search

p1atdev / LECO

Low-rank adaptation for Erasing COncepts from diffusion models.
https://arxiv.org/abs/2303.07345
Apache License 2.0
307 stars 23 forks source link
diffusion lora stable-diffusion

readme

LECO ✏️

Open In Colab

Low-rank adaptation for Erasing COncepts from diffusion models.

The original repository: Erasing Concepts from Diffusion Models

and the project page: https://erasing.baulab.info/

(Not only for erasing concepts, but also emphasizing or swapping them by devising prompts and LoRA weight. See ConceptMod for more details)

Setup

conda create -n leco python=3.10
pip install torch torchvision --index-url https://download.pytorch.org/whl/cu118
pip install xformers
pip install -r requirements.txt

Train

You need 8GB VRAM at least.

python ./train_lora.py --config_file "./examples/config.yaml"

config.yaml:

prompts_file: "./prompts.yaml"

pretrained_model:
  name_or_path: "stabilityai/stable-diffusion-2-1" # you can also use .ckpt or .safetensors models
  v2: true # true if model is v2.x
  v_pred: true # true if model uses v-prediction

network:
  type: "lierla" # or "c3lier"
  rank: 4
  alpha: 1.0

train:
  precision: "bfloat16"
  noise_scheduler: "ddim" # or "ddpm", "lms", "euler_a"
  iterations: 500
  lr: 1e-4
  optimizer: "AdamW"
  lr_scheduler: "constant"

save:
  name: "van_gogh"
  path: "./output"
  per_steps: 200
  precision: "bfloat16"

logging:
  use_wandb: false
  verbose: false

other:
  use_xformers: true

prompts.yaml:

- target: "van gogh" # what word for erasing the positive concept from
  positive: "van gogh" # concept to erase
  unconditional: "" # word to take the difference from the positive concept
  neutral: "" # starting point for conditioning the target
  action: "erase" # erase or enhance
  guidance_scale: 1.0
  resolution: 512
  dynamic_resolution: false
  batch_size: 2

See the example config for more details.

Note: You can use float16 but it is unstable and not recommended. Please use bfloat16 or float32.

Pretrained weights

You can use the pretrained weights on AUTOMATIC1111's webui.

🤗 HuggingFace: https://huggingface.co/p1atdev/leco

SDv1.5

Results of oil painting of van gogh by himself:

van gogh

Generation settings ```yaml oil painting of van gogh by himself Steps: 20, Sampler: Euler a, CFG scale: 7, Seed: 3870472781, Size: 512x512, Model hash: cc6cb27103, Model: v1-5-pruned-emaonly, Clip skip: 2, AddNet Enabled: True, AddNet Module 1: LoRA, AddNet Model 1: van_gogh_4_last(db68853d039b), AddNet Weight A 1: -1.0, AddNet Weight B 1: -1.0, Script: X/Y/Z plot, X Type: AddNet Weight 1, X Values: "-1, 0, 1", Version: v1.3.0 ```

Results of painting of scenery by monet:

monet

Generation settings ```yaml painting of scenery by monet Steps: 20, Sampler: Euler a, CFG scale: 7, Seed: 1284787312, Size: 512x512, Model hash: cc6cb27103, Model: v1-5-pruned-emaonly, Clip skip: 2, AddNet Enabled: True, AddNet Module 1: LoRA, AddNet Model 1: van_gogh_4_last(db68853d039b), AddNet Weight A 1: -1.0, AddNet Weight B 1: -1.0, Script: X/Y/Z plot, X Type: AddNet Weight 1, X Values: "-1, 0, 1", Version: v1.3.0 ```

