Doubao-Seedance连续视频生成

Doubao-Seedance连续视频生成 is the pack's answer to the hardest problem in API video generation: getting more than one scene out of a model that only makes short clips. Instead of generating a single shot, it takes a list of scene descriptions, generates them in sequence, and - crucially - makes each scene a continuation of the last, so the story flows instead of jumping. It's the "tell me a longer story" node.

How it works

You give it a base_prompt that sets the world (characters, setting, style) and a prompts_text box where each line is one scene - the default demo is a girl and a fox, three scenes, and it reads exactly like a storyboard:

女孩抱着狐狸，女孩睁开眼，温柔地看向镜头
女孩和狐狸在草地上奔跑，阳光明媚
女孩和狐狸坐在树下休息，女孩轻轻抚摸狐狸

The node generates the first clip with a text-to-video model (first_video_model, default the lite t2v), then each subsequent scene is generated image-to-video using the last frame of the previous clip as its reference (subsequent_video_model, default the lite i2v). That hand-off is what gives you continuity - the next scene starts from where the last one ended. video_count sets how many scenes to run (1–10). If merge_videos is on (default), the clips get stitched into a single video; merge_videos false gives you separate clips instead.

The rest of the controls are the usual Seedance set: duration (3s–12s), resolution (480p–1080p), aspect_ratio, fps, camera_fixed, watermark, and seed. There's also an initial_image input if you want the whole chain to start from a specific image instead of pure text.

Outputs: first_video and AFVIDEO (the first clip as VIDEO), all_video_urls (a list of the per-scene clip URLs), response_text, video_info, and merged_video (the stitched result when merge is on). Wire the merged output into a video-save node and you have a multi-scene short.

Install & setup

Same pack, one install:

cd ComfyUI/custom_nodes
git clone https://github.com/xuchenxu168/ComfyUI_doubao_seed
cd ComfyUI_doubao_seed
pip install -r requirements.txt

Restart, configure your key in SeedReam4_config.json, and pick a mirror.

Gotchas

Cost multiplies here: video_count scenes means video_count separate billed video generations (plus a merge step). A 10-scene run is a serious bill and a genuinely long wait - the README's "5–15 minutes" per clip applies per scene, so budget accordingly. Second, the continuity trick depends on the i2v hand-off holding the character, which it does most of the time; when a scene drifts, it's usually the base_prompt that was too vague. Spell out the character and setting in base_prompt even if it feels repetitive, and keep merge_videos on so the seam isn't your problem. Third, same moderation and metering warnings as the rest of the pack - this is closed-model API generation, so if a scene fails instantly, suspect the prompt filter before your config.