Gemini Omni 1.1 Flash: Reference to Video
Generate video that incorporates specific subjects from multiple reference images using Gemini Omni 1.1 Flash's reference mode: characters, products, and objects from your images appear in the scene while the image itself is never shown as a frame. Bind each image to a role with tags like <IMAGE_REF_0>, combine character references with style references, and direct the scene in plain language with native audio. Ideal for character-driven stories, product placement, and brand-consistent content.
About
This workflow teaches you how to steer Gemini Omni 1.1 Flash to generate a video that features specific people, products, or objects without ever cutting to the original photos. It does this by binding each loaded reference image to a role tag (for example, <IMAGE_REF_0>, <IMAGE_REF_1>) and then using those tags inside your scene description so the model knows which visual subject to synthesize into the shot. The core is the GeminiVideoOmniV2 node running in reference mode, fed by one or more LoadImage nodes for character and style/product references, and finalized by a SaveVideo node to write the result to disk. A MarkdownNote in the graph documents the prompt pattern and tag usage.
What you provide
- Reading the workflow
Runs in ComfyUI. This workflow can't run on this page. Open it in Comfy Cloud with everything loaded, or download it for your own ComfyUI.
Try in CloudQuestions & answers
Load each image with a LoadImage node and connect it to the GeminiVideoOmniV2 image reference inputs. In your prompt, refer to them with the exact tags used by the node (e.g., <IMAGE_REF_0>, <IMAGE_REF_1>). Immediately assign a role: "Use <IMAGE_REF_0> as the product" or "Use <IMAGE_REF_1> as the presenter." Keep those tags near the start of the prompt so the model prioritizes them.
No. In this workflow the images are used as visual references only. Gemini Omni 1.1 Flash conditions on their content to synthesize new frames; it does not insert the source photos as frames.
Yes. Provide separate LoadImage nodes for character/product and style boards, bind each to its own tag, and clarify their roles in the prompt (e.g., "<IMAGE_REF_0> character, <IMAGE_REF_1> product, <IMAGE_REF_2> style reference"). Avoid conflicting instructions (e.g., opposing lighting or wardrobe notes) to maintain fidelity.
Common causes are unclear role assignment, low-quality references, or conflicting prompt details. Put the tags at the start, describe each role plainly, use high-resolution, well-lit references where the subject is large in frame, and reduce competing style constraints. If needed, shorten the scene description and focus on essentials (location, key action, lighting).