Feed forward VQGAN-CLIP model, where the goal is to eliminate the need for optimizing the latent space of VQGAN for each input prompt. This is done by training a model that takes as input a text prompt, and returns as an output the VQGAN latent space, which is then transformed into an RGB image. The model is trained on a dataset of text prompts and can be used on unseen text prompts. The loss function is minimizing the distance between the CLIP generated image features and the CLIP input text features. Additionally, a diversity loss can be used to make increase the diversity of the generated images given the same prompt.
-
09 July 2022
- Version 0.4 released. New models that can be tried in the notebook and on Replicate:
- Three models (one for 256x256 resolution, one for 512x512 resolution) trained with OpenCLIP Vit-B/32 (LAION-2B-en model)
- One model for ViT-B/32 trained with CLIP and pixel reconstruction
- Update Colab Notebook with easy to use widgets for image and video generation
- Support CLOOB models trained by @crowsonkb on LAION-400M
- Support Training/using priors to generate image embeddings from text embeddigs using Net2Net, this can be used to generate multiple images for the same text prompt. See here for an example. Also available in the Colab Notebook and Replicate.
- Support L2 loss and Total variation losses as regularizations
- Support OpenCLIP models
- Support evaluation in the training phase
- Version 0.4 released. New models that can be tried in the notebook and on Replicate:
-
04 January 2022
- Support CLOOB as an alternative OpenAI's CLIP models
- Support X-transformer as an alternative to MLP Mixer and VitGAN
- New models (can be tried in the notebook and on Replicate) ) released (see 0.3 version in https://github.com/mehdidc/feed_forward_vqgan_clip#pre-trained-models):
- New model which use text and real image embeddings as a loss instead of just text embeddings, trained on conceptual captions 12M
- New model based on CLOOB
- First 512x512 model based on X-transformer
- Update notebook with Real-ESRGAN to upscale the images
-
22 September 2021
- New models released (see 0.2 version in https://github.com/mehdidc/feed_forward_vqgan_clip#pre-trained-models)
- New Colab notebook for training from scratch or fine-tuning
- Web interface from Replicate AI to use the models
Links:
- https://github.com/mehdidc/feed_forward_vqgan_clip/releases/download/0.1/vqgan_imagenet_f16_16384.ckpt (vqgan_imagenet_f16_16384.ckpt)
- https://github.com/mehdidc/feed_forward_vqgan_clip/releases/download/0.1/vqgan_imagenet_f16_16384.yaml (vqgan_imagenet_f16_16384.yaml)
conda create -n ff_vqgan_clip_env python=3.8
conda activate ff_vqgan_clip_env
# Install pytorch/torchvision - See https://pytorch.org/get-started/locally/ for more info.
(ff_vqgan_clip_env) conda install pytorch torchvision torchaudio cudatoolkit=11.1 -c pytorch -c nvidia
(ff_vqgan_clip_env) pip install -r requirements.txtconda deactivate # Make sure to use your global python3
python3 -m pip install venv
python3 -m venv ./ff_vqgan_clip_venv
source ./ff_vqgan_clip_venv/bin/activate
$ (ff_vqgan_clip_venv) python -m pip install -r requirements.txt- If you want to use priors (see 09 July 2022 release), please install Net2Net, e.g. with
pip install git+https://github.com/CompVis/net2net
$ (ff_vqgan_clip_venv) python main.py tokenize data/list_of_captions.txt cembeds 128
Modify configs/example.yaml as needed.
$ (ff_vqgan_clip_venv) python main.py train configs/example.yaml
Loss will be output for tensorboard.
# in a new terminal/session
(ff_vqgan_clip_venv) pip install tensorboard
(ff_vqgan_clip_venv) tensorboard --logdir resultsAfter downloading a model (see Pre-trained models available below) or finishing training your own model, you can test it with new prompts, e.g.:
wget https://github.com/mehdidc/feed_forward_vqgan_clip/releases/download/0.2/cc12m_32x1024_vitgan.thpython -u main.py test cc12m_32x1024_vitgan.th "Picture of a futuristic snowy city during the night, the tree is lit with a lantern"
You can also use the priors to generate multiple images for the same text prompt, e.g.:
wget https://github.com/mehdidc/feed_forward_vqgan_clip/releases/download/0.4/cc12m_32x1024_mlp_mixer_openclip_laion2b_ViTB32_256x256_v0.4.thwget https://github.com/mehdidc/feed_forward_vqgan_clip/releases/download/0.4/prior_cc12m_2x1024_openclip_laion2b_ViTB32_v0.4.thpython main.py test cc12m_32x1024_mlp_mixer_openclip_laion2b_ViTB32_256x256_v0.4.th "bedroom from 1700" --prior-path=prior_cc12m_2x1024_openclip_laion2b_ViTB32_v0.4.th --nb-repeats=4 --images-per-row=4
You can also try all the models in the Colab Notebook and in Replicate. Using the notebook, you can generate images from pre-trained models and do interpolations between text prompts to create videos, see for instance video 1 or video 2 or video 3.
