How to run SD Forge WebUI on Google Colab

Updated Categorized as Tutorial Tagged , , 21 Comments on How to run SD Forge WebUI on Google Colab

Stable Diffusion Forge WebUI has emerged as a popular way to run Stable Diffusion and Flux AI image models. It is optimized to run fast and pre-installed with many essential extensions. If you don’t have a powerful GPU card, you can run Forge on Google Colab.

What is SD Forge?

Once upon a time, AUTOMATIC1111 WebUI (A1111) was the go-to software for running Stable Diffusion locally. Armed with a large “Generate” button, A1111 is perfect for interactive AI image creation.

However, the development of A1111 has lagged behind. It doesn’t support the latest local models like Flux. Optimized to run fast and supporting Flux, SD Forge WebUI has gained popularity.

This is a detailed guide for Forge on Google Colab. You can access the notebook by getting the Quick Start Guide.

If you are a site member, log in and access the notebook below.

Become a member of this site to see this content

Already a member? Log in here.

This notebook shares models with the following notebooks in Google Drive.

Using SD Forge on Google Colab

Google Colab (Google Colaboratory) is an interactive computing service offered by Google. It is a Jupyter Notebook environment that allows you to execute code.

Due to the computing resources required (High RAM), you need a Google Pro and Pro+ to run Forge on Colab.

I recommend using the Colab Pro plan. It gives you 100 compute units per month on T4, which are about 50 hours on a standard GPU. (It’s a steal)

Alternatives

Think Diffusion provides fully managed Forge/ComfyUI/AUTOMATIC1111 online service. They cost a bit more than Colab but provide a better user experience by installing models and extensions. They offer 20% extra credit to our readers. (Affiliate link)

Running SD Forge on Colab

Step 0: Sign up

Sign up a Google Colab Pro or Pro+ plans. (I use Pro.)

Step 1: Open the Forge Colab notebook

Open the Forge Colab notebook in the Quick Start Guide. You should see the notebook with the second cell below.

Note: For quick start, you can skip the following steps and run the notebook with the default settings.

Set the username and password. You will need to enter them before using Forge.

Step 2: Select models

Review which models you want to use.

The more you select, the longer it takes to download. They will be downloaded to the Colab drive, not your Google Drive.

Step 3: Run the notebook

Click the Play button on the left of the cell to start.

Start-up should complete within a few minutes. How long it takes depends on how many models you include. When it is done, you should see the message below.

Step 4: Start Forge

Follow the gradio.live link to start Forge.

Enter the username and password you specified in the notebook.

You should see the Forge GUI after you log in.

Put in “a cat” in the prompt text box and press Generate to test using Stable Diffusion. You should see it generates an image of a cat.

Speeding up image generation

You can pick a faster runtime type to speed up the generation, which costs more per hour.

Click downward caret on the top right and then select Change runtime type.

This notebook only supports GPU. Below is the approximate performance for Flux.1 Dev.

  • T4 GPU: ~ 1.5 mins per image.
  • L4 GPU: ~ 30 seconds per image.

ngrok (Optional)

If you run into display issues with the GUI, you can try using ngrok instead of Gradio to establish the public connection. It is a more stable alternative to the default gradio connection.

You will need to set up a free account and get an authoken.

  1. Go to https://ngrok.com/
  2. Create an account
  3. Verify email
  4. Copy the authoken from https://dashboard.ngrok.com/get-started/your-authtoken and paste it into the NGROK field in the notebook.

The Stable Diffusion cell in the notebook should look like the one below after you put in your ngrok authtoken.

Click the play button on the left to start running. When it is done loading, you will see a link to  ngrok.io in the output under the cell. Click the ngrok.io link to start AUTOMATIC1111. The first link in the example output below is the ngrok.io link.

When you visit the ngrok link, it should show a message like below

 Click on Visit Site to Start AUOTMATIC1111 GUI. Occasionally, you will see a warning message that the site is unsafe to visit. It is likely because someone used the same ngrok link to put up something malicious. Since you are the one who created this link, you can ignore the safety warning and proceed.

When you are done

When you finish using the notebook, don’t forget to click “Disconnect and delete runtime” in the top right drop-down menu. Otherwise, you will continue to consume compute credits.

Computing resources and compute units

Computing units and usage rate.

To view computing resources and credits, click the downward caret next to the runtime type (E.g. T4, High RAM) on the top right. You will see the remaining compute units and usage rate.

Models available

For your convenience, the notebook has options to load some popular models. You will find a brief description of them in this section.

Flux models

Flux AI is a state-of-the-art AI model that produces stunning images. You can use

  • Flux.1 Dev: The full development Flux model.
  • Flux.1 Schnell: The fast version.

v1.5 models

Stable Diffusion 1.5

The Stable Diffusion 1.5 model is the officially released model which is trained with diverse styles.

Realistic Vision

Realistic Vision v2 is good for generating anything realistic, whether they are people, objects, or scenes.

F222

F222

F222 is good at generating photo-realistic images. It is good at generating females with correct anatomy.

