Creating a Synthetic Dataset

If you don't have real photos of a model, you can create a "Digital Twin" from scratch using AI. This guide explains how to generate a consistent dataset for LoRA training.

The Strategy: "The Perfect Seed"

The goal is to find one specific random seed that generates a face you love, and then re-use that seed across different prompts to keep the face consistent.

Step 1: Find the Face

1. Craft a detailed prompt: Include specific facial features (e.g., "hazel eyes", "small scar on eyebrow", "freckles"). > "A hyper-realistic close-up portrait of a 25-year-old woman with dirty blonde messy hair and hazel eyes. She has a symmetric face, sun-kissed skin texture, natural freckles. Soft natural lighting, 85mm lens."
2. Generate with FLUX.1: Use Replicate to run this prompt multiple times.
3. Select the Winner: Pick the face that embodies your character perfectly. Save the Seed number.

Step 2: Generate Variations

Now, generate 15-20 images using the SAME SEED but slightly modifying the context.

• Close-ups (10 images): Keep the prompt almost identical, maybe change lighting slightly.
• Half-Body (5 images): "A medium shot of [description...] wearing a t-shirt".
• Full-Body (5 images): "A full body shot of [description...] walking in the city".

Crucial: Do not change the facial description part of the prompt. Discard any images where the face looks different.

Step 3: Prepare for Training

1. Crop: Ensure faces are clear.
2. Rename: sienna_01.jpg, sienna_02.jpg, etc.
3. Zip: Compress them into dataset.zip.

You are now ready to train your LoRA!