Model
Here you define the base model you use for training and the output model you want to achieve.
- Base Model (default: hugging face link to base): either provide the path of a saved base model or provide a link to it in the huggingface format (example: stabilityai/stable-diffusion-xl-base-1.0).
- Vae (default:blank): If you want to use a custom VAE, it must be a huggingface link (example: madebyollin/sdxl-vae-fp16-fix)
- Model Output destination: set any local folder with its name and type (ex C:/OneTrainer/Output/xxx.safetensors) Note - models that produce diffusor outputs expect a directory, not a safetensor file. This currently includes Pixart Alpha.
- Output Format (default:safetensors): Here you can choose between the default safetensors and the optional checkpoint format.
- Data Type: 4 options are available. From the least to most precise they are: float8, bfloat16, float16 and float32. The more precision you use, the more VRAM you will require to load the weights.
| DType for | Fine Tune | LoRA | Embedding | Fine Tune VAE | Comment |
|---|---|---|---|---|---|
| Output | All | All | All | All | Comment |
| Weight | All | All | All | All | Comment |
| Override Text Encoder | float32 | All | Not relevant | Not relevant | Comment |
| Override Unet | All | All | Not relevant | Not relevant | Comment |
| Override VAE | float32, bfloat16 | float32, bfloat16 | float32, bfloat16 | float32, bfloat16 | Comment |
If all samples are black, try to set the VAE data type to float32 or bfloat16, or use the VAE override to load the fixed VAE.
Note: The presets (top left dropdown: #SD1.5 LoRA, #SD1.5 Embedding, ...) will set up defaults values, you can stick to it, they are working.