FLUX Klein fine-tuning hits Kohya wall as trainers seek cloud templates

Kohya_ss, the standard fine-tuning interface for Stable Diffusion and SDXL, doesn't support FLUX Klein or Zbase out of the box, leaving trainers scrambling for cloud-compatible alternatives.

A creator who previously fine-tuned Illustrious into the Dixar checkpoint posted this week about the gap. They've assembled a 15,000-image dataset from both open-source and proprietary outputs, ready to train a FLUX Klein variant. They've used Vast.ai with Kohya for Illustrious fine-tunes in the past but found no ready-made template for the newer architectures. OneTrainer, a more recent option, failed to launch on Vast instances. The workflow relies on Google Drive for dataset storage and Vast for high-spec GPU access, making template availability the critical bottleneck.

FLUX Klein, Black Forest Labs' 9-billion-parameter open-weight model released in November 2024, uses a different architecture than SDXL or Illustrious. Zbase, a community-driven base model, shares the same compatibility gap. Kohya_ss has historically supported LoRA and full fine-tuning for Stable Diffusion 1.x, SDXL, and Pony variants, but the codebase hasn't caught up to FLUX's MMDiT blocks or the newer tokenizer stack. The FLUX Klein model card on HuggingFace lists training details but doesn't link to a public fine-tuning script.

Community members pointed to ai-toolkit and SimpleTuner as alternatives. Both support FLUX fine-tuning but require manual Docker setup or local Python environments — a friction point for trainers who rely on one-click Vast templates. The creator emphasized preferring a GUI-driven workflow with minimal code, the same approach that made Kohya popular for SDXL. Without a maintained Vast template, FLUX Klein fine-tuning remains accessible only to users comfortable writing launch scripts or building custom containers.