Photoreal stills

Produce still images that a person would believe came out of a camera, not a model. The hard part is not resolution, it is defeating the "AI look": waxy skin, plastic textures, eerie symmetry, blown-out studio light, and over-saturated everything. The lever is prompting for real optical and lighting behavior plus deliberate imperfection, then judging the result against an is-this-a-photo bar.

Inputs to collect

• The scene / subject and the genre it belongs to (candid portrait, editorial, documentary, food, architectural, lifestyle). Genre sets the lens, light, and framing defaults.
• The vibe / reference: a era, film stock, magazine, or photographer style if the user has one. (Ask only if the request is generic.)
• Aspect ratio and orientation for the intended use (vertical for phone/social, 3:2 for editorial, square for grid).
• Optional: a real reference image to match, brand-safety constraints (no readable on-image text or logos unless supplied), how many variants.

Models

• Default: FLUX.2 [max] (bfl:7@1), flagship text-to-image with the strongest prompt adherence and the most convincing micro-texture and lighting. Best general pick for photoreal.
• Strong alternative: Qwen-Image-2.0-Pro (alibaba:qwen-image@2.0-pro), high visual fidelity and reliable text/iconography, good for editorial and lifestyle frames that include signage or packaging.
• ByteDance Seedream for a different texture and fast iteration: Seedream 4.5 (bytedance:seedream@4.5) for high-fidelity 2K to 4K, or Seedream 5.0 Lite (bytedance:seedream@5.0-lite) as a cheaper, faster draft tier.
• Confirm each is live and inspect its schema before calling. Use runware-models (live lookup) + runware-run. Never hardcode a stale choice. Models change weekly.

Workflow