Seedance 2.0 Prompt Enhancer (Cloud / Local GGUF)

Seedance 2.0 is a closed, paid video model, and like most of them it's extremely literal about its prompt grammar. Edit a clip and it wants @Video 1 references; generate between two frames and it wants first-frame plus last-frame wired in; describe an eight-shot sequence and it wants 镜头N ordering, not absolute timestamps. Get the intent phrasing wrong and the API charges you anyway. This node takes a sentence of what you want plus whatever media you've got, and returns a single enhanced_prompt STRING written to Seedance 2.0's official task vocabulary.

It's the second of the three core enhancers in T8mars' pack, sharing the exact same channel plumbing as the H3 node - the author's seedance.nz gateway by default, the t8star workshop, any OpenAI-compatible endpoint, or a fully offline local Qwen GGUF. The prompt protocols are kept completely separate, though: nothing H3-flavored leaks in here.

How it works

The node turns your intent into one of the official task types via task_intent: AUTO, T2V (text-to-video), I2V (first frame), FL-I2V (first + last frame), multimodal reference generation, video editing, video extension, track filling (bridging 2–3 clips), or combined tasks (edit one video while referencing another). In AUTO it reads your prompt and what's connected and picks. The media sockets follow the task - first_frame and last_frame for the frame tasks, reference_images and reference_videos for anything borrowing a subject or camera move.

The rest is Seedance-idiom control. complexity_mode decides between one tight paragraph (simple jobs) and explicit 镜头1 / 镜头2 / 镜头3 sequences (complex ones). reference_syntax picks the official Volcengine @图片N/@视频N style or Seedance.nz's English @Image N/@Video N. subtitle_policy and stability_constraints are exactly what they sound like, both AUTO by default. shot_count goes AUTO or a fixed 1–20, though high counts are a soft constraint - the upstream model may merge or drop shots and the node won't fail you for it. duration_seconds only shapes content density; nothing here fabricates per-shot second marks.

The inputs that matter

On a first pass you touch prompt (the only required field), task_intent (leave AUTO), and the media sockets for whatever you're doing. If the frames can't show a relationship or role, reference_roles lets you write things like @图片1=人物外观；@视频1=动作和运镜. reference_template migrates structure and rhythm only - never characters, props or plot. api_key here is again the prompt-enhancement LLM's key, not your Seedance 2.0 generation key.

Wire the enhanced_prompt output into your downstream Seedance 2.0 video node's prompt input, and you're done.

The audio boundary, stated plainly

Seedance 2.0 the model can handle audio. This node cannot analyze audio files, and the README is refreshingly blunt about why: a real capability probe on 2026-08-05 showed the default vision model rejecting OpenAI-compatible audio input outright ("audio input is not supported by this model"). So there's no AUDIO socket, and neither the cloud nor the local path claims to have heard your track. What you can do is describe dialogue, ambience and music in text, or keep text-only @音频N intent markers - they're understood as text, which is still useful.

Install and the gotchas

Same pack, same steps - Manager search "MiniMax H3 / Seedance 2.0 / Music 3 Prompt Enhancer (T8)", or:

cd ComfyUI/custom_nodes
git clone https://github.com/T8mars/comfyui-minimax-h3-prompt-enhancer-T8.git

Restart, Ctrl+F5. Cloud needs no extra deps; local mode pulls a ~16 GiB Qwen GGUF plus projector via python install_local_qwen.py and wants ~24 GB VRAM.

• Don't click "save to workflow" on the key before sharing a JSON.
• Cloud video uploads cap at 50 MB; local sampling is 2 fps visual frames only.
• Editing and combined tasks default to @视频1 as the edit target - if a different clip is the victim, say so in reference_roles.
• Gateway 5xx/Cloudflare errors retry up to three times automatically; 401s and empty balances don't.