Installation:
Install this ComfyUI custom node
ComfyUI_LLM_SDXL_Adapter_Additions.
Alongside the required
ComfyUI_LLM_SDXL_Adapter
.
Relevant Details for Usage
-
-
Use the LLMAdapterLoaderUnfused node at the node path llm_sdxl/additions/ to load the adapter.
-
The noise_aug V1 to V3 loras are experimental and shouldn't be used (unless you want to test).
-
noise_aug_V4 and chenkin03unetTrainedOnAdapter can be used together.
Tagging Information:
Positive:
masterpiece, best quality,
Negative:
worst quality, low quality, bad anatomy
Artists: Tagged with the prefix by, e.g., by someArtistHere.
- Date Tags:
- Content Rating Tags:
Rating: ExplicitRating: QuestionableRating: SensitiveRating: General
Training Information:
~ 80,000 unbatched steps per epoch
batch size: 28
gradient accumulation: 1
learning_rate: 2e-6
learning warmup steps: 80
lr_scheduler: cosine
max_train_epochs: 4
Final used model epochs: 2 / 4
precision: mixed precision BFloat16
optimizer: AdamW8kahan (AdamW8Bit with Kahan summation)
optimizer_args: None (Weight decay code was broken when training this.)
flow model arguments:
- flow_model
- flow_use_ot
- flow_timestep_distribution uniform
- flow_uniform_static_ratio 3.0
loss_type: L2
loss_objective Flow Matching
Trained by freezing everything in the adapter except:
- pooling_attention
- pooled_projection
- Last two narrow_attention_blocks
with a max learning rate of 2e-6 used.