SDv2.1-768

Results of mona lisa with jewelry:

mona lisa

Generation settings ```yaml mona lisa with jewelry Steps: 20, Sampler: Euler a, CFG scale: 7, Seed: 3630495347, Size: 512x512, Model hash: 832eb50c0c, Model: v2-1_768-ema-pruned, Clip skip: 2, AddNet Enabled: True, AddNet Module 1: LoRA, AddNet Model 1: mona_lisa2_last(393beb35c4b1), AddNet Weight A 1: -1.0, AddNet Weight B 1: -1.0, Script: X/Y/Z plot, X Type: AddNet Weight 1, X Values: "-1, 0, 1", Version: v1.3.0 ```

Results of photo of a cute cat:

mona lisa

Generation settings ```yaml photo of a cute cat Steps: 20, Sampler: Euler a, CFG scale: 7, Seed: 900866192, Size: 512x512, Model hash: 832eb50c0c, Model: v2-1_768-ema-pruned, Clip skip: 2, AddNet Enabled: True, AddNet Module 1: LoRA, AddNet Model 1: mona_lisa2_last(393beb35c4b1), AddNet Weight A 1: -1.0, AddNet Weight B 1: -1.0, Script: X/Y/Z plot, X Type: AddNet Weight 1, X Values: "-1, 0, 1", Version: v1.3.0 ```

WD1.5 beta3

Cat ears will be attached forcibly when using with 1.0~3.0 weight.

If -1.0~-3.0, cat ears will never appear.

Training settings: see configs.

cat ears

Generation settings ```yaml masterpiece, best quality, exceptional, best aesthetic, anime, 1girl, school uniform, upper body, smile Negative prompt: worst quality, low quality, bad aesthetic, oldest, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, jpeg artifacts, signature, watermark, username, blurry Steps: 20, Sampler: Euler a, CFG scale: 7, Seed: 4103955758, Size: 512x512, Model hash: d38e779546, Model: wd-beta3-base-fp16, Clip skip: 2, Script: X/Y/Z plot, X Type: AddNet Weight 1, X Values: "0, 2, 3, 4", Version: v1.3.0 ```

Training settings: see configs.

With "real life, instagram":

unreal

Generation settings ```yaml real life, instagram, masterpiece, best quality, exceptional, best aesthetic, 1girl, cat ears, blue hair, school uniform, upper body Negative prompt: worst quality, low quality, bad aesthetic, oldest, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, jpeg artifacts, signature, watermark, username, blurry Steps: 20, Sampler: Euler a, CFG scale: 7, Seed: 757542759, Size: 768x768, Model hash: d38e779546, Model: wd-beta3-base-fp16, Clip skip: 2, AddNet Enabled: True, AddNet Module 1: LoRA, AddNet Model 1: unreal_6_many_prompts_200steps(fff5917285da), AddNet Weight A 1: -1.0, AddNet Weight B 1: -1.0, Script: X/Y/Z plot, X Type: AddNet Weight 1, X Values: "-1, 0, 1", Version: v1.3.0 ```

Without "real life, instagram":

unreal

Generation settings ```yaml masterpiece, best quality, exceptional, best aesthetic,, 1girl, aqua eyes, baseball cap, blonde hair, closed mouth, earrings, green background, hat, hoop earrings, jewelry, looking at viewer, shirt, short hair, simple background, solo, upper body, yellow shirt, Negative prompt: worst quality, low quality, bad aesthetic, oldest, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, jpeg artifacts, signature, watermark, username, blurry Steps: 20, Sampler: Euler a, CFG scale: 7, Seed: 2867636749, Size: 768x768, Model hash: d38e779546, Model: wd-beta3-base-fp16, Clip skip: 2, AddNet Enabled: True, AddNet Module 1: LoRA, AddNet Model 1: unreal_6_many_prompts_200steps(fff5917285da), AddNet Weight A 1: -1.0, AddNet Weight B 1: -1.0, Script: X/Y/Z plot, X Type: AddNet Weight 1, X Values: "-1, 0, 1", Version: v1.3.0 ```

References

I am deeply inspired by and my work relies on the outstanding efforts of the following projects. I want to express my profound gratitude to these projects and their developers: