[Flux.1 dev] LuciGrimace Concept LoRa

Details

Download Files (1)

About this version

accelerate launch "./sd-scripts/flux_train_network.py" ^
--pretrained_model_name_or_path "D:\kohya_ss\flux\unet\flux1DevFp8_v10.safetensors" ^
--clip_l "D:/stable-diffusion-webui-forge-main/models/text_encoder/flux/clip_l.safetensors" ^
--t5xxl "D:/stable-diffusion-webui-forge-main/models/text_encoder/flux/flux1T5TextEncoder_stock.safetensors" ^
--ae "D:/stable-diffusion-webui-forge-main/models/VAE/flux/flux1DevVAE_stock.safetensors" ^
--train_data_dir "D:\kohya_ss\flux\img" ^
--reg_data_dir "D:\kohya_ss\flux\regularization\img" ^
--output_dir "D:/kohya_ss/flux/models/Lucignolo_grimace_flux1_withLabel" ^
--output_name "Lucignolo_grimace_flux1_withLabel" ^
--save_model_as safetensors ^
--sdpa ^
--sample_sampler "euler" ^
--sample_prompts "D:\kohya_ss\lucignolo_grimace_prompts.txt" ^
--sample_every_n_epochs 1 ^
--max_data_loader_n_workers 2 ^
--seed 777 ^
--gradient_checkpointing ^
--mixed_precision bf16 ^
--save_precision bf16 ^
--network_module networks.lora_flux ^
--network_dim 16 ^
--network_alpha 32 ^
--network_args train_blocks=single ^
--optimizer_type Prodigy ^
--lr_scheduler constant ^
--lr_scheduler_num_cycles 1 ^
--lr_scheduler_power 1 ^
--learning_rate 1 ^
--network_train_unet_only ^
--timestep_sampling sigmoid ^
--max_grad_norm 1 ^
--model_prediction_type raw ^
--apply_t5_attn_mask ^
--guidance_scale 1.0 ^
--min_bucket_reso 256 ^
--max_bucket_reso 2048 ^
--bucket_reso_steps 64 ^
--resolution 512,512 ^
--discrete_flow_shift 3 ^
--cache_text_encoder_outputs ^
--cache_text_encoder_outputs_to_disk ^
--cache_latents ^
--fp8_base ^
--enable_bucket ^
--split_mode ^
--caption_extension .txt ^
--keep_tokens 0 ^
--min_snr_gamma 7 ^
--save_every_n_epochs 1 ^
--train_batch_size "3" ^
--noise_offset 0.05 ^
--max_train_epochs 30 ^
--max_timestep 1000 ^
--adaptive_noise_scale 0

Model description

trigger word: grimace

recommended weight: 1.0

About

This model was based on a phenomenon that occurs mostly in the Lucignolo Merge Checkpoint ( made by ALIENHAZE )

Note for ALIENHAZE: I got this grimace phenomenon to work with same prompt in a few XL checkpoint merges, and merge only checkpoints, but your model is the only that can generate it very well. This could be a signal that people are merging their models to your model, but i can't prove anything. Idk where this grimace come from, it's very weird.

The Lucignolo Phenomenon

when you set the "correct parameters" for ksampler and the prompt to:

incredibly absurd unrelatable content without context, inexplicable, uncoordinated

It generates a very uncommon grimace in all images, as you can see, (those are from my dataset):

So, i just loved it very much to put effort to train a LoRA for it.
I liked to set it more realistic, which made it much creepier instead artful 3D from dataset.

I just added general regularization of many creatures, so it could apply to anything that lives.

This maybe final model version is actually a merge of epoch 13# trained without any caption and epoch 2# with the one word only caption "grimace", so it could be applied with much more efficiency using the word "grimace" in the prompt.

Images made by this model

Sort by

Model Type	LORA
Base Model	Flux.1 D
Published	2024-09-25
Trained Words	grimace