Caution: F222 is prone to generating explicit images. Suppress explicit images with a prompt “dress” or a negative prompt “nude”.

Dreamshaper

Dreamshaper

Model Page

Dreamshaper is easy to use and good at generating a popular photorealistic illustration style. It is an easy way to “cheat” and get good images without a good prompt!

Open Journey Model

Open Journey Model.

Model Page

Open Journey is a model fine-tuned with images generated by Mid Journey v4. It has a different aesthetic and is a good general-purpose model.

Triggering keyword: mdjrny-v4 style

Anything v3

Anything v3 model.

Model Page

Anything V3 is a special-purpose model trained to produce high-quality anime-style images. You can use danbooru tags (like 1girl, white hair) in the text prompt.

It’s useful for casting celebrities to amine style, which can then be blended seamlessly with illustrative elements.

Inkpunk Diffusion

Inkpunk Diffusion model.

Inkpunk Diffusion is a Dreambooth-trained model with a very distinct illustration style.

Model Page

Use keyword: nvinkpunk

SDXL model

SDXL

This Coalb notebook supports SDXL 1.0 base and refiner models.

Select SDXL_1 to load the SDXL 1.0 model.

Important: Don’t use VAE from v1 models. Go to Settings > Stable Diffusion. Set SD VAE to AUTOMATIC or None.

Check out some SDXL prompts to get started.

ControlNet models

Forge comes with the ControlNet extension installed but you still need to download the ControlNet models.

Alternatively, you can put the ControlNet models in the Google Drive folder AI_PICS > models > ControlNet.

Installing models

There are two ways to install models that are not on the model selection list.

  1. Use the Checkpoint_models_from_URL and LoRA_models_from_URL fields.
  2. Put model files in your Google Drive.

Install models using URLs

You can only install checkpoint or LoRA models using this method.

Put in the download URL links in the field. The link you initiate the file download when you visit it in your browser.

  • Checkpoint_models_from_URL: Use this field for checkpoint models.
  • LoRA_models_from_URL: Use this field for LoRA models.

Some models on CivitAi needs an API key to download. Go to the account page on CivitAI to create a key and put it in Civitai_API_Key.

Below is example of getting the download link on CivitAI.

Put it in the Model_from_URL field.

Installing models in Google Drive

After running the notebook for the first time, you should see the folder AI_PICS > models created in your Google Drive. The folder structure inside this folder mirrors AUTOMATIC1111‘s and is designed to share models with:

Put your model files in the corresponding folder. For example,

  • Put checkpoint model files in AI_PICS > models > Stable-diffusion.
  • Put LoRA model files in AI_PICS > models > Lora.

You will need to restart the notebook to see the new models on Forge.

Extensions

The notebook loads the following extensions by default.

  • Infinite Image browser – For viewing and searching generated images. The secret key is SDA.
  • Aspect ratio helper – Extra buttons for setting the aspect ratio of images.

Installing extensions from URL

You can install any number of extensions by using this field. You will need the URL of the Github page of the extension.

For example, put in the following if you want to install the Civitai model extension.

https://github.com/civitai/sd_civitai_extension

You can also install multiple extensions. The URLs need to be separated with commas. For example, the following URLs install the Civitai and the multi-diffusion extensions.

https://github.com/civitai/sd_civitai_extension,https://github.com/pkuliyi2015/multidiffusion-upscaler-for-automatic1111

Extra arguments to webui

You can add extra arguments to the Web-UI by using the Extra_arguments field.

Other useful arguments are

Frequently asked questions

Do I need a paid account to use the notebook?

Yes, you need a paid Google Colab account to use this notebook. Google has blocked the free usage of Stable Diffusion.

Is there any alternative to Google Colab?

Think Diffusion provides fully-managed Forge/AUTOMATIC1111/ComfyUI WebUI web service. They offer 20% extra credit to our readers. (Affiliate link)

How do you resolve an out-of-memory error?

The T4 runtime type has 15 GB of VRAM. You can select a runtime type of higher memory, such as L4 or A100. (They cost more.)

Do I need to use ngrok?

You don’t need to use ngrok to use the Colab notebook. In my experience, ngrok provides a more stable connection between your browser and the GUI. If you experience issues like buttons not responding, you should try ngrok.

What is the password for the Infinite Image Browser?

SDA

Can I use the checkpoint and LoRA models I trained?

Yes, put the model file in the corresponding folder in Google Drive.

  • Checkpoint models: AI_PICS > models > Stable-diffusion.
  • LoRA models: AI_PICS > models > Lora.

Why do my SDXL images look garbled?

Check to make sure you are not using a VAE from v1 models. Check Settings > Stable Diffusion > SD VAE. Set it to None or Automatic.

Next Step

If you are new to Stable Diffusion, check out the Absolute beginner’s guide.

Avatar

By Andrew

Andrew is an experienced engineer with a specialization in Machine Learning and Artificial Intelligence. He is passionate about programming, art, photography, and education. He has a Ph.D. in engineering.

