525 lines
15 KiB
Plaintext
525 lines
15 KiB
Plaintext
{
|
|
"cells": [
|
|
{
|
|
"cell_type": "markdown",
|
|
"id": "aa2c1ada",
|
|
"metadata": {
|
|
"id": "aa2c1ada"
|
|
},
|
|
"source": [
|
|
"# Dreambooth\n",
|
|
"### Notebook implementation by Joe Penna (@MysteryGuitarM on Twitter) - Improvements by David Bielejeski\n",
|
|
"\n",
|
|
"### Instructions\n",
|
|
"- Sign up for RunPod here: https://runpod.io/?ref=n8yfwyum\n",
|
|
" - Note: That's my personal referral link. Please don't use it if we are mortal enemies.\n",
|
|
"\n",
|
|
"- Click *Deploy* on either `SECURE CLOUD` or `COMMUNITY CLOUD`\n",
|
|
"\n",
|
|
"- Follow the rest of the instructions in this video: https://www.youtube.com/watch?v=7m__xadX0z0#t=5m33.1s\n",
|
|
"\n",
|
|
"Latest information on:\n",
|
|
"https://github.com/JoePenna/Dreambooth-Stable-Diffusion"
|
|
]
|
|
},
|
|
{
|
|
"cell_type": "markdown",
|
|
"id": "7b971cc0",
|
|
"metadata": {
|
|
"id": "7b971cc0"
|
|
},
|
|
"source": [
|
|
"## Build Environment"
|
|
]
|
|
},
|
|
{
|
|
"cell_type": "code",
|
|
"execution_count": null,
|
|
"id": "2AsGA1xpNQnb",
|
|
"metadata": {
|
|
"id": "2AsGA1xpNQnb"
|
|
},
|
|
"outputs": [],
|
|
"source": [
|
|
"# If running on Vast.AI, copy the code in this cell into a new notebook. Run it, then launch the `dreambooth_runpod_joepenna.ipynb` notebook from the jupyter interface.\n",
|
|
"!git clone https://github.com/JoePenna/Dreambooth-Stable-Diffusion"
|
|
]
|
|
},
|
|
{
|
|
"cell_type": "code",
|
|
"execution_count": null,
|
|
"id": "9e1bc458-091b-42f4-a125-c3f0df20f29d",
|
|
"metadata": {
|
|
"id": "9e1bc458-091b-42f4-a125-c3f0df20f29d",
|
|
"scrolled": true
|
|
},
|
|
"outputs": [],
|
|
"source": [
|
|
"#BUILD ENV\n",
|
|
"!pip install omegaconf\n",
|
|
"!pip install einops\n",
|
|
"!pip install pytorch-lightning==1.6.5\n",
|
|
"!pip install test-tube\n",
|
|
"!pip install transformers\n",
|
|
"!pip install kornia\n",
|
|
"!pip install -e git+https://github.com/CompVis/taming-transformers.git@master#egg=taming-transformers\n",
|
|
"!pip install -e git+https://github.com/openai/CLIP.git@main#egg=clip\n",
|
|
"!pip install setuptools==59.5.0\n",
|
|
"!pip install pillow==9.0.1\n",
|
|
"!pip install torchmetrics==0.6.0\n",
|
|
"!pip install -e .\n",
|
|
"!pip install protobuf==3.20.1\n",
|
|
"!pip install gdown\n",
|
|
"!pip install pydrive\n",
|
|
"!pip install -qq diffusers[\"training\"]==0.3.0 transformers ftfy\n",
|
|
"!pip install -qq \"ipywidgets>=7,<8\"\n",
|
|
"!pip install huggingface_hub\n",
|
|
"!pip install ipywidgets"
|
|
]
|
|
},
|
|
{
|
|
"cell_type": "code",
|
|
"execution_count": null,
|
|
"id": "dae11c10",
|
|
"metadata": {
|
|
"id": "dae11c10"
|
|
},
|
|
"outputs": [],
|
|
"source": [
|
|
"## Login to stable diffusion\n",
|
|
"from huggingface_hub import notebook_login\n",
|
|
"\n",
|
|
"notebook_login()"
|
|
]
|
|
},
|
|
{
|
|
"cell_type": "code",
|
|
"execution_count": null,
|
|
"metadata": {
|
|
"collapsed": false
|
|
},
|
|
"outputs": [],
|
|
"source": [
|
|
"## Download the 1.4 sd model\n",
|
|
"from huggingface_hub import hf_hub_download\n",
|
|
"downloaded_model_path = hf_hub_download(\n",
|
|
" repo_id=\"CompVis/stable-diffusion-v-1-4-original\",\n",
|
|
" filename=\"sd-v1-4.ckpt\",\n",
|
|
" use_auth_token=True\n",
