ITADN

feat: add reference image support for imagePrompt in imageParams.images

#1314Closedisamu 创建于 2026-03-20
I
isamucommented
## Problem When generating reference images via `imagePrompt` in `imageParams.images`, there is no way to reference another image as input (for image-to-image generation, style transfer, etc.). Currently: - `imagePrompt`: generates from text prompt only — no image reference - `image`: loads from source (URL/path/base64) — no generation - `moviePrompt`: supports `imageName` to reference other `imageRefs` keys — **this pattern should be extended to `imagePrompt`** ## Proposed Solution Add two new optional fields to `mulmoImagePromptMediaSchema`: ### 1. `referenceImageName` — reference another imageRefs key ```json { "imageParams": { "images": { "base_photo": { "type": "image", "source": { "kind": "path", "path": "./photo.jpg" } }, "stylized": { "type": "imagePrompt", "prompt": "Same scene in anime style", "referenceImageName": "base_photo" } } } } ``` Uses the same 2-stage resolution pattern as `moviePrompt.imageName`: 1. Stage 1: resolve `image` and non-referencing `imagePrompt` entries 2. Stage 2: resolve `imagePrompt` entries with `referenceImageName` (can reference Stage 1 results) ### 2. `referenceImage` — direct source specification ```json { "stylized": { "type": "imagePrompt", "prompt": "Same scene in anime style", "referenceImage": { "kind": "path", "path": "./photo.jpg" } } } ``` Direct file path/URL/base64, no dependency on other `images` keys. ### Resolution priority When both are specified, `referenceImageName` takes precedence (consistent with `firstFrameImageName` precedence pattern). ## Implementation Plan 1. **Schema**: add `referenceImageName` (imageIdSchema) and `referenceImage` (mediaSourceSchema) to `mulmoImagePromptMediaSchema` 2. **image_references.ts**: split `getMediaRefs` into 2 stages — first resolve non-referencing images, then resolve referencing `imagePrompt` entries 3. **generateReferenceImage**: accept optional `referenceImagePath` and pass to image generation agent as reference input 4. **Image agents**: pass reference image to OpenAI (edit_image) / Google (image reference) / Replicate APIs ## User Prompt - imageParams.imagesで画像を生成するとき(reference)、他の画像を参照する機能がない - 直接ファイルのpath or urlを指定するか、imagesでの他のpath or urlの画像を参照して生成できるようにしたい
关闭于 2026-03-21 0 条评论