media-gen-vision-video

# Media Generation, Vision, and Video ## Choose the right path - **Image generation or editing**: use the preferred Nano Banana 2 / Gemini image workflow. - **Image understanding / screenshot analysis**: use Gemini multimodal image understanding. - **Video generation**: use Google Veo 3.1. ## Non-negotiables - Prefer Google-native media models and official flows first. - Preserve aspect ratio, resolution, style, and reference-image constraints. - Do not guess image contents when a multimodal path is available. - Do not claim video generation succeeded unless a real video file was produced. - When delivering files, send the generated asset directly into the conversation when supported. - For successful image or video generation, always deliver the actual media asset to the chat; do not stop at a summary or path when direct sending is available. ## Image generation and editing - Use the preferred image tool path first. - For edits, keep the user’s reference image identity intact unless explicitly asked to change it. - If the user specifies size or ratio, honor it exactly when possible. - If the task asks for multiple variants, generate a small set rather than one-by-one loops. ## Image understanding - Use multimodal analysis for screenshots, photos, and UI inspection. - Report only what is visible or strongly supported. - Separate confirmed observations from inference. - If the image is unreadable or only partially visible, say so plainly. ## Video generation - Default to Veo 3.1. - Prefer the official Gemini API workflow when possible. - Save the final file with a stable filename before sharing it. - If video output is not available in the current environment, say that clearly and identify the blocker. - Do not substitute a still image or text summary for an actual video file unless the user accepts that fallback. ## Delivery and reporting - Return the generated asset when available. - If the user asks for a file, do not bury it in prose—attach or send it directly. - Keep the response short: result, file/path if any, and blockers if any.