|
|
")"
|
|
]
|
|
},
|
|
{
|
|
"cell_type": "code",
|
|
"execution_count": null,
|
|
"metadata": {
|
|
"collapsed": false
|
|
},
|
|
"outputs": [],
|
|
"source": [
|
|
"## Move the sd-v1-4.ckpt to the root of this directory as \"model.ckpt\"\n",
|
|
"actual_locations_of_model_blob = !readlink -f {downloaded_model_path}\n",
|
|
"!mv {actual_locations_of_model_blob[-1]} model.ckpt"
|
|
]
|
|
},
|
|
{
|
|
"cell_type": "markdown",
|
|
"id": "17d1d11a",
|
|
"metadata": {
|
|
"id": "17d1d11a"
|
|
},
|
|
"source": [
|
|
"# Regularization Images (Skip this section if you are uploading your own or using the provided images)"
|
|
]
|
|
},
|
|
{
|
|
"cell_type": "markdown",
|
|
"id": "ed07a5df",
|
|
"metadata": {
|
|
"id": "ed07a5df"
|
|
},
|
|
"source": [
|
|
"Training teaches your new model both your token **but** re-trains your class simultaneously.\n",
|
|
"\n",
|
|
"From cursory testing, it does not seem like reg images affect the model too much. However, they do affect your class greatly, which will in turn affect your generations.\n",
|
|
"\n",
|
|
"You can either generate your images here, or use the repos below to quickly download 1500 images."
|
|
]
|
|
},
|
|
{
|
|
"cell_type": "code",
|
|
"execution_count": null,
|
|
"id": "67f9ff0c-b529-4c7c-8e26-8388d70a5d91",
|
|
"metadata": {
|
|
"id": "67f9ff0c-b529-4c7c-8e26-8388d70a5d91"
|
|
},
|
|
"outputs": [],
|
|
"source": [
|
|
"# GENERATE 200 images\n",
|
|
"!python scripts/stable_txt2img.py \\\n",
|
|
" --seed 10 \\\n",
|
|
" --ddim_eta 0.0 \\\n",
|
|
" --n_samples 1 \\\n",
|
|
" --n_iter 200 \\\n",
|
|
" --scale 10.0 \\\n",
|
|
" --ddim_steps 50 \\\n",
|
|
" --ckpt model.ckpt \\\n",
|
|
" --prompt \"person\"\n",
|
|
"\n",
|
|
"# If you don't want to train against \"person\", change it to whatever you want above, and on some of the cells below:"
|
|
]
|
|
},
|
|
{
|
|
"cell_type": "code",
|
|
"execution_count": null,
|
|
"id": "3d1c7e1c",
|
|
"metadata": {
|
|
"id": "3d1c7e1c"
|
|
},
|
|
"outputs": [],
|
|
"source": [
|
|
"# zip up the files for downloading and reuse.\n",
|
|
"!apt-get install -y zip\n",
|
|
"!zip -r all_images.zip outputs/\n",
|
|
"\n",
|
|
"# Download this file locally so you can reuse during another training on this dataset"
|
|
]
|
|
},
|
|
{
|
|
"cell_type": "markdown",
|
|
"id": "mxPL2O0OLvBW",
|
|
"metadata": {
|
|
"id": "mxPL2O0OLvBW"
|
|
},
|
|
"source": [
|
|
"## Download pre-generated regularization images\n",
|
|
"\n",
|
|
"We've created the following image sets\n",
|
|
"\n",
|
|
"* man_euler - provided by Niko Pueringer (Corridor Digital) - euler @ 40 steps, CFG 7.5\n",
|
|
"* man_unsplash - pictures from various photographers\n",
|
|
"* person_ddim\n",
|
|
"* woman_ddim - provided by David Bielejeski - ddim @ 50 steps, CFG 10.0\n",
|
|
"\n",
|
|
"`person_ddim` is recommended"
|
|
]
|
|
},
|
|
{
|
|
"cell_type": "code",
|
|
"execution_count": 1,
|
|
"id": "e7EydXCjOV1v",
|
|
"metadata": {
|
|
"id": "e7EydXCjOV1v"
|
|
},
|
|
"outputs": [],
|
|
"source": [
|
|
"# Grab the existing regularization images\n",
|
|
"# Choose the dataset that best represents what you are trying to do and matches what you used for your token\n",
|
|
"# man_euler, man_unsplash, person_ddim, woman_ddim\n",
|
|
"dataset=\"person_ddim\"\n",
|
|
