I have galleries on Deviant Art and Discord. But I stopped linking to them because the legal climate is clearly changing. I am probably very close to deleting everything I have public. Kinda 2025 is already very-very different from ~2021
I can't even access to D.A. since long with a vpn...I have galleries on Deviant Art and Discord. But I stopped linking to them because the legal climate is clearly changing. I am probably very close to deleting everything I have public. Kinda 2025 is already very-very different from ~2021
I mean even F95 took down the AI thread - that's a sign if one is looking
The D.A. is changing itself, after AI event everything is changed, starting without control and now they have to put something to limit AI. but is too late, there are many model, different train for specific role. See the llm model, there are really variety, some for coding, some for rp, etc.I have galleries on Deviant Art and Discord. But I stopped linking to them because the legal climate is clearly changing. I am probably very close to deleting everything I have public. Kinda 2025 is already very-very different from ~2021
I mean even F95 took down the AI thread - that's a sign if one is looking
You right, and the effect of political stuff and money (steam; civitai, for example) is visible to all.Yet I agree, it's like they are trying to put the hat on it since that's where the big money is now, or maybe it's just political, being China the leader in this technology... and I don't even start talking about the EU
Looks amazingly easy, you may want to make a github page for it.
I repeat, I'm new to stable diffusion
I created a json configuration file to train the lora
{
"model_name": "GiannaR_lora_v1",
"pretrained_model_name_or_path": "C:/Users/nameuser/Desktop/AI_Gen/Data/Models/StableDiffusion/perfection25DILXLIllustrious_33.safetensors",
"train_data_dir": "C:/Users/nameuser/Desktop/Training_GiannaR/Training_GiannaR1/img",
"output_dir": "C:/Users/nameuser/Desktop/Training_GiannaR/Training_GiannaR1/model",
"logging_dir": "C:/Users/nameuser/Desktop/Training_GiannaR/Training_GiannaR1/log",
"resolution": "1024,1024",
"train_batch_size": 1,
"max_train_epochs": 10,
"save_every_n_epochs": 1,
"gradient_accumulation_steps": 1,
"learning_rate": 0.0001,
"text_encoder_lr": 5e-05,
"unet_lr": 0.0001,
"lr_scheduler": "cosine",
"lr_scheduler_num_cycles": 1,
"lr_warmup_steps": 0,
"network_module": "networks.lora",
"network_dim": 64,
"network_alpha": 32,
"enable_bucket": true,
"min_bucket_reso": 256,
"max_bucket_reso": 2048,
"bucket_reso_steps": 64,
"bucket_no_upscale": true,
"optimizer_type": "AdamW8bit",
"min_snr_gamma": 5,
"mixed_precision": "fp16",
"save_precision": "fp16",
"save_model_as": "safetensors",
"cache_latents": true,
"cache_latents_to_disk": true,
"gradient_checkpointing": true,
"xformers": true,
"no_half_vae": true,
"caption_extension": ".txt",
"shuffle_caption": true,
"keep_tokens": 1,
"prior_loss_weight": 1.0,
"seed": 42
}Concentrate on Illustrious for animated (and is getting better for realistic) and SDXL for realistic, pony is over.
For SDXL:
these have been good to me(especially 1A). forgot where i got them from. i made changes and renamed them before i realized that. anyhow, take them to for a spin to test. you'll probably have to find the models. of course u can use what u have in your stash. also install missing nodes(if any). and change prompts to what u want. they are I2V workflows. using gguf models.
Auto1111, Fooocus, Forge etc.. are pre-built ready to go interfaces where all you have to do is change parameters and load the supported models (supported by the inbuilt workflows).