Model

Here you define the base model you use for training, data types and the desired name of output model (This applies to both LoRA and Finetune)

• Hugging Face Token: you can specify here your HF token, required to download gated models from Hugging Face (SD3, Flux). It will be saved locally in secrets.json and reused when loading any preset.
• Base Model (default: Hugging Face repo): either keep the default or provide the path to a saved model (in safetensor format or a directory for diffused models). If you did a large scale finetune and then wanted to continue training it this is where you would place it.
• Override transformer / GGUF GGUF support for Lora, this field appears when selecting a Flux/Chroma/Qwen preset, see the note below for its usage and purpose.
• Vae Override (default: blank): If you want to use a custom VAE, provide a Hugging Face link or a path to a local file.
• Model Output Destination: file name or directory where the output model is saved. In case of a directory, OT use the save prefix set in the backup tab and a timestamp to name the file.
• Output Format (default: safetensors): Here you can choose between the default safetensors and the optional checkpoint format.
• Data Types: several options are available. The default presets (#SD1.5 LoRA, #SD1.5 Embedding, ...) will set up defaults values, you can stick to it, they work fine. Dont touch unless you have a reason.

Note: To restore a specific backup (and not the latest), select the specific backup epoch folder you care about as the base model path.

Note about GGUF support for Lora (Flux/Chroma/Qwen only):

• If you use GGUF during inference, you can more accurately train to that checkpoint by using the same model during training.
• By using a high-quality GGUF such as Q8: train in higher quality than using float8.
• By using a mid-quality GGUF such as Q4_K_S: train in a higher quality than nfloat4. nfloat4 has shown issues on several models. GGUF models of similar size might be better.
• By using a low-quality GGUF such as Q2 or Q3: Extremely low VRAM training. Using the same GGUF during inference as during training is recommended.

Usage:

• Pick a GGUF file (Flux.-dev-GGUF, Chroma1-HD-GGUF, Qwen-image-GGUF) and put it in Override transformer / GGUF, can be a link or a local file.