"!git clone https://github.com/djbielejeski/Stable-Diffusion-Regularization-Images-{dataset}.git\n",
|
|
"\n",
|
|
"!mkdir -p outputs/txt2img-samples/samples/{dataset}\n",
|
|
"!mv -v Stable-Diffusion-Regularization-Images-{dataset}/{dataset}/*.* outputs/txt2img-samples/samples/{dataset}"
|
|
]
|
|
},
|
|
{
|
|
"cell_type": "markdown",
|
|
"id": "zshrC_JuMXmM",
|
|
"metadata": {
|
|
"id": "zshrC_JuMXmM"
|
|
},
|
|
"source": [
|
|
"# Upload your training images\n",
|
|
"Upload 10-20 images of someone to\n",
|
|
"\n",
|
|
"```\n",
|
|
"/workspace/Dreambooth-Stable-Diffusion/training_samples\n",
|
|
"```\n",
|
|
"\n",
|
|
"WARNING: Be sure to upload an *even* amount of images, otherwise the training inexplicably stops at 1500 steps.\n",
|
|
"\n",
|
|
"* 2-3 full body\n",
|
|
"* 3-5 upper body \n",
|
|
"* 5-12 close-up on face\n",
|
|
"\n",
|
|
"The images should be:\n",
|
|
"\n",
|
|
"- as close as possible to the kind of images you're trying to make (most of the time, that means no selfies).\n",
|
|
"- "
|
|
]
|
|
},
|
|
{
|
|
"cell_type": "code",
|
|
"execution_count": null,
|
|
"metadata": {
|
|
"collapsed": false
|
|
},
|
|
"outputs": [],
|
|
"source": [
|
|
"#@markdown Add here the URLs to the images of the subject you are adding\n",
|
|
"urls = [\n",
|
|
" \"https://i.imgur.com/test1.png\",\n",
|
|
" \"https://i.imgur.com/test2.png\",\n",
|
|
" \"https://i.imgur.com/test3.png\",\n",
|
|
" \"https://i.imgur.com/test4.png\",\n",
|
|
" \"https://i.imgur.com/test5.png\",\n",
|
|
" # You can add additional images here -- about 20-30 images in different \n",
|
|
"]"
|
|
]
|
|
},
|
|
{
|
|
"cell_type": "code",
|
|
"execution_count": null,
|
|
"metadata": {
|
|
"collapsed": false
|
|
},
|
|
"outputs": [],
|
|
"source": [
|
|
"#@title Download and check the images you have just added\n",
|
|
"import os\n",
|
|
"import requests\n",
|
|
"from io import BytesIO\n",
|
|
"from PIL import Image\n",
|
|
"\n",
|
|
"\n",
|
|
"def image_grid(imgs, rows, cols):\n",
|
|
" assert len(imgs) == rows*cols\n",
|
|
"\n",
|
|
" w, h = imgs[0].size\n",
|
|
" grid = Image.new('RGB', size=(cols*w, rows*h))\n",
|
|
" grid_w, grid_h = grid.size\n",
|
|
"\n",
|
|
" for i, img in enumerate(imgs):\n",
|
|
" grid.paste(img, box=(i%cols*w, i//cols*h))\n",
|
|
" return grid\n",
|
|
"\n",
|
|
"def download_image(url):\n",
|
|
" try:\n",
|
|
" response = requests.get(url)\n",
|
|
" except:\n",
|
|
" return None\n",
|
|
" return Image.open(BytesIO(response.content)).convert(\"RGB\")\n",
|
|
"\n",
|
|
"images = list(filter(None,[download_image(url) for url in urls]))\n",
|
|
"save_path = \"./training_samples\"\n",
|
|
"if not os.path.exists(save_path):\n",
|
|
" os.mkdir(save_path)\n",
|
|
"[image.save(f\"{save_path}/{i}.png\", format=\"png\") for i, image in enumerate(images)]\n",
|
|
"image_grid(images, 1, len(images))"
|
|
]
|
|
},
|
|
{
|
|
"cell_type": "markdown",
|
|
"id": "ad4e50df",
|
|
"metadata": {
|
|
"id": "ad4e50df"
|
|
},
|
|
"source": [
|
|
"## Training\n",
|
|
"\n",
|
|
"If training a person or subject, keep an eye on your project's `logs/{folder}/images/train/samples_scaled_gs-00xxxx` generations.\n",
|
|
"\n",
|
|
"If training a style, keep an eye on your project's `logs/{folder}/images/train/samples_gs-00xxxx` generations."
