Workflow: photorealistic product UGC with GPT-Image 2.5 via JSON style prompts
sven_ai · x · 2026-09-09
A complete workflow for photorealistic AI product photos:
- Problem: GPT-Image 2.5 still looks "AI-ish" without references — mostly due to color grading and grain.
- Style extraction: Upload a reference photo to GPT-6 Astra and ask for a detailed JSON describing color grading, lighting, and exact colors.
- Generation: Feed the JSON plus your product image to GPT-Image 2.5 to generate a person holding your product; compare against unreferenced outputs to see the lighting/grading difference.
- Character consistency: Save a satisfactory character photo as a reference and attach it in future generations to keep facial features consistent, building a reusable UGC library across products.
- The JSON drives lighting/color direction, the character reference keeps identity, and prompts can be iterated with Astra.
More from Multimodal
- OpenAI reportedly ships ChatGPT Images 2.5 with 50% lower latency, free for all tiers — eyishazyer · 2026-09-09
- SuperSplat browser viewer showcases stunning user-uploaded 3D Gaussian captures — willeastcott · 2026-09-09
- 5 ChatGPT Image Prompts That Turn Casual Photos Into Pro Shots (Face Preserved) — TawohAwa · 2026-09-09
- Sculpt Canvas Brings 10-Brush 3D Sculpting Inside ComfyUI Graphs — skbphy · 2026-09-09
- 'Generate for the Edit' Accepted to SIGGRAPH Asia 2026 Posters — ring_hyacinth · 2026-09-09
- 10 Hours, 21 Revisions, 1.3M Polygons: Artist Builds Vintage Gym in Blender via GPT-6 Astra — petewoodbridge · 2026-09-09