Skip to content
This repository has been archived by the owner on Mar 23, 2024. It is now read-only.

Latest commit

 

History

History
202 lines (176 loc) · 27.2 KB

MANUAL.md

File metadata and controls

202 lines (176 loc) · 27.2 KB
Raw

Cagliostro Colab UI User Manual

Welcome to the Cagliostro Colab UI User Manual! This guide will walk you through the basics of using Cagliostro Colab UI, an innovative and powerful notebook designed to launch Automatic1111's Stable Diffusion Web UI in Google Colab. With its advanced features, customizability, and flexibility, Cagliostro Colab offers a seamless and efficient way to utilize Stable Diffusion Web UI for your projects.

To get started with Cagliostro Colab UI, you'll need to have a Google account. Once you're logged in, you can open up Google Colab by visiting the website and signing in with your Google credentials.

Table of Contents

Features:

Feature Description
Faster installation Unpacks a pre-installed repository instead of installing from scratch, saving time and reducing the risk of installation errors.Uses pre-installed Python dependencies and unpacks them to /usr/local/lib/python3.9/dist-packages/ instead of downloading them during installation, further reducing installation time.
Faster downloader Uses the aria2c downloader instead of wget.Downloads packages with up to 16 parallel connections, 16 threads per connection, and a 1MB chunk size for faster downloads.Used to download pre-installed repo, dependencies, and models.
Better UI/UX Uses Anapnoe's forked version of Automatic1111's Stable Diffusion Web UI that focuses on developing better UI/UX for Stable Diffusion Web UI. But users can still use Automatic1111 by disabling the use_anapnoe_ui option.Note that Anapnoe's still uses an old commit, so if users want to experience Gradio 3.23.0, it's better to disable the use_anapnoe_ui option.
Up-to-date Enabled with git pull for both the repository and all extensions by default to increase user experience.Some extensions might be skipped from updating due to upgrading its Gradio to 3.23.
Integrated to Google Drive Provides a way to enhance the user experience for Google Drive users.Includes options to mount Google Drive, save output to Google Drive, load the model directory from Google Drive by using unionfs-fuse to merge folder A and folder B to folder C, and load the model file from Google Drive by copying the model to Automatic1111's Stable Diffusion Web UI.
Colab Optimization Loads the model in VRAM, merges in VRAM, and uses VRAM instead of CPU for most things to create a better environment for Colab free tier users.
Better Default Model Provides the top trending default model and the best choice in quality.Chooses Huggingface server instead of CivitAI because it's more stable and faster.
All-in-One ControlNet model Provides not only the ControlNet model from SD1.5 but also SD2.1.Users can set the max model, config, and adapter config before starting the Web UI.
Advanced Custom Download Provides seven fields to download models, including custom models, VAE, embedding, LoRA, hypernetwork, control, and extension.Users can prune their model by typing fp16:url to prune with FP16 precision and fp32:url for FP32 precision.Users can also easily load models from Google Drive by merging folders with the fuse:path command.
More Tunnels! Uses sd-webui-tunnels extension (forked by Camenduru) and set --multiple by default. No more boring gradio.live!We're also support ngrok.

Main Cell Explained

Install Stable Diffusion Web UI

This cell installs stable-diffusion-webui repository in Colab. It also includes several configuration options, such as mounting Google Drive, updating extensions, and choosing between different versions of the web UI.

Options Default Value Description
Drive Config
mount_drive False Mount your Google Drive to /content/drive/MyDrive/ to load models.
output_to_drive False Save generation outputs to your Google Drive instead of /content/stable-diffusion-webui/outputs.
output_drive_folder False Set your default drive folder name to save your generation outputs, default: cagliostro-colab-ui/outputs
Web UI Config
use_anapnoe_ui True Use Anapnoe's forked repository instead of Automatic1111's.
update_webui True Update the web UI to the latest version. Because we're using pre-installed repo, this option is crucial if you want to try latest update.
update_extensions True Update all extensions to the latest version.
commit_hash '' Go back to a specific commit hash to prevent errors when the web UI is updated.
colab_optimizations False Load SDv2 the model into VRAM, optimizes demo.queue() for Colab, and loads the model onto GPU memory if available.

Note:

  • The options colab_optimizations are not relevant anymore as the new argument --lowram serves the same purpose, allowing the model to be loaded onto the GPU memory instead of VRAM if available. It also not recommended to set this True if you have Colab Pro subscription.

Download Model and VAE

This code block is responsible for downloading pre-trained models and VAEs (Variational Autoencoders) from Hugging Face's model hub. The available models and VAEs are listed as boolean checkboxes, and the user can select the ones they want to download.