|
|
]
|
|
},
|
|
{
|
|
"cell_type": "markdown",
|
|
"metadata": {
|
|
"collapsed": false
|
|
},
|
|
"source": [
|
|
"## Edit the personalized.py file\n",
|
|
"Execute this cell `%load ldm/data/personalized.py`\n",
|
|
"\n",
|
|
"Change `joepenna` to whatever you want it to be (but keep the {})\n",
|
|
"\n",
|
|
"```\n",
|
|
"training_templates_smallest = [\n",
|
|
" 'joepenna {}',\n",
|
|
"]\n",
|
|
"```\n",
|
|
"\n",
|
|
"I recommend using the name of a celebrity that:\n",
|
|
"1) kinda looks like you.\n",
|
|
"2) Stable Diffusion generates well (you can check by typing their name on DreamStudio)\n",
|
|
"\n",
|
|
"Then paste this at the very top of the cell:\n",
|
|
"```\n",
|
|
"%%writefile ldm/data/personalized.py\n",
|
|
"```\n",
|
|
"\n",
|
|
"Then run the cell again. This will save your changes.\n"
|
|
]
|
|
},
|
|
{
|
|
"cell_type": "code",
|
|
"execution_count": null,
|
|
"metadata": {
|
|
"collapsed": false
|
|
},
|
|
"outputs": [],
|
|
"source": [
|
|
"%load ldm/data/personalized.py"
|
|
]
|
|
},
|
|
{
|
|
"cell_type": "code",
|
|
"execution_count": null,
|
|
"id": "6fa5dd66-2ca0-4819-907e-802e25583ae6",
|
|
"metadata": {
|
|
"id": "6fa5dd66-2ca0-4819-907e-802e25583ae6",
|
|
"tags": []
|
|
},
|
|
"outputs": [],
|
|
"source": [
|
|
"# START THE TRAINING\n",
|
|
"project_name = \"project_name\"\n",
|
|
"\n",
|
|
"# oops, this next variable should be called \"max_steps\".\n",
|
|
"# It's how long you want your training to go.\n",
|
|
"# If you're seeing this message, I'm literally at my computer right now fixing this up:\n",
|
|
"batch_size = 1000\n",
|
|
"\n",
|
|
"class_word = \"person\" # << match this word to the class word from regularization images above\n",
|
|
"reg_data_root = \"/workspace/Dreambooth-Stable-Diffusion/outputs/txt2img-samples/samples/\" + dataset\n",
|
|
"\n",
|
|
"!rm -rf training_samples/.ipynb_checkpoints\n",
|
|
"!python \"main.py\" \\\n",
|
|
" --base configs/stable-diffusion/v1-finetune_unfrozen.yaml \\\n",
|
|
" -t \\\n",
|
|
" --actual_resume \"model.ckpt\" \\\n",
|
|
" --reg_data_root {reg_data_root} \\\n",
|
|
" -n {project_name} \\\n",
|
|
" --gpus 0, \\\n",
|
|
" --data_root \"/workspace/Dreambooth-Stable-Diffusion/training_samples\" \\\n",
|
|
" --batch_size {batch_size} \\\n",
|
|
" --class_word class_word"
|
|
]
|
|
},
|
|
{
|
|
"cell_type": "markdown",
|
|
"id": "dc49d0bd",
|
|
"metadata": {},
|
|
"source": [
|
|
"## Pruning (12GB to 2GB)\n",
|
|
"We are working on having this happen automatically (TODO: PR's welcome)"
|
|
]
|
|
},
|
|
{
|
|
"cell_type": "code",
|
|
"execution_count": null,
|
|
"metadata": {
|
|
"collapsed": false
|
|
},
|
|
"outputs": [],
|
|
"source": [
|
|
"directory_paths = !ls -d logs/*"
|
|
]
|
|
},
|
|
{
|
|
"cell_type": "code",
|
|
"execution_count": null,
