🌱 Seedance 2.0 Omni Reference

Seedance 2.0 Omni Reference is the flagship of this pack and the one node that does everything: text plus up to nine reference images, three reference videos, and three reference audio clips, all mixed into a single generation. If you've been building reference-heavy workflows on the other generators and hitting their ceilings - "I need a video reference and audio" - this is where you land.

The mental model is a single prompt with typed placeholders. You write @image1 for images, @video1/@video2/@video3 for video references, and @audio1/@audio2/@audio3 for audio, and the node resolves those tokens against what you've wired into the matching ports. The default prompt shows the pattern: "A person @image1 walking on the beach at sunset, cinematic lighting."

The inputs that matter

Same generation core as the others - prompt, aspect_ratio, resolution (endpoint-dependent for 1080p/4k), duration 4–15s, seed, batch_count - plus the multi-modal extras:

• image_1–image_9 - IMAGE tensors from Load Image, referenced as @image1–@image9.
• video_ref_1–video_ref_3 - the SEEDANCE2_VIDEO_REF type, fed from the S3 upload/browse nodes or Video Reference URL. Referenced as @video1–@video3. This is the key rule: BytePlus does not accept local video files in the request - a private file sitting in ComfyUI/input/ will not work. You must upload it to S3 first and pass the resulting video_ref. The pack actively errors out on legacy local-video inputs to stop you from trying.
• delete_s3_references_after_generation - when true, the node deletes connected S3 video_ref objects after a successful generation. Handy for cleanup, but note it only triggers on success.
• audio_file_1–audio_file_3 - dropdowns of mp3/wav files in ComfyUI/input/.
• audio_url_1–audio_url_3 - a public http(s) URL, an asset:// ID, a data:audio/... URL, or an absolute local path that gets Base64-encoded.

BytePlus's audio limits are worth memorizing: 2–15 seconds per clip, max 3 clips, total ≤ 15 seconds, each file ≤ 15 MB. Clips outside those bounds are the most common cause of an Omni failure.

How it works

Images are uploaded and encoded; video refs are resolved to fetchable URLs (or error out if local); audio is validated and encoded; then everything is assembled into one task payload and submitted through the standard BytePlus cycle. Outputs: video_url, first_frame (IMAGE), request_id, batch_json.

The trap to plan around

Real-person reference video is the sharpest edge. Community reports on BytePlus's Seedance API are consistent: reference footage containing identifiable real people frequently trips InputImageSensitiveContentDetected.PrivacyInformation - the safety filter treats recognizable faces, tattoos, name tags, even license plates as privacy risk. The practical workaround is anonymized or synthetic subjects in your references. It's ByteDance's moderation, running on their servers, and no node setting gets around it.

Install: ComfyUI Manager → Install via Git URL → https://github.com/nnnnkatsu/seedance2-comfyui-byteplus, restart. The S3 path needs boto3, which the pack's requirements.txt installs. And remember the ~24h URL expiry on anything it outputs - save the keepers.