| Name | Type | Size | Dataset | Link | Author |
|---|---|---|---|---|---|
| cc12m_32x1024_mlp_mixer_clip_ViTB32_pixelrecons_256x256 | MLPMixer | 1.2GB | Conceptual captions 12M | Download | @mehdidc |
| cc12m_32x1024_mlp_mixer_openclip_laion2b_ViTB32_256x256 | MLPMixer | 1.2GB | Conceptual captions 12M | Download | @mehdidc |
| cc12m_32x1024_mlp_mixer_openclip_laion2b_imgEmb_ViTB32_256x256 | MLPMixer | 1.2GB | Conceptual captions 12M | Download | @mehdidc |
| cc12m_1x1024_mlp_mixer_openclip_laion2b_ViTB32_512x512 | MLPMixer | 580MB | Conceptual captions 12M | Download | @mehdidc |
| prior_cc12m_2x1024_openclip_laion2b_ViTB32 | Net2Net | 964MB | Conceptual captions 12M | Download | @mehdidc |
| prior_cc12m_2x1024_clip_ViTB32 | Net2Net | 964MB | Conceptual captions 12M | Download | @mehdidc |
| Name | Type | Size | Dataset | Link | Author |
|---|---|---|---|---|---|
| cc12m_32x1024_mlp_mixer_clip_ViTB32_256x256 | MLPMixer | 1.19GB | Conceptual captions 12M | Download | @mehdidc |
| cc12m_32x1024_mlp_mixer_cloob_rn50_256x256 | MLPMixer | 1.32GB | Conceptual captions 12M | Download | @mehdidc |
| cc12m_256x16_xtransformer_clip_ViTB32_512x512 | Transformer | 571MB | Conceptual captions 12M | Download | @mehdidc |
| Name | Type | Size | Dataset | Link | Author |
|---|---|---|---|---|---|
| cc12m_8x128 | MLPMixer | 12.1MB | Conceptual captions 12M | Download | @mehdidc |
| cc12m_32x1024 | MLPMixer | 1.19GB | Conceptual captions 12M | Download | @mehdidc |
| cc12m_32x1024 | VitGAN | 1.55GB | Conceptual captions 12M | Download | @mehdidc |
| Name | Type | Size | Dataset | Link | Author |
|---|---|---|---|---|---|
| cc12m_8x128 | VitGAN | 12.1MB | Conceptual captions 12M | Download | @mehdidc |
| cc12m_16x256 | VitGAN | 60.1MB | Conceptual captions 12M | Download | @mehdidc |
| cc12m_32x512 | VitGAN | 408.4MB | Conceptual captions 12M | Download | @mehdidc |
| cc12m_32x1024 | VitGAN | 1.55GB | Conceptual captions 12M | Download | @mehdidc |
| cc12m_64x1024 | VitGAN | 3.05GB | Conceptual captions 12M | Download | @mehdidc |
| bcaptmod_8x128 | VitGAN | 11.2MB | Modified blog captions | Download | @afiaka87 |
| bcapt_16x128 | MLPMixer | 168.8MB | Blog captions | Download | @mehdidc |
NB: cc12m_AxB means a model trained on conceptual captions 12M, with depth A and hidden state dimension B
- The training code is heavily based on the VQGAN-CLIP notebook https://colab.research.google.com/drive/1ZAus_gn2RhTZWzOWUpPERNC0Q8OhZRTZ, thanks to all the authors who contributed to the notebook (@crowsonkb, @advadnoun, @Eleiber, @Crimeacs, @Abulafia)
- Thanks to @lucidrains, the MLP mixer model (
mlp_mixer_pytorch.py) is from https://github.com/lucidrains/mlp-mixer-pytorch. - Thanks to CompVis for Taming Transformers https://github.com/CompVis/taming-transformers, the code uses VQGAN pre-trained model and VGG16 feature space perceptual loss https://github.com/CompVis/taming-transformers/blob/master/taming/modules/losses/lpips.py
- Thanks to @afiaka87 for all the contributions to the repository's code and for providing the blog captions dataset for experimentation
- Thanks to VitGAN authors, the VitGAN model is from https://github.com/wilile26811249/ViTGAN
- Thanks to Replicate team, especially @chenxwh and @andreasjansson for making and hosting a browser based text to image interface using the model and for all the support
- Thanks to the authors of CLOOB for the code and the pre-trained models
- Thanks to @crowsonkb, code/models for CLOOB pre-trained on LAION-400M are based on cloob-training
- Thanks to OpenCLIP authors for CLIP-like code/models pre-trained on LAION-400M and LAION-2B
- Thanks to CompVis's Net2Net (https://github.com/CompVis/net2net), it was used to train text to image embedding priors
- Models were trained on JURECA-DC supercomputer at Jülich Supercomputing Centre (JSC), many thanks for the compute provided to train the models.