Train Your First LoRA on FLUX 3 Image Klein: Dataset, Captions, Settings and Picking a Checkpoint
Black Forest Labs' official guide trains a style LoRA on its open FLUX 3 Image Klein model with ostris' AI Toolkit, using 65 public-domain Halloween postcards as the example. It covers the dataset, captions and trigger word, the settings that matter, and how to compare checkpoints to pick one.
Key points
- You need a GPU with at least 24 GB of VRAM, AI Toolkit and access to FLUX 3 Image Klein Base on Hugging Face; check the FLUX Kommunity License before you share or sell a LoRA
- Captions describe what is in each image and leave the style out, because the trigger word absorbs the look
- Use a meaningless trigger such as vqx: with pdhalloween the model read the word halloween and added pumpkins on its own
- Rank 32 gives a 161 MB LoRA; keep the learning rate at 1e-4 with constant_with_warmup or cosine
- timestep_type: shift had the biggest effect of any setting; if samples barely change after a few hundred steps, check it first
- Do a 50-step smoke test, then train 4,000 steps, saving and sampling every 100 with fixed seeds; BFL's favourites were at 2,000, and 4,000 began copying training cards