Training Tool Comparisons

last edited

Training tool comparisons cover open-source trainers for adapters and fine-tuning. Each comparison weighs kohya-ss sd-scripts against other trainers on model support, realism and setup effort. Adapter basics are on the LoRA page, and rank settings are in the advanced techniques page.

Kohya ecosystem

The name kohya covers two projects: sd-scripts and a separate interface. The sd3 branch was merged into main through a pull request. Main therefore covers FLUX.1 and SD3 training in most tool comparisons.

Recent kohya-ss sd-scripts releases:

  • v0.12.0 (24 September 2026) added ARM64 support and truncated_logit_normal timestep sampling.
  • v0.12.2 (20 October 2026) added H2D-only block swap for LoRA training.
  • v0.12.3 (26 October 2026) added a torch.compile fix.

The interface released v26.0.0 in July 2026 with native low-rank variants for SDXL. Its cache latents setting blocks augmentation, except flip_aug. Optimizer choice is covered on the optimizers page.

Kohya vs OneTrainer

In training tool comparisons, sd-scripts is favored for generalization. Its adapters hold up across diverse prompts and suit broad style or concept work.

OneTrainer supports masked training, which limits the loss to chosen image regions. It can also train the FLUX developer model with its text encoders. A published FLUX fine-tuning test favors OneTrainer for photorealistic likeness.

OneTrainer is developed on its master branch and has no tagged releases. Its commit history is the reference for current changes.

Other trainers and selection

AI-Toolkit is MIT-licensed and runs as a web interface or command line. A commit on 10 November 2026 added an image model. Community users often rank it first for beginners.

SimpleTuner v4.11.0 organizes examples by model family, flavour and VRAM requirement. Its release notes list the attention default.

Selection guide for trainers in tool comparisons: OneTrainer suits real-person likeness, sd-scripts suits flexible adapters, AI-Toolkit offers easy setup, and SimpleTuner sorts configs by VRAM.

Failed runs are analysed in Civitai training post-mortems.

Sources