/lora-training

LoRA training model packaged with Cog

Primary LanguagePythonMIT LicenseMIT

LoRA training Cog model

Use on Replicate

Easy-to-use model pre-configured for faces, objects, and styles:

Replicate

Advanced model with all the parameters:

Replicate

Feed the trained model into this inference model to run predictions:

Replicate

If you want to share your trained LoRAs, please join the #lora channel in the Replicate Discord.

Use locally

First, download the pre-trained weights with your Hugging Face auth token:

cog run script/download-weights <your-hugging-face-auth-token>

Then, you can run train your dreambooth:

cog predict -i instance_data=@my-images.zip

The resulting LoRA weights file can be used with patch_pipe function:

from diffusers import StableDiffusionPipeline
from lora_diffusion import patch_pipe, tune_lora_scale, image_grid
import torch

model_id = "runwayml/stable-diffusion-v1-5"

pipe = StableDiffusionPipeline.from_pretrained(model_id, torch_dtype=torch.float16).to(
    "cuda:1"
)

patch_pipe(pipe, "./my-images.safetensors")
prompt = "detailed photo of <s1><s2>, detailed face, a brown cloak, brown steampunk corset, belt, virtual youtuber, cowboy shot, feathers in hair, feather hair ornament, white shirt, brown gloves, shooting arrows"

tune_lora_scale(pipe.unet, 0.8)
tune_lora_scale(pipe.text_encoder, 0.8)

imgs = pipe(
    [prompt],
    num_inference_steps=50,
    guidance_scale=4.5,
    height=640,
    width=512,
).images
...