Step-by-step: Training a product-specific LoRA with 50 images.

Trained a product-focused LoRA on exactly 50 product shots. Goal was consistent style and branding across generated images. Used Kohya_SS.

Key parameters and results:

**Dataset:**
* 50 images, 512x512 resolution.
* Captioned with BLIP, manually verified. Tags: product name, material (e.g., `acrylic_frame`), background color.
* `instance_token`: `prod_a1b2`

**Training Config (kohya_ss `train_network.py`):**
```bash
--network_module=networks.lora
--network_dim=32
--network_alpha=16
--train_batch_size=4
--num_epochs=10
--learning_rate=1e-4
--optimizer_type=AdamW8bit
--lr_scheduler=cosine
--save_every_n_epochs=2
--mixed_precision=fp16
```

**Outcome:**
* Training time: ~2.1 hours on an RTX 4090 (24GB VRAM).
* Final model size: 8.4 MB.
* Effective after 8 epochs; last 2 showed minimal loss decrease.
* Inference test: 95% consistency on product shape and key features when using `prod_a1b2` in prompt. Background and lighting variations remain controllable.

Pitfall: Initial run with `network_dim=128` and 20 epochs led to overfitting—model replicated training images. Lower dim and earlier stopping solved it.