{"lid":12,"links":[{"from":{"node":"main","port":"image"},"id":"l1","to":{"node":"survey","port":"image"}},{"from":{"node":"main","port":"image"},"id":"l2","to":{"node":"edit","port":"image"}},{"from":{"node":"refjacket","port":"image"},"id":"l3","to":{"node":"edit","port":"image2"}},{"from":{"node":"refshades","port":"image"},"id":"l4","to":{"node":"edit","port":"image3"}},{"from":{"node":"refdog","port":"image"},"id":"l5","to":{"node":"edit","port":"image4"}},{"from":{"node":"instruction","port":"text"},"id":"l6","to":{"node":"edit","port":"prompt"}},{"from":{"node":"edit","port":"image"},"id":"l7","to":{"node":"reread","port":"image"}},{"from":{"node":"instruction","port":"text"},"id":"l8","to":{"node":"j1","port":"a"}},{"from":{"node":"survey","port":"text"},"id":"l9","to":{"node":"j1","port":"b"}},{"from":{"node":"j1","port":"text"},"id":"l10","to":{"node":"j2","port":"a"}},{"from":{"node":"reread","port":"text"},"id":"l11","to":{"node":"j2","port":"b"}},{"from":{"node":"j2","port":"text"},"id":"l12","to":{"node":"checker","port":"prompt"}}],"nid":13,"nodes":[{"fields":{"color":"yellow","text":"Edit with references. One main photo, up to three reference pictures, and one instruction that names each picture — MiniMax H3-Image (the model at #1 on Hugging Face this week; its open weights are ~135 GB and its license skips the US, EU, UK and Korea, so here it just runs in the browser) fuses them into a single scene. A survey reads the subject first, and a checker at the end grades whether every piece landed and the subject survived. Swap any 📷 input for your own photos and rewrite the instruction to match."},"id":"c-intent","type":"comment","w":720,"x":-40,"y":-380},{"fields":{"image":""},"id":"main","name":"Main photo","type":"upload","w":320,"x":-40,"y":0},{"fields":{"image":""},"id":"refjacket","name":"Reference 1 · outfit","type":"upload","w":320,"x":-40,"y":330},{"fields":{"image":""},"id":"refshades","name":"Reference 2 · accessory","type":"upload","w":320,"x":-40,"y":660},{"fields":{"image":""},"id":"refdog","name":"Reference 3 · companion","type":"upload","w":320,"x":-40,"y":990},{"fields":{"model":"google/gemini-3.8-flash","q":"You are the surveyor in a photo-editing pipeline. Look at this photo and write SUBJECT ANCHORS in this exact shape:\nFace: <face shape, skin tone, distinguishing features, one line>\nHair: <colour, length, style, one line>\nOutfit now: <what the person is wearing, one line>\nPose: <stance and where the arms and legs are, one line>\nBackground: <what is behind and under the person, one line>\nPlain factual lines only. No advice, no markdown, no extra lines."},"id":"survey","name":"Subject anchors — Gemini reads the photo","type":"vision","w":340,"x":380,"y":-40},{"fields":{"model":"minimax-h3-image","prompt":"","size":"1k"},"id":"edit","name":"Fuse the references — MiniMax H3-Image","type":"edit","w":380,"x":380,"y":320},{"fields":{"text":"Picture 1 is the main photo. Keep the person in Picture 1 exactly as they are — same face, same long dark hair, same pose, same jeans and shoes, same plain grey wall behind them. Dress the person in the red leather jacket from Picture 2, worn open over the white t-shirt. Put the round gold-framed sunglasses from Picture 3 on the person's face. Sit the Dalmatian puppy from Picture 4 on the floor beside the person's legs, looking at the camera. One single photorealistic scene with soft, even light."},"id":"instruction","name":"Edit instruction (names each picture)","type":"text","w":380,"x":380,"y":700},{"fields":{"model":"google/gemini-3.8-flash","q":"You are a blind reader in a photo-editing pipeline. Describe this picture factually in this exact shape:\nPerson: <who is in the picture — face and hair as visible, one line>\nWearing: <every garment you can see, one line>\nAccessories: <anything worn on the face, head, hands or body beyond clothing, or 'none'>\nAnimals: <any animal — kind, colouring, where it sits, or 'none'>\nBackground: <setting behind the subject, one line>\nPlain factual lines only. Do not guess at anything you cannot see."},"id":"reread","name":"Blind re-read — Gemini describes the result","type":"vision","w":340,"x":860,"y":320},{"fields":{"sep":"\n\nSUBJECT ANCHORS (read from the main photo before the edit):\n"},"id":"j1","name":"Instruction + anchors","type":"join","w":300,"x":860,"y":700},{"fields":{"sep":"\n\nRE-READ (a blind reader described the edited picture; it never saw the originals):\n"},"id":"j2","name":"Everything + re-read","type":"join","w":300,"x":1240,"y":700},{"fields":{"maxTokens":"400","model":"z-ai/glm-5.3-flash","reasoningEffort":"low","system":"You are the edit checker in a multi-reference photo-editing pipeline. The input gives the EDIT INSTRUCTION that was sent to the editor, then SUBJECT ANCHORS read from the main photo before the edit, then a RE-READ a blind reader wrote from the edited picture alone. Output exactly this shape:\nLANDED: <each element the instruction asked for that the re-read confirms, comma-separated, or 'nothing landed'>\nMISSING: <each asked-for element the re-read does not confirm or contradicts, comma-separated, or 'nothing missing'>\nSUBJECT: SURVIVED if the re-read's person matches the anchors (face, hair, pose, background), otherwise DRIFTED — <one line saying what changed>\nJudge only from the three texts. If the re-read is ambiguous about an element, count it as MISSING and say 'unclear in the re-read'. No markdown, no extra lines."},"id":"checker","name":"Edit report — GLM grades the fusion","type":"llm","w":340,"x":1240,"y":320}],"v":1}