Training character/face LoRAs on FLUX.2-dev with Ostris AI-Toolkit - full setup after 5+ runs, looking for feedback

A developer has documented a successful methodology for training high-fidelity character and face LoRAs on Stability AI's next-generation FLUX.2-dev model using the Ostris AI-Toolkit. Running on RunPod with 1x H100 SXM 80GB instances ($2.69/hr), the setup achieved production-quality results in approximately 3 hours per run at a cost of ~$8, reaching 0.75 InsightFace cosine similarity scores on validation tests.

The technical configuration reveals crucial differences between FLUX.1 and FLUX.2-dev implementations. Unlike FLUX.1, FLUX.2-dev requires specifying 'arch: flux2' instead of the 'is_flux: true' flag, and must enable 'quantize_te: true' to handle its Mistral 24B text encoder. The optimal training used LoRA rank 32 with linear_alpha 16, 5e-5 learning rate, 3500 steps, and multi-resolution bucketing at [768, 1024]. A key discovery was that multi-GPU (2x H100) provided zero speedup for LoRA training, making single-GPU setups more cost-effective.

The developer trained two fictional characters, with Character A achieving 0.753 similarity at step 3250 using 25 images. Performance analysis showed headshots scoring 0.83-0.86 similarity while full-body images dropped to 0.53-0.69. The most significant lesson emerged from Character A's training: dataset images with consistent accessories (gold jewelry in most photos) caused the LoRA to permanently bake these elements in, impossible to remove via negative prompting. For Character B, the dataset was improved with 8-10 accessory-free images and more varied outfits, demonstrating the critical importance of clean training data.