Adaptive Fashion E-commerce Hero Banner Pipeline

What we're building A pipeline that composites 2-3 individually-generated model+garment shots into one shared lifestyle scene for an e-commerce hero banner built so each model's outfit is an independent swappable layer, not baked into one generation, so changing what "model 2" wears later doesn't touch anyone else in the image. Problem it solves Multi-person hero banners (the "3 models, 3 outfits, one scene" look every fashion e-commerce homepage uses) can't be generated reliably in one AI call current models garble anatomy and scale when asked to place multiple full figures together. Splitting it into individual generations plus manual composition avoids that failure mode entirely, and as a side effect gives you a genuinely reusable asset: swap one garment, regenerate one layer, done.

Workflow Steps Stage A - Individual model + garment composites (repeat per model) Step 1 - Model Reference (Image Node) Upload a neutral model photo. Label "Model Reference 1." Repeat for each model Step 2 - Garment Reference (Image Node) Upload that model's garment design/reference. Label "Garment Reference 2," "Garment Reference 3," etc. one per model. Step 3 - Dress Change Request (Text-to-Text Node, Claude Sonnet 5 / Opus 5) Connect that model's Step 1 + Step 2. Prompt: "Change the dress of the model in image 1 from the current outfit to the new dress design in image 2, ensuring the new dress style is reflective of the provided dress details." Claude returns a full construction brief silhouette, neckline, fabric, drape not just a yes/no, giving the render step real detail to work from (visible in your "Picturing the transformation..." and "Here's the styling direction applied..." outputs). Step 4 - Render the composite (Image Node, Seedream 5 Pro / Nano Banana 2) Connected to Step 3's output. Produces "Dress Transformation Design" and "Gown Styling Details" - each model now in their new outfit, individually, on a plain background. This is Output 1 repeat Steps 1-4 once per model. Step 5 - Quality check Check each model's identity and garment Stage B - Combine into one posed duo composition Step 6 - Posture Reference (Image Node) A Reference photo of two people standing together in the pose/interaction you want. Label "two girl standing.jpg." Step 7 - Posture Analysis (Image-to-Text Node, Claude Sonnet 5) Connect Step 6. Prompt: "Analyze the body posture, standing style, and facial expression of each person in this image — ignore hairstyle and clothing entirely. For each person, describe: their stance (weight distribution, foot placement, leg position), the angle of their body relative to camera, arm placement, head/neck tilt, shoulder position, posture energy, and facial expression. Also describe the spatial relationship and interaction between them." This deliberately strips out garment/hair so the output is a pure pose-and-mood brief, reusable regardless of what either model is wearing. Step 8 - Generate the duo composition (Image Node, Nano Banana 2) Connect Step 7's posture brief + both models' Output 1 (Step 4, one per model). Produces "Posing Composition" / "Dynamic Duo Pose" — Output 2, both dressed models combined into one image, holding the exact stance/interaction from the reference. Step 9 - Quality check Check both models' identity and garments held correctly in the combined shot Stage C - Place the duo into a real-world background Step 10 - Background Reference (Image Node) Upload a real photo of two people in the actual location/lighting/composition you want the final hero image to have. Label "Background reference.jpg." Step 11 - Image Replacement Task (Image-to-Text Node, Claude Opus 5) Connect Step 10. Prompt: "Replace the two people in this reference image with these two people [Output 2], keeping their exact clothing, hairstyle, and appearance as shown. Preserve the reference image's exact pose, positioning, physical interaction, camera angle, and composition — only change who the two people are. Match the natural lighting, color grade, and background from the reference." Step 12 - Generate the final hero image (Image Node, Nano Banana 2) Connect Step 11's brief + Output 2 (Step 8) + Background Reference (Step 10). Produces "Image Swap Description" / "Position Swap Brief" the two dressed, posed models now composited into the real-world scene, matching its lighting, pose, and framing exactly. Step 13 - Quality check Check and confirm identity, garments, and pose all held through the final placement, and that lighting/color grade genuinely matches the background reference rather than looking pasted in. Step 14 - Export