Model Name Description Model Page
Stable Diffusion V1.x Model
anything_v3_0 State-of-the-art merged/overfitted/vae broken model when it was first launched in November, it has influenced most anime models since then. Link
anime_pastel_dream The best alternative model to pastelmix, anime_pastel_dream was developed to provide a better user experience. Link
anylora (default) The best alternative model to anything_v4_5, anylora was developed to provide a better user experience. Link
chilloutmix_ni The most used non-anime model for now, it generates real person pictures but is still influenced by anime models. Link
Stable Diffusion V2.x Model
replicant_v2 An anime model based off Waifu Diffusion V1.5 Beta 2, which performs better than waifu_diffusion_v1_5_e2_aesthetic. Link
waifu_diffusion_v1_5_e2_aesthetic An overtrained version of Waifu Diffusion V1.5 Beta 2 trained off stable diffusion v2.1 768. It works great to generate anime art. Link
VAE models
anime (default) A popular anime VAE used by many model, such as AbyssOrangeMixs and Anything V3. 🔐
waifu_diffusion Anime VAE developed by the Waifu Diffusion developer based on the stable diffusion VAE. It provides contrast color compared to the anime VAE. Link
stable_diffusion The original Stable Diffusion VAE trained by Stability AI. It is intended to be used with the original CompVis Stable Diffusion. Link

ControlNet V1.1

ControlNet is a neural network structure to control diffusion models by adding extra conditions. ControlNet was introduced in Adding Conditional Control to Text-to-Image Diffusion Models by Lvmin Zhang and Maneesh Agrawala. It introduces a framework that allows for supporting various spatial contexts that can serve as additional conditionings to Diffusion models such as Stable Diffusion. It basically allow user to control Stable Diffusion generation. ControlNet 1.1 has been released nightly at this github repository and Cagliostro Colab UI is now support it!

Option Default Description
pre_download_annotator True Specifies whether to download ControlNet pre-processor/annotator before starting the Web UI. The pre-processor/annotator includes 13 items in total, and will be downloaded to the specified path.
control_v11_sd15_model True Specifies whether to download all 14 extracted ControlNet v1.1 model developed by Lvmin Zhang
t2i_adapter_sd15_model False Specifies whether to download all 12 Text to Image Adapter model developed by TencentArcLab.
control_v10_sd21_model False Specifies whether to download all 5 extracted SDv21 ControlNet v1.0 model developed by thibaud.
control_v10_wd15_model False Specifies whether to download all 3 extracted Waifu Diffusion 1.5 Controlnet v1.0 model developed by furusu.
control_net_max_models_num 2 Specifies the maximum number of ControlNet tabs for multi-controlnet generation.
control_net_model_adapter_config sketch_adapter_v14.yaml Specifies the configuration file to use for the ControlNet model adapter. Different models require different configurations. The default configuration is sketch_adapter_v14.yaml.

T2I Adapter is a simple and small (~70M parameters, ~300M storage space) network that can provide extra guidance to pre-trained text-to-image models while freezing the original large text-to-image models. Controlnet was introduced in T2I-Adapter: Learning Adapters to Dig out More Controllable Ability for Text-to-Image Diffusion Models. It basically a cheapest ControlNet that allow user to control Text to Image generation.

t2i_adapter_model control_net_model_adapter_config
t2iadapter_canny_sd14v1.pth sketch_adapter_v14.yaml
t2iadapter_sketch_sd14v1.pth sketch_adapter_v14.yaml
t2iadapter_seg_sd14v1.pth image_adapter_v14.yaml
t2iadapter_keypose_sd14v1.pth image_adapter_v14.yaml
t2iadapter_openpose_sd14v1.pth image_adapter_v14.yaml
t2iadapter_color_sd14v1.pth t2iadapter_color_sd14v1.yaml
t2iadapter_style_sd14v1.pth t2iadapter_style_sd14v1.yaml

Custom Download Corner

This is a cell that allows you to download custom models, VAEs, embeddings, LoRA, hypernetworks, upscalee and install extensions by providing URLs to the files you want to download. This cell downloads custom files from various sources, including Google Drive, Huggingface, CivitAI, and other direct download links.

Feature Description How to Use Example
Multiple Downloads Download multiple files at once. Fill in the URL fields with the links to the files you want to download. Separate multiple URLs with a comma. url1, url2, url3
Auto-prune Prune models after downloading Add fp16: or fp32: before URLs. fp16:url1
Copy from Google Drive Copy models from Google Drive and load them in the session. Make sure you have already mounted Google Drive. Type the path to your model/lora/embedding from Google Drive. /content/drive/MyDrive/path/to/folder
Fusing Folder Fuse models/embeddings/LoRA folder to /content/fused/{category}. Make sure you have already mounted Google Drive. Add fuse: before the path to the folder. fuse:/path/to/gdrive/folder
Auto-extract Extract files after downloading. Add links/to/file ending with .zip or .tar.lz4. Extract files to specified destination directory. https//link.com/to/file
Install Extensions Install extensions for Stable Diffusion Web UI. Add the link to the Github repository to custom_extension_url. https://github.com/user/repo

Once the download is complete, you can proceed to the next step.

