Model Guides¶

Training examples — Examples are organised as model_family/model_flavour/vram/recipe. Choose a model family, then its flavour and a memory tier such as 24g, 48g, or 80g. The default tier has no specific VRAM target; a default flavour uses the model’s configured default. Multi-GPU tiers such as 8x80g require eight GPUs with that memory per GPU.

In the WebUI environment creator and setup wizard, select the model family first. The example selector shows only that family’s recipes. Changing the family clears the previous example selection.

Step-by-step guides for training each supported model architecture.

Image Models¶

Flow Matching¶

DiT / Transformer¶

U-Net¶

Image Editing¶

Video Models¶

License notes: The commercial-use status covers model weights, derivative checkpoints, fine-tunes, and hosted model use. Generated-output rights can differ; read the linked license text before commercial deployment.

¹ OpenRAIL-style licenses generally permit commercial use with usage restrictions that remain attached to the model and derivatives.

² Stability AI Community License is available for qualifying users below the revenue threshold; larger commercial use needs Stability enterprise terms.

³ Flux.1 varies by flavour: Schnell and LibreFlux are Apache-2.0, while Dev, Krea, and Kontext use BFL non-commercial terms; review FluxBooru upstream metadata before commercial use.

⁴ Flux.2 varies by flavour: Klein 4B is Apache-2.0, while Dev and Klein 9B use BFL non-commercial terms.

⁵ Public non-commercial model terms do not permit commercial use of weights, derivative checkpoints, or hosted model services without a separate license.

⁶ Krea 2 Community License permits commercial use under its revenue cap (under $1M annual revenue) and safety/filtering requirements; otherwise an enterprise license is required.

⁷ Abandonware means the original vendor has effectively left the model behind and there is no reliable permission path; end users must decide whether that risk is acceptable.

⁸ AuraFlow supports Apache-2.0 upstream flavours and a Pony flavour with a separate custom license; check the selected flavour.

⁹ NVIDIA Open Model License permits commercial use but includes agreement, acceptable-use, and export-control terms.

¹⁰ LTX Video 0.9.5 uses OpenRAIL-M; LTX Video 2 uses LTX community terms with a revenue threshold for commercial use.

¹¹ Tencent Hunyuan Community License includes territorial exclusions and a commercial threshold for very large services.

¹² MiniMax H3 Community License excludes the United States, European Union, United Kingdom, and South Korea from the standard applicable territory; those territories require separate authorization.

Audio Models¶

Choosing a Model¶

For beginners:

• Start with Flux.1 for high-quality image generation
• Use LoRA training to reduce memory requirements

For production:

• SD3 or SDXL for broad compatibility
• Flux.2 for maximum quality (requires more VRAM)

For video:

• Wan Video for best quality/resource balance
• Hunyuan Video for I2V with super-resolution

For specific use cases:

• Flux Kontext for image editing/conditioning
• ACE-Step for text-to-music LoRA training (v1 and v1.5)
• HeartMuLa for autoregressive text-to-audio

Qwen Image 2.1: use the dedicated quickstart for assisted REPA presets, multi-scale data, auto shift and the fixed VAE. Assistant creation recipes are separate experiments.