Seedance 2.5: Reference to Video
Generate reference-driven videos with Seedance 2.5 by supplying up to 20 images, 6 video clips, and 6 audio clips to control character consistency, visual style, and motion rhythm in a single 4–30 second output with synchronized sound and lip-synced dialogue. Ideal for e-commerce product showcases, brand advertising, and character-driven short series.
例を試す
このワークフローについて
Seedance 2.5: Reference to Video is a multimodal ComfyUI workflow that turns your curated references into a single, coherent video clip with synchronized sound. You can feed up to 20 still images for character identity and visual style, plus up to 6 video clips to guide motion, and up to 6 audio clips to drive rhythm and lip-sync. The core of the pipeline is the ByteDance2ReferenceNode, which fuses these inputs to generate a 4–30 second sequence of frames and an aligned audio track. LoadImage nodes bring your still references into the graph, and SaveVideo encodes the generated frames (and accompanying audio from the generator) into a final video file.
よくある質問
Begin with a small, curated set: 3–6 strong images for identity/style, 1–2 motion clips, and a single clean audio track. Add more only if you need additional cues. Too many overlapping references can cause conflicts and visual jitter.
Use clear, front-facing images with consistent lighting and aspect ratio, place your best identity shots early in the list, and increase the image-reference influence in ByteDance2ReferenceNode. Avoid mixing drastically different hairstyles, angles, or outfits unless you intend to change appearance.
Use a single, clean vocal track without heavy reverb or background music for best lip-sync. Ensure your target duration and FPS match your audio plan; if the audio is longer than the selected duration, trim it to fit. Re-queue with a fixed seed after adjustments for consistent comparisons.
Resolution, frame rate, duration, and the number of references all increase compute and memory use. For faster previews, lower resolution and FPS, shorten duration, and limit references. When satisfied with motion and sync, scale up quality for the final render.