Start Stable Diffusion Web UI

This is a cell for launching Stable Diffusion Web UI after a long configuration process above. However, it still has its own configuration requirements, such as defining arguments, using alternative tunnels, selecting a theme, and so on.

Option Default Description
Alternative Tunnels Note: Recommended Tunnels: ngrok > cloudflared > remotemoe > localhostrun > googleusercontent > gradio
tunnel multiple Allow users to use alternative tunnels for shared links such as cloudflared, remotemoe, localhostrun, and googleusercontent.
ngrok_token ' ' Enable ngrok tunnel for shared links. Users can get their ngrok token from here. If ngrok_token is enabled, it will automatically disable alt_tunnels.
ngrok_region ap Specify the desired region for ngrok tunnel. Users can choose between ["us", "eu", "au", "ap", "sa", "jp", "in"].
Launch Arguments
theme_selector ogxBGreen This section used to change preferred theme for Anapnoe's Stable Diffusion Web UI/UX
use_gradio_auth False If enabled, every time a user opens shared links, it will ask for authentication. The username is set to cagliostro and the password will be randomly generated with 6 ASCII characters and numbers. Authentication information will be printed when the user launches the Web UI.
accelerator xformers To accelerate your generation and training, we offer 3 arguments. People commonly use xformers, but opt-sdp-attention would be a nice choice to have after Colab updates PyTorch to 2.0. It's claimed to be better than xformers but only for generating at lower resolution. So we choose to use xformers by default.
quiet_mode True This is a quality-of-life option. If enabled, it will use --no-hashing to skip model hashing, which is useful for not double-checking the model if it doesn't have the same hash and for merging the model. If enabled, it will also use --disable-console-progressbars to only have 1 line progress bars instead of updating new lines every generation.
auto_select_model False This argument allows searching for a model and randomly selecting the model in the model directory.
auto_select_VAE True This argument allows searching for a VAE and randomly selecting the VAE in the VAE directory.
no_half_VAE True This argument prevents black generation caused by certain VAEs. This usually happens because certain VAEs like anime have FP32.
additional_args --no-download-sd-model --gradio-queue This argument allows the user to add custom args from here.

Download Generated Images

This cell useful to download outputs generated from Stable Diffusion Web UI as a compressed .zip

Option Default Description
use_drive True If enabled, this option will ask to mount your Google Drive account and save the compressed file to a preferred folder name specified by folder_name. It will also provide a download link after the process is completed.
folder_name "AI-generated Art" This option is used to specify the preferred folder name in Google Drive if use_drive is enabled.
filename "waifu.zip" This option allows the user to specify the filename of the zipfile. If use_drive is enabled and a duplicate filename is found in Google Drive, it will be renamed to filename(n+1).zip.

Extra Cell Explained

Download Generated Images V2

Same as Download Generated Images, the difference is it send the zipfile to Huggingface instead of Google Drive. Personally, it's better than save it to Google Drive.

Option Default Value Description
write_token "" User needs to specify a WRITE token from Hugging Face settings.
repo_name "ai-art-dump" Specifies where the repository is located. If no repository is available, one will be created automatically.
private_repo False Specifies whether the repository is private or public to everyone.
project_name "waifu" Same as Download Generated Images, this option allows the user to specify the name of the zipfile.

Resources

Default Extensions

Extensions Github Repository
ashen-sensored/sd_webui_stealth_pnginfo Link
hnmr293/sd-webui-cutoff Link
KohakuBlueleaf/a1111-sd-webui-locon Link
DominikDoom/a1111-sd-webui-tagcomplete Link
etherealxx/batchlinks-webui Link
mcmonkeyprojects/sd-dynamic-thresholding Link
kohya-ss/sd-webui-additional-networks Link
thomasasfk/sd-webui-aspect-ratio-helper Link
Mikubill/sd-webui-controlnet Link
camenduru/sd-webui-tunnels Link
bbc-mc/sdweb-merge-block-weighted-gui Link
bbc-mc/sdweb-xyplus Link
opparco/stable-diffusion-webui-composable-lora Link
AlUlkesh/stable-diffusion-webui-images-browser Link
arenatemp/stable-diffusion-webui-model-toolkit Link
opparco/stable-diffusion-webui-two-shot Link
Coyote-A/ultimate-upscale-for-automatic1111 Link

Default Custom Upscalers

ESRGAN Upscaler Mirror Link
lollypop.pth Link
4x-AnimeSharp.pth Link
4x_foolhardy_Remacri.pth Link
4x-UltraSharp.pth Link

Default Negative Embeddings

Negative Embeddings Link
SDv1.x Negative Embeddings
EasyNegative.safetensors Link
bad-artist-anime.pt Link
bad-hands-5.pt Link
bad-artist.pt Link
bad_prompt_version2.pt Link
bad_prompt.pt Link
bad-image-v2-39000.pt Link
ng_deepnegative_v1_75t.pt Link
SDv2.x Negative Embeddings
rev2-badprompt.safetensors Link
re-badprompt.safetensors Link
wdbadprompt.pt Link