 2.0's. Full Seedance 2.0 accepts up to nine reference images and gives you 1080p and 4K; Mini is the 720p-capped tier, and its fidelity to a reference is proportionate to the price.

And the standard cautions: cancelling a run doesn't stop or refund the server-side job, and a content_rejected result comes from the model's own filter with no local workaround. Set a Max credits per run and check the quote rather than the badge.