Start here, then generate in Vision Lab.
First test pack
Use these as copy/paste starters. Run them one at a time. If a variant improves realism, we keep the useful phrase and test it again on another scene.
Control
Neutral baseline. Use this first so the comparison has a reference.
Anti-HDR
Tests whether the model can stop lifting every shadow and polishing every highlight.
Texture
Pushes skin, fabric, pavement, glass, and food away from the clean AI finish.
Phone Snapshot
Tests whether a less editorial camera instruction creates a more believable picture.
Can prompt language reduce the visible AI finish?
The target is not “cinematic” by default. The target is believable: small exposure mistakes, real fabric, ordinary lens behavior, imperfect location light, skin that has texture without becoming harsh, and backgrounds that do not feel art-directed by a machine.
Photorealistic editorial travel image of an adult subject in a real city location. Natural daylight, believable camera exposure, real fabric behavior, skin texture preserved, no beauty retouching, no HDR, no plastic skin, no glossy render finish, no over-smoothed surfaces, no cinematic color grade.
What we are hunting
- Plastic skin or poreless face finish.
- Overexposed highlights and lifted grey shadows.
- Too-clean clothing, streets, counters, glass, or food.
- Unnatural bokeh that feels pasted on.
- Fake editorial lighting in ordinary places.
- Hands, text, signs, and background objects that reveal the model.
Variables to test
- Exposure language: “metered for skin” versus “available light.”
- Texture language: fabric weave, skin texture, dusty surfaces.
- Camera language: phone snapshot, compact camera, documentary still.
- Anti-render language: no HDR, no glossy finish, no beauty retouching.
- Lighting language: flat overcast, mixed tungsten, fluorescent, hard noon.
Realism rubric
How findings graduate
- Run a control prompt in Vision Lab.
- Run one realism variant at a time.
- Score results against the rubric.
- Write the winning phrase into the experiment ledger.
- Promote only stable language into Traveler prompts.
Use Vision Lab as the bench
Start from a fixed scene, then run the same prompt across Flux, GPT Image, and any challenger models. The point is to separate model weakness from prompt weakness.
Flux Ultra plasticity test
Recommended first run: take one Traveler-style prompt and create four variants: control, anti-HDR, texture-heavy, and phone-camera realism. Compare whether the “AI sheen” decreases or whether the model ignores the language.
Add: available light only, imperfect real-world exposure, slight sensor noise in shadows, natural skin texture, matte fabric, no HDR, no commercial retouching, no plastic skin, no synthetic gloss.