Character creation through photography

This case study is still being iterated on as we refine the production notes.

Casting-level photography before prompting

Instead of inventing a character through prompts, we treated the shoot just like casting and wardrobe tests: controlled angles, consistent lighting, and a clear reference library.

Step 1: Photograph the character

• Front portrait
• Three-quarter portrait
• Side profile
• Full body standing
• Full body walking motion
• Emotion variations

Consistent lighting prevents the training data from introducing noise, just like keeping wardrobe tests under the same key light.

Step 2: Capture lens and distance variations

Close/medium/wide framings teach the LoRA how the character should read across shot sizes, mirroring how cinematographers test multiple lenses.

Step 3: Caption with extreme detail

Captions describe face structure, hairstyle, body proportions, clothing textures, color tones, and lighting so the model learns identity rather than guessing.

Step 4: Train the identity model

Once the dataset is stable, the identity model is trained and observed over multiple passes to detect drift. The early results surfaced the same pain point: open-source training tools are powerful but require filmmaker-friendly abstractions.