|
|
"metadata": {
|
|
"collapsed": false
|
|
},
|
|
"outputs": [],
|
|
"source": [
|
|
"# This version should automatically prune around 10GB from the ckpt file\n",
|
|
"last_checkpoint_file = directory_paths[-1] + \"/checkpoints/last.ckpt\"\n",
|
|
"!python \"prune_ckpt.py\" --ckpt {last_checkpoint_file}"
|
|
]
|
|
},
|
|
{
|
|
"cell_type": "code",
|
|
"execution_count": null,
|
|
"metadata": {
|
|
"collapsed": false
|
|
},
|
|
"outputs": [],
|
|
"source": [
|
|
"last_checkpoint_file_pruned = directory_paths[-1] + \"/checkpoints/last-pruned.ckpt\"\n",
|
|
"training_samples = !ls training_samples\n",
|
|
"date_string = !date +\"%Y-%m-%dT%H-%M-%S\"\n",
|
|
"file_name = date_string[-1] + \"_\" + project_name + \"_\" + str(len(training_samples)) + \"_training_images_\" + str(batch_size) + \"_batch_size_\" + class_word + \"_class_word.ckpt\"\n",
|
|
"!mkdir -p trained_models\n",
|
|
"!mv {last_checkpoint_file_pruned} trained_models/{file_name}"
|
|
]
|
|
},
|
|
{
|
|
"cell_type": "code",
|
|
"execution_count": null,
|
|
"metadata": {},
|
|
"outputs": [],
|
|
"source": [
|
|
"# Download your trained model file from `trained_models` and use in your favorite Stable Diffusion repo!"
|
|
]
|
|
},
|
|
{
|
|
"cell_type": "markdown",
|
|
"id": "9a90ac5c",
|
|
"metadata": {},
|
|
"source": [
|
|
"# Big Important Note!\n",
|
|
"\n",
|
|
"The way to call your token"
|
|
]
|
|
},
|
|
{
|
|
"cell_type": "markdown",
|
|
"id": "d28d0139",
|
|
"metadata": {},
|
|
"source": [
|
|
"## Generate Images With Your Trained Model!"
|
|
]
|
|
},
|
|
{
|
|
"cell_type": "code",
|
|
"execution_count": null,
|
|
"id": "80ddb03b",
|
|
"metadata": {},
|
|
"outputs": [],
|
|
"source": [
|
|
"!python scripts/stable_txt2img.py \\\n",
|
|
" --ddim_eta 0.0 \\\n",
|
|
" --n_samples 1 \\\n",
|
|
" --n_iter 4 \\\n",
|
|
" --scale 7.0 \\\n",
|
|
" --ddim_steps 50 \\\n",
|
|
" --ckpt \"/workspace/Dreambooth-Stable-Diffusion/trained_models/\" + {file_name} \\\n",
|
|
" --prompt \"joepenna person as a masterpiece portrait painting by John Singer Sargent in the style of Rembrandt\""
|
|
]
|
|
}
|
|
],
|
|
"metadata": {
|
|
"colab": {
|
|
"collapsed_sections": [],
|
|
"provenance": []
|
|
},
|
|
"kernelspec": {
|
|
"display_name": "Python 3 (ipykernel)",
|
|
"language": "python",
|
|
"name": "python3"
|
|
},
|
|
"language_info": {
|
|
"codemirror_mode": {
|
|
"name": "ipython",
|
|
"version": 3
|
|
},
|
|
"file_extension": ".py",
|
|
"mimetype": "text/x-python",
|
|
"name": "python",
|
|
"nbconvert_exporter": "python",
|
|
"pygments_lexer": "ipython3",
|
|
"version": "3.10.6"
|
|
},
|
|
"vscode": {
|
|
"interpreter": {
|
|
"hash": "b0fa6594d8f4cbf19f97940f81e996739fb7646882a419484c72d19e05852a7e"
|
|
}
|
|
}
|
|
},
|
|
"nbformat": 4,
|
|
"nbformat_minor": 5
|
|
}
|