Millie Bobby Brown SDXL v2
Trained on BigLust v1.6
Updated to work with DMD2
Description
FAQ
Comments (11)
how did you train these loras?
kohya
I use it too, did you follow a specific guide? The face accuracy in your loras are impressive for being based SDXL
@tacocat not really, but here are the settings i used for millie
{
"LoRA_type": "Standard",
"adaptive_noise_scale": 0,
"additional_parameters": "",
"block_alphas": "",
"block_dims": "",
"block_lr_zero_threshold": "",
"bucket_no_upscale": true,
"bucket_reso_steps": 64,
"cache_latents": true,
"cache_latents_to_disk": false,
"caption_dropout_every_n_epochs": 0.0,
"caption_dropout_rate": 0,
"caption_extension": ".txt",
"clip_skip": 1,
"color_aug": false,
"conv_alpha": 1,
"conv_block_alphas": "",
"conv_block_dims": "",
"conv_dim": 1,
"decompose_both": false,
"dim_from_weights": false,
"down_lr_weight": "",
"enable_bucket": true,
"epoch": 100,
"factor": -1,
"flip_aug": false,
"full_bf16": true,
"full_fp16": false,
"gradient_accumulation_steps": 1,
"gradient_checkpointing": true,
"keep_tokens": 0,
"learning_rate": 1.0,
"logging_dir": "D:/pinokio/api/kohya.pinokio.git/kohya_ss/loras/_done/_dmd2/millie/log",
"lora_network_weights": "",
"lr_scheduler": "cosine",
"lr_scheduler_args": "",
"lr_scheduler_num_cycles": "1",
"lr_scheduler_power": "1",
"lr_warmup": 0,
"max_bucket_reso": 2048,
"max_data_loader_n_workers": "0",
"max_resolution": "1024,1024",
"max_timestep": 1000,
"max_token_length": 75,
"max_train_epochs": "",
"max_train_steps": "0",
"mem_eff_attn": false,
"mid_lr_weight": "",
"min_bucket_reso": 256,
"min_snr_gamma": 5,
"min_timestep": 0,
"mixed_precision": "bf16",
"model_list": "custom",
"module_dropout": 0,
"multires_noise_discount": 0.3,
"multires_noise_iterations": 0,
"network_alpha": 1,
"network_dim": 128,
"network_dropout": 0,
"no_token_padding": false,
"noise_offset": 0.07,
"noise_offset_type": "Original",
"num_cpu_threads_per_process": 2,
"optimizer": "Prodigy",
"optimizer_args": "d_coef=0.75 safeguard_warmup=True use_bias_correction=True weight_decay=0.01",
"output_dir": "D:/pinokio/api/kohya.pinokio.git/kohya_ss/loras/_done/_dmd2/millie/model",
"output_name": "Millie",
"persistent_data_loader_workers": false,
"pretrained_model_name_or_path": "D:/downloads/forge/webui/models/Stable-diffusion/bigLust_v16.safetensors",
"prior_loss_weight": 1.0,
"random_crop": false,
"rank_dropout": 0,
"reg_data_dir": "",
"resume": "",
"sample_every_n_epochs": 0,
"sample_every_n_steps": 0,
"sample_prompts": "high resolution photograph of person name. She is topless and taking a selfie photo with her iPhone 12. She has small sized breasts. You can see her nipples. --w 896 --h 1152 --n ugly, old, weird",
"sample_sampler": "euler_a",
"save_every_n_epochs": 5,
"save_every_n_steps": 0,
"save_last_n_steps": 0,
"save_last_n_steps_state": 0,
"save_model_as": "safetensors",
"save_precision": "fp16",
"save_state": false,
"scale_v_pred_loss_like_noise_pred": false,
"scale_weight_norms": 0,
"sdxl": true,
"sdxl_cache_text_encoder_outputs": false,
"sdxl_no_half_vae": false,
"seed": "0",
"shuffle_caption": true,
"stop_text_encoder_training_pct": 0,
"text_encoder_lr": 0.0,
"train_batch_size": 1,
"train_data_dir": "D:/pinokio/api/kohya.pinokio.git/kohya_ss/loras/_done/_dmd2/millie/img",
"train_on_input": true,
"training_comment": "",
"unet_lr": 1.0,
"unit": 1,
"up_lr_weight": "",
"use_cp": false,
"use_wandb": false,
"v2": false,
"v_parameterization": false,
"v_pred_like_loss": 0,
"vae_batch_size": 0,
"wandb_api_key": "",
"weighted_captions": false,
"xformers": "xformers"
}
Personally, I found SDXL face accuracy to be generally worse than SD1.5 and Flux in most loras. But yours seem to be pretty good. Maybe it's helped by the BigLust model. Just starting test making loras myself with that one. Thanks for the json info, I'll give your config a shot.
@tacocat no problem. dmd2 helps as well with accuracy
Just thoughs around conversation
I found, in Lora, it's important:
1. quality and coherence of the dataset
2. which checkpoint you train on
3. on which checkpoint you use (not always Lora works better on the checkpoint you trained on).
Most of the tutorials do not correspond to reality. Only practice.
another paremeters should be set in reasonable values, there are no clear rules. Optimizers simplify a lot, you can concentrate on the number of epochs and repetitions and other(but i do not much). There are to many combinations, and haven't tried even a fraction of them.
@rustyshackleford69420 ,nice Loras by the way :).
@DRZ3000 Agreed. I've had a lot of experience training loras, hundreds of them. The face accuracy in mine, and everyone else's, never looked all that great with SDXL in my opinion. I had more luck with SD 1.5 full model training and Flux loras, but those have "limitations" on the kinds of images they can produce. But these latest SDXL loras are looking way better now, probably due to updates to the training apps, better training settings, and more refined models. A lot of the loras trained on big lust are impressive so far, considering it's still SDXL.
@tacocat I find there is no definite config, I have to do some range in repeats, epochs, other sht. I have now found that some of my Loras (may be all) works better with DMD2 Lora and my world has been turned upside down, AGAIN, WTF
@rustyshackleford69420 I didn't quite get the point, are you preparing config for DMD2 in any particular way? I just noticed that my casual Lora works better in parallel with DMD2's Lora. I had the idea of training my Lora in the same dim and then merging it with DMD2, but I don't know if that makes sense yet
@DRZ3000 you probably don't need to do anything to make your loras work with dmd2. it's just my very first loras had some setting wrong that made them look bad with dmd2. still don't know what it was tbh
Looks like we don't have an active mirror for this file right now.
CivArchive is a community-maintained index — we catalog mirrors that volunteers upload to HuggingFace, torrents, and other public hosts. Looks like no one has uploaded a copy of this file yet.
Some files do get recovered over time through contributions. If you're looking for this one, feel free to ask in Discord, or help preserve it if you have a copy.
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.