Do I have to use the LLM, or can I write the Seedance prompt myself?

You can do either. The LLM (GeminiNode) is optional—use it to draft a Seedance 2.0–style prompt, then edit it in MarkdownNote. If you already have a strong prompt, skip the LLM and feed your text directly to ByteDance2ReferenceNode.

Why two reference images?

Using two references gives Seedance 2.0 a clearer target for subject, style, and context (e.g., front and three-quarter views, or subject plus environment). This typically improves consistency and reduces drift compared to a single reference.

The LLM isn’t running—what should I check?

Make sure GeminiNode is configured with a valid API key and internet access, and that you’re only running the LLM at first (bypass ByteDance2ReferenceNode with Ctrl+B). If the request exceeds model limits, shorten your description or resize references before retrying.

How do I improve quality or stabilize motion?

Start with better references (clear subject, lighting, and style). In ByteDance2ReferenceNode, increase steps or guidance for sharper detail, and adjust reference weighting to keep the look locked in. If available, reduce resolution or frame count during exploration and scale up once you like the motion.

Workflow

Seedance 2.0 + LLM Prompt Helper

This workflow uses two reference images and a simple text description to automatically generate a detailed Seedance 2.0-style prompt using an LLM. Instead of spending time figuring out complex prompting structure, the workflow analyzes your references and builds the prompt for you, making it much faster and easier to create high-quality results. The included LLM helper is designed to match the style and formatting that works well with Seedance 2.0, helping streamline the entire process while still giving you full control over the final output. Just drop in your references, describe the scene or motion you want, and the workflow handles the prompt generation automatically.

Comfy Cloudで試すワークフローをダウンロード

作成者

enigmatic_e 1 month ago

Seedance 2.0 + LLM Prompt Helper is a ComfyUI workflow that turns two reference images and a short scene description into a detailed, Seedance-optimized prompt—then renders your video. It uses GeminiNode to analyze your references and expand your brief into the style and structure Seedance 2.0 expects, so you spend less time wrestling with prompt formatting and more time iterating on look, motion, and composition. The Seedance 2.0 generation runs through ByteDance2ReferenceNode, giving you reference-driven consistency from first frame to last.

Technically, the Reference Images group uses LoadImage to bring in two guides and ImageConcanate to combine them for quick visual context, previewed with PreviewImage. Your short description and the two images feed the LLM group, where GeminiNode produces a refined Seedance-style prompt that’s displayed via MarkdownNote for easy review and edits. That text flows to the Seedance 2.0 group: ByteDance2ReferenceNode conditions on the two references plus the generated prompt to synthesize frames, which you can inspect in PreviewAny and export with SaveVideo. The result is a streamlined loop: drop images, describe the motion, let the LLM draft the prompt, tweak if needed, and render.

FAQ

よくある質問

すべてのワークフローを見る