PVC style LoRA for
AuraFlow v0.3.
Only tested with ComfyUI, and I don't know if this LoRA works on other WebUIs.
Long danbooru tags style prompt is preferred.
This LoRA was trained with danbooru tags, but not of all danbooru tags would work.
Please consider to use natural language to align. (but too long natural language prompts sometimes break outputs.)
-
Training tool:
auraflow-ft
-
- Quantization: bitsandbytes NF4
- Target keys: all except the exclude keys
- Exclude keys:
["t_embedder", "final_linear", "modF"]
-
Peft method: LoRA
- LoRA settings
- Rank: 16
- Alpha: 1
- Zero-init: lora_down
- Target keys:
[".attn.", ".mlp.", ".modC.", ".modC.", ".modX."]
- Exclude keys:
["text_encoder", "vae", "t_embedder", "final_linear"]
-
Optimizer:
RAdamScheduleFree from
schedulefree
-
Learning rate: 1e-3
- Learning rate scheduler: not used
-
Loss type: L2
-
Timestep sampling: Logit-norm
-
Train epochs: 10
-
Dataset
- Num images: 7452
- Num repeats: 2
- Batch size: 2
- Bucket settings: (maybe buggy)
- Maximum resolution: around 1024x1024
- Minimum resolution: 384x384
- Bucket size step: 128
- Caption augumentation
- Shuffle tags: Yes
- Tag dropout: No
- Caption dropout: No
-
Gradient checkpointing: Yes
-
Training device: 1x RTX 4070Ti Super (16GB VRAM)
-
Training hours: around 62 hours
-
Weights for this model are available in Safetensors format.
Download them in the Files & versions tab.
