This is a huge collection (to be uploaded gradually) of aesthetic DoRAs for science fiction cities, buildings, interiors, and vehicles. I've tried to generate the same set of prompts and seeds for each one so that you have a good basis for comparison. Pick the ones you like, or just download all of them. :)
These DoRAs are for Envy Starlight, but will probably work with other SDXL models as well.
Description
FAQ
Comments (2)
Hey there Envy! I have a question regarding DoRas. Do you train them locally? I've written multiple posts here and on Reddit asking about online services, where I could train DoRas easily, but no one could give me an answer. The Ally mentioned that DoRa training might be coming to CivitAI in the future. But in the meantime, do you know any website, where I could train DoRas? (I only have an 8GB RTX 2060 so I won't even try locally).
Thank you! ^^
I train them locally, although anything with sufficient ram that can run the current version of kohya's sd-scripts or kohya_ss can probably do it. This is the (windows) command line that I've had luck with:
accelerate launch --mixed_precision="fp16" --num_processes=1 --num_machines=1 --num_cpu_threads_per_process=2 "./sd-scripts/sdxl_train_network.py" --bucket_no_upscale --bucket_reso_steps=64 --cache_latents --caption_extension=".txt" --enable_bucket --min_bucket_reso=256 --max_bucket_reso=2048 --gradient_checkpointing --huber_c="0.1" --huber_schedule="snr" --learning_rate="1.0" --loss_type="l2" --lr_scheduler="constant" --lr_scheduler_num_cycles="8" --max_data_loader_n_workers="0" --max_grad_norm="1" --resolution="1024,1024" --max_train_steps="67" --mem_eff_attn --min_snr_gamma=5 --min_timestep=0 --mixed_precision="fp16" --network_alpha="4" --network_args preset="full" conv_dim="8" conv_alpha="4" rank_dropout="0" bypass_mode="False" dora_wd="True" module_dropout="0" use_tucker="False" use_scalar="False" rank_dropout_scale="False" algo="locon" train_norm="False" --network_dim=8 --network_module=lycoris.kohya --no_half_vae --optimizer_type="Prodigy" --output_dir="c:/training_data" --output_name="{outdir}/{name}{suffix}" --pretrained_model_name_or_path="D:/ai/models/Stable-diffusion/!!SDXL/!EnvyStarlightXL01+vae.fp16.safetensors" --save_every_n_epochs="1" --save_model_as=safetensors --save_precision="bf16" --scale_weight_norms="1" --text_encoder_lr=1 --train_batch_size="12" --train_data_dir="{directory}/{name}" --unet_lr=1 --xformers
It may need to be altered slightly for Linux. Note that each epoch should have a hundred images or so (repeats are fine -- it can work on as little as one image).