Browser Based Imagen

This projects contains the scripts required for running imagen in the browser (i.e. using the actual chrome runtime). The GIF is sped up around 1000x, so in reality it isn't practical in any way, but hopefully this repo will be useful to someone else trying to put text2img models in the browser.

How it Works

Imagen works by taking a text prompt, running it through a t5 tokenizer, and then a t5 encoder (a.k.a "transformer"), before passing the encoded text plus some random noise to a unet, and performing a long sampling loop.

This is the process I followed:

1. Trained a very small imagen model on cifar-10 classes using google colab (see links)
2. Converted the unet portion into onnx format using torch.onnx (see export_unet.ipynb)
3. Downloaded an existing t5 transformer/tokenizer combo and exported these to ONNX/json (see export_t5.ipynb) respectively
4. Loaded all 3 (tokenizer, transformer, unet) into a chrome extension and used them to create a 250-step javascript inference loop where the output at each step is projected onto a html canvas

Links

The actual extension (as well as the pretrained pytorch Imagen model) can be found on huggingface. Simply download, unzip locally, and then load the build directory into chrome using "load unpacked". You can also find a pretrained imagen model here (pretrained for the cifar-10 classes).

The script for training a tiny imagen model can be found in this colab.

The wanb report from some successful training runs can be found here.

Replication

There is very little actual code in this repo. It's mostly just piping together different existing APIs built by the smarty pants people.

1. Train the model, just run all cells of imagen_koiboi.ipynb, if you run outside of colab you will need to remove references to google.colab
   • In this case I am training using the cifar-10 training dataset, which consists of 50,000 images
   • Because the images are only 32x32 we only use a single Unet, and we reduce its parameters significantly and only perform 250 steps (rather than the usual 1000) per sample. We rely heavily on lucidrains excellent imagen-pytorch.
   • It took 67,000 steps at a batch size of 64 to get pretty average results (see the wanb report above)
   • We use a t5-base embedding of the sentence "a photo of a class-name" as the conditioning embedding for each class
   • You can download it from the hugging face link above
2. Convert the model to ONNX
   • Using export_unet.ipynb import the best model you trained and overwrite some of the internal functions so that all instances of the expm1 function are removed from the model's internal graph (and replaced with an equivalent faux_expm1 function). This is important since ONNX has no equivalent operation for exmp1. My math knowledge is very weak, so I used ChatGPT to write faux_expm1.
   • ONNX can only convert models which have a forward method, so we create a new torch.nn.Module class ImagenOnnx which wraps the Imagen instance and performs a single simplified Imagen sampling step using its forward method.
   • We then convert an ImagenOnnx instance to ONNX using the standard torch.onnx.export.
3. Download the t5 tokenizer and transformer
   • The Tokenizer can be loaded from a json file, so all we need to do is download that file
   • The transformer we download and then convert into ONNX format using export_t5.ipynb
4. Load the tokenizer, transformer and imagen unet into the chrome extension (toymodel/public/model.js)
   • I used chrome-extension-cli to generate the boilerplate for the extension and npm run watch to keep the toymodel/build directory (which is the actual extension you load into chrome) in sync with my changes.
   • We import toymodel/public/tokenizers.js from the excellent transformers-js repo into toymodel/public/model.html above model.js, which allows us to load the t5 tokenizer from json.
   • The transformer and unet are sitting in ONNX files which we import using onnxruntime-web, which in this case is sitting in toymodel/public/ort.js and imported into model.html. We end up with two ONNX sessions.*
5. Perform inference
   • Once the models are loaded we convert the prompt into an embedding using the tokenizer and the transformer session and then perform 250 sampling steps on this embedding with the unet session. Results are drawn onto a 32x32 canvas on the popup each step. On my computer it takes around 7 minutes for 250 steps to complete, and if you un-focus the popup you need to start again.

*importing ONNX models into chrome extensions.

1. I found that the only reliable method was to use https://cdn.jsdelivr.net/npm/onnxruntime-web/dist/ort.js. Since this script relies on several generic javascript built-in constructs like URL, and Worker which are missing in both "content scripts" and "service workers", the only way I was able to get the import working was through an "extension page", in this case the popup itself. I also got hit with CORS errors if I tried to request this file from the web so I had to download it and include it in toymodel/public.
2. onnxruntime-web compiles the model into web-assembly, so I found it necessary to add the toymodel/public/ort-wasm-simd.wasm to toymodel/public to prevent errors complaining about there being no "wasm" backend. I found this file in node_modules\onnxruntime-web\dist\ort-wasm-simd.wasm after npm install onnxruntime-web and simply copied it over.
3. Executing web assembly code counts as "unsafe-eval" under the standard Content Security Policy and chrome's new Manifest v3 prevents most extensions from even loading if they have a CSP which permits "unsafe-eval". To get around this you need to set chrome extension sandbox permissions so that unsafe-eval is allowed inside sandboxes, and then specify that the page you are using to run any web assembly code is a sandbox. See toymodel/public/manifest.json for more details. If you don't do this you will get hit with CSP violations.

Citations

@inproceedings{Saharia2022PhotorealisticTD,
    title   = {Photorealistic Text-to-Image Diffusion Models with Deep Language Understanding},
    author  = {Chitwan Saharia and William Chan and Saurabh Saxena and Lala Li and Jay Whang and Emily L. Denton and Seyed Kamyar Seyed Ghasemipour and Burcu Karagol Ayan and Seyedeh Sara Mahdavi and Raphael Gontijo Lopes and Tim Salimans and Jonathan Ho and David Fleet and Mohammad Norouzi},
    year    = {2022}
}

@TECHREPORT{Krizhevsky09learningmultiple,
    author = {Alex Krizhevsky},
    title = {Learning multiple layers of features from tiny images},
    institution = {},
    year = {2009}
}