21 comments

  1. As for the extensions:
    They are loaded to the virtual drive, but I want to include an extension that uses wildcards for prompts.
    I tried to load the gdrive extension folder but somehow it failed.
    Any suggestions on this?

    And btw:
    thx a lot for that notebook – it’s an epiphany! Spent the whole weekend to get it up and going.
    The trick was to manually change the graphic card in Google Collab from T4 to L4, which is also possible in the pre-paid plan!

      1. It’s the dynamic prompt extension:
        https://github.com/adieyal/sd-dynamic-prompts

        The default libraries are loaded to the virtual drive and work from there.
        But my custom libraries need to be in the folder “wildcards” and atm the only way to make it work is to upload them to the temporary disk, which is quite time consuming.

        As for extensions/loras beeing loaded through the script in general, it would be cool to have a switch between “Load to temporary virtual drive” and “Load to GDrive (this will take disk space of your GDrive)” and a checkbox to “Load existing extensions from GDrive”. Just an idea.

  2. Hi Andrew, thanks for sharing the notebook. As far as I can see it works. But I am running into an issue: From all my Loras and models I see only a selection: 5 Loras and 4 models. But is the problem here?

    1. Do you mean you cannot see all the models in Google Drive? You can try pressing the refresh button. You need to stop and rerun the cell (but no need to restart the notebook) if you just uploaded the models to Google Drive.

  3. Thx for sharing the notebook and for explaining everything.
    Works great with the Flux versions loaded through colab.
    But I ran into an issue with some custom model from Civitai.

    Basically I should be able to solve it by adding the folder /models/text_encoder for those files:
    clip_l.safetensors
    t5xxl_fp8_e4m3fn.safetensors
    t5xxl_fp16.safetensors

    But the notebook doesn’t read or find that folder.
    Any idea on how to solve this?

    1. Found out that I could add it to the code.
      Add this line for text_encoder
      f'{models_gdrive_path}/text_encoder’: f'{forge_local_path}/models/text_encoder’,
      }

  4. Hey Andrew, love that you put this together! Question.. I already have all of my controlnet, checkpoint and lora models uploaded into my Google Drive. Google Drive takes such an incredible amount of time to upload anything, so can you please let me know how would I point go about pointing Forge to those folders instead of AI_PICS > models??

    1. Moving the models to the AI_PICS/models folder in Google Drive should be quick.

      If you cannot move the models, you can create symbolic links in Colab’s terminal. (This is a more technical solution. I unfortunately cannot provide the instructions)

  5. The presets in Forge for the type of model – SD, XL, Forge – produce default image sizes that aren’t 1:1 nor 512 or 1024. Is the default optimal in terms of processing the image or should we revert to the standard ratios/sizes?
    Also, using Flux for the first time or the preset defaults to Euler as the sampler rather than DPM++2M – is one better than the other?

    1. The choice of sampler is not very important for flux. I leave it as euler for simplicity.

      The optimal image size for the models are 512×512, or 1024×1024 because this is how they are trained. But the square aspect ratios are normally not what we want. The different aspect ratio is a compromise that it still works reasonably well.

  6. Hello Andrew, maybe you can help me. I am using your Forge Colab Notebook with Colab Pro. If I want to use inpaint in Img2img and load a Lora into the prompts, nothing happens. The result changes only marginally and in no way corresponds to the Lora. What am I doing wrong or what am I overlooking?

      1. You had to remove “Automatic” from the low bits setting and select “Lora”. Otherwise Loras are apparently not recognized.

  7. When I run any Flux model, the notebook always automatically aborts the program without saving the image, I tested the notebook on SDXL model, it works fine, the issue only appears with Flux models generation.

    0% 0/20 [00:00<?, ?it/s]
    5% 1/20 [00:02<00:56, 2.96s/it]
    10% 2/20 [00:07<01:07, 3.77s/it]
    15% 3/20 [00:11<01:09, 4.07s/it]
    20% 4/20 [00:16<01:07, 4.21s/it]
    25% 5/20 [00:20<01:04, 4.31s/it]
    30% 6/20 [00:25<01:01, 4.38s/it]
    35% 7/20 [00:29<00:57, 4.43s/it]
    40% 8/20 [00:34<00:53, 4.46s/it]
    45% 9/20 [00:38<00:49, 4.50s/it]
    50% 10/20 [00:43<00:45, 4.54s/it]
    55% 11/20 [00:48<00:41, 4.57s/it]
    60% 12/20 [00:52<00:36, 4.61s/it]
    65% 13/20 [00:57<00:32, 4.64s/it]
    70% 14/20 [01:02<00:28, 4.67s/it]
    75% 15/20 [01:06<00:23, 4.70s/it]
    80% 16/20 [01:11<00:18, 4.75s/it]
    85% 17/20 [01:16<00:14, 4.77s/it]
    90% 18/20 [01:21<00:09, 4.80s/it]
    95% 19/20 [01:26<00:04, 4.80s/it]
    100% 20/20 [01:31<00:00, 4.55s/it]
    ^C

Leave a comment

Your email address will not be published. Required fields are marked *