No blowjob model for SDXL? Time to change that!
Trained on 300+ images for 13500 steps.
base prompt:
woman, sucking a cock
nice simple prompt example:
photo of a elven woman, sucking a cock, cold light
But of course you can get crazy with SDXL. It's very powerful!. From fantasy, ordinary photos to paintings. It just works!
I generated my images with ComfyUI. Just drop the images into ComfyUI to get the node setup. It's a super basic setup, but works really well with that lora.
Sometimes adding 'closeup' helps if strange compositions happen.
If your prompt gets to complex and the blowjob isn't happening anymone, then raise the lora strenght to around 1.3.
Have fun ;D
Description
First version. Works best on close ups.
FAQ
Comments (25)
add XL to the name to make it obvious
Done ;)
Soon to be the most downloaded LORA here
I hope so ;)
Is it just me or does this already look miles ahead of any blowjob 1.5 could have produced?
the quality and variety of SDXL is just amazing! Immediately elevades the Lora into top tier regions!
@WhatTheGuy How paradoxical... You're saying that going down goes up!
@vestin Ah took a while, but you are right, hah!
Remember a fortnight ago when people were claiming SDXL was so censored it couldn't be trained to do even a nipple....
@countlippe yeah that was when they tried to get nudity out of the censored online generator with all kind of wonky prompts that then of course gave strange results
I'm having a hell of a time training a decent LORA on SDXL. Could you share the hyperparameters you used to train this?
Secret sauce was unchecking 'use buckets' . Since then everything came out better. Hera are all setings:
{
"LoRA_type": "LyCORIS/LoHa",
"adaptive_noise_scale": 0,
"additional_parameters": "--max_grad_norm=0",
"block_alphas": "",
"block_dims": "",
"block_lr_zero_threshold": "",
"bucket_no_upscale": true,
"bucket_reso_steps": 64,
"cache_latents": true,
"cache_latents_to_disk": true,
"caption_dropout_every_n_epochs": 0.0,
"caption_dropout_rate": 0,
"caption_extension": ".txt",
"clip_skip": "1",
"color_aug": false,
"conv_alpha": 4,
"conv_alphas": "",
"conv_dim": 8,
"conv_dims": "",
"decompose_both": false,
"dim_from_weights": false,
"down_lr_weight": "",
"enable_bucket": false,
"epoch": 30,
"factor": -1,
"flip_aug": false,
"full_fp16": false,
"gradient_accumulation_steps": 1.0,
"gradient_checkpointing": true,
"keep_tokens": "0",
"learning_rate": 1.0,
"logging_dir": "F:/kohya/TRAINING/sdxl/w sucking a cock- all v3 - NC/log",
"lora_network_weights": "",
"lr_scheduler": "cosine",
"lr_scheduler_num_cycles": "",
"lr_scheduler_power": "",
"lr_warmup": 0,
"max_data_loader_n_workers": "0",
"max_resolution": "1024,1024",
"max_timestep": 1000,
"max_token_length": "75",
"max_train_epochs": "",
"mem_eff_attn": false,
"mid_lr_weight": "",
"min_snr_gamma": 0,
"min_timestep": 0,
"mixed_precision": "bf16",
"model_list": "custom",
"module_dropout": 0,
"multires_noise_discount": 0.2,
"multires_noise_iterations": 8,
"network_alpha": 16,
"network_dim": 32,
"network_dropout": 0,
"no_token_padding": false,
"noise_offset": 0.0357,
"noise_offset_type": "Original",
"num_cpu_threads_per_process": 2,
"optimizer": "Prodigy",
"optimizer_args": "decouple=True weight_decay=0.5 betas=0.9,0.99 use_bias_correction=False",
"output_dir": "F:/kohya/TRAINING/sdxl/w sucking a cock- all v3 - NC/model",
"output_name": "woman, sucking a cock - v3 -nc",
"persistent_data_loader_workers": false,
"pretrained_model_name_or_path": "F:/ComfyUI/ComfyUI_windows_portable/ComfyUI/models/checkpoints/sd_xl_base_1.0.safetensors",
"prior_loss_weight": 1.0,
"random_crop": false,
"rank_dropout": 0,
"reg_data_dir": "F:/kohya/TRAINING/sdxl/w sucking a cock- all v3 - NC/reg",
"resume": "",
"sample_every_n_epochs": 1,
"sample_every_n_steps": 0,
"sample_prompts": "woman, sucking a cock --w 1024 --h 1024",
"sample_sampler": "euler_a",
"save_every_n_epochs": 1,
"save_every_n_steps": 0,
"save_last_n_steps": 0,
"save_last_n_steps_state": 0,
"save_model_as": "safetensors",
"save_precision": "bf16",
"save_state": false,
"scale_v_pred_loss_like_noise_pred": false,
"scale_weight_norms": 1,
"sdxl": true,
"sdxl_cache_text_encoder_outputs": false,
"sdxl_no_half_vae": true,
"seed": "12345",
"shuffle_caption": false,
"stop_text_encoder_training_pct": 0,
"text_encoder_lr": 1.0,
"train_batch_size": 8,
"train_data_dir": "F:/kohya/TRAINING/sdxl/w sucking a cock- all v3 - NC/img",
"train_on_input": true,
"training_comment": "",
"unet_lr": 1.0,
"unit": 1,
"up_lr_weight": "",
"use_cp": true,
"use_wandb": false,
"v2": false,
"v_parameterization": false,
"vae_batch_size": 0,
"wandb_api_key": "",
"weighted_captions": false,
"xformers": true
}
Thank you!
This kinda blew my mind (pun intended), I really didn't have hope in XL doing XXX until this. The diversity in positions, flexibility in styles, and ability to swap gender this can do on top of Dreamshaper is quite impressive. Not to mention no 1.5 BJ lora can generate accurate male genitalia like this does out of the box! Excited for more, thanks for this
Glad you like it! I always felt like poeple want to ignore XL because it is -as a base model- not as good as those highly finetuned 1.5 models, not understanding that we are behind at the moment, but will surpass 1.5 in no time, archiving multiple time the quality of it easily ... So I made that as a kind of lighthouse project to show the power that so many seem not to see. But I have to say I'm also surprised that it will became THAT good, hah. I hope it will get the bandwagon rolling and people see the huge advantage of continuing with SDXL from now on!
Thanks for this.. i have question.. did you get strange results when using refine?
yes I'm just using it with denoise level of 0.05. Otherwise it's messing stuff up, because the Refiner isn't trained for that
The Name 🤣🤣🤣👍🏻
Hah thanks, took a while unitl I had it xD
Please add more stuff for SDXL !
actually... I have something that I could just upload. I will look into that!
best of its kind.
did you use regularization? the other models seem to mess up the world, this one only slightly
I used no regulation images
This tool doesn't work on Macbook Pro with M2 CPU. Automatic1111 throws error "BFloat16 is not supported on MPS". Is there another version of this model implemented with float16 rather than bfloat16?
no sorry, that's all I have