Epoch Lab Control Center
Text-to-Image Realism Lab

Make generated photos stop announcing that they were generated.

This room is for controlled prompt experiments around naturalism: exposure, texture, skin, imperfect optics, city light, fabric behavior, and the quiet defects that make a picture feel like it came from a camera instead of a renderer.

How to use this room
This page is the method. Vision Lab is the machine that generates the test images.
5-minute workflow

Start here, then generate in Vision Lab.

Pick one scene.Use one Traveler-style scene or write a simple test scene. Do not change the subject while testing realism language.
Copy one prompt variant.Start with Control, then run Anti-HDR, Texture, and Phone Snapshot as separate tests.
Open Vision Lab.Paste the prompt into the main prompt box, keep No Character Lock selected, choose models, then click Generate.
Score the result.Judge skin, light, material, and place. The winner is the phrase that makes the image less synthetic, not merely prettier.
Open Vision Lab generate
Copy prompts

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.

Experiment protocol
This lab is not for pretty one-offs. It is for finding repeatable prompt language that makes images less synthetic.
Core question

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.

Baseline control prompt
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.
Failure signals

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.
Prompt families

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.
Scoring

Realism rubric

SkinTexture and anatomy look photographed, not polished.
LightExposure feels plausible for the scene, not globally optimized.
MaterialCloth, glass, food, pavement, and metal hold real-world defects.
PlaceThe location has human mess and specificity without becoming chaotic.
Promotion path

How findings graduate

  1. Run a control prompt in Vision Lab.
  2. Run one realism variant at a time.
  3. Score results against the rubric.
  4. Write the winning phrase into the experiment ledger.
  5. Promote only stable language into Traveler prompts.
Working links
This room defines the method. Vision Lab runs generic model comparisons. S Consistency Lab is separate.
Run tests

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.

Open Vision Lab bench Open S Consistency Lab identity
First experiment

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.

Variant seed
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.