November 30, 2023

• Following the launch of SDXL-Turbo, we are releasing SD-Turbo.

November 28, 2023

• We are releasing SDXL-Turbo, a lightning fast text-to image model. Alongside the model, we release a technical report

  • Usage:
    • Follow the installation instructions or update the existing environment with pip install streamlit-keyup.
    • Download the weights and place them in the checkpoints/ directory.
    • Run streamlit run scripts/demo/turbo.py.

  

November 21, 2023

• We are releasing Stable Video Diffusion, an image-to-video model, for research purposes:

  • SVD: This model was trained to generate 14 frames at resolution 576x1024 given a context frame of the same size. We use the standard image encoder from SD 2.1, but replace the decoder with a temporally-aware deflickering decoder.
  • SVD-XT: Same architecture as SVD but finetuned for 25 frame generation.
  • We provide a streamlit demo scripts/demo/video_sampling.py and a standalone python script scripts/sampling/simple_video_sample.py for inference of both models.
  • Alongside the model, we release a technical report.

  

July 26, 2023

• We are releasing two new open models with a permissive CreativeML Open RAIL++-M license (see Inference for file hashes):
  • SDXL-base-1.0: An improved version over SDXL-base-0.9.
  • SDXL-refiner-1.0: An improved version over SDXL-refiner-0.9.

July 4, 2023

• A technical report on SDXL is now available here.

June 22, 2023

• We are releasing two new diffusion models for research purposes:
  • SDXL-base-0.9: The base model was trained on a variety of aspect ratios on images with resolution 1024^2. The base model uses OpenCLIP-ViT/G and CLIP-ViT/L for text encoding whereas the refiner model only uses the OpenCLIP model.
  • SDXL-refiner-0.9: The refiner has been trained to denoise small noise levels of high quality data and as such is not expected to work as a text-to-image model; instead, it should only be used as an image-to-image model.

If you would like to access these models for your research, please apply using one of the following links: SDXL-0.9-Base model, and SDXL-0.9-Refiner. This means that you can apply for any of the two links - and if you are granted - you can access both. Please log in to your Hugging Face Account with your organization email to request access. We plan to do a full release soon (July).

Modularity is king. This repo implements a config-driven approach where we build and combine submodules by calling instantiate_from_config() on objects defined in yaml configs. See configs/ for many examples.

For training, we use PyTorch Lightning, but it should be easy to use other training wrappers around the base modules. The core diffusion model class (formerly LatentDiffusion, now DiffusionEngine) has been cleaned up:

• No more extensive subclassing! We now handle all types of conditioning inputs (vectors, sequences and spatial conditionings, and all combinations thereof) in a single class: GeneralConditioner, see sgm/modules/encoders/modules.py.
• We separate guiders (such as classifier-free guidance, see sgm/modules/diffusionmodules/guiders.py) from the samplers (sgm/modules/diffusionmodules/sampling.py), and the samplers are independent of the model.
• We adopt the "denoiser framework" for both training and inference (most notable change is probably now the option to train continuous time models):
  • Discrete times models (denoisers) are simply a special case of continuous time models (denoisers); see sgm/modules/diffusionmodules/denoiser.py.
  • The following features are now independent: weighting of the diffusion loss function (sgm/modules/diffusionmodules/denoiser_weighting.py), preconditioning of the network (sgm/modules/diffusionmodules/denoiser_scaling.py), and sampling of noise levels during training (sgm/modules/diffusionmodules/sigma_sampling.py).
• Autoencoding models have also been cleaned up.

git clone https://github.com/Stability-AI/generative-models.git
cd generative-models

This is assuming you have navigated to the generative-models root after cloning it.

NOTE: This is tested under python3.10. For other python versions, you might encounter version conflicts.

PyTorch 2.0

# install required packages from pypi
python3 -m venv .pt2
source .pt2/bin/activate
pip3 install -r requirements/pt2.txt

pip3 install .

pip3 install -e git+https://github.com/Stability-AI/datapipelines.git@main#egg=sdata

This repository uses PEP 517 compliant packaging using Hatch.

To build a distributable wheel, install hatch and run hatch build (specifying -t wheel will skip building a sdist, which is not necessary).

pip install hatch
hatch build -t wheel

You will find the built package in dist/. You can install the wheel with pip install dist/*.whl.

Note that the package does not currently specify dependencies; you will need to install the required packages, depending on your use case and PyTorch version, manually.

We provide a streamlit demo for text-to-image and image-to-image sampling in scripts/demo/sampling.py. We provide file hashes for the complete file as well as for only the saved tensors in the file ( see Model Spec for a script to evaluate that). The following models are currently supported:

• SDXL-base-1.0

  File Hash (sha256): 31e35c80fc4829d14f90153f4c74cd59c90b779f6afe05a74cd6120b893f7e5b
  Tensordata Hash (sha256): 0xd7a9105a900fd52748f20725fe52fe52b507fd36bee4fc107b1550a26e6ee1d7
  

• SDXL-refiner-1.0

  File Hash (sha256): 7440042bbdc8a24813002c09b6b69b64dc90fded4472613437b7f55f9b7d9c5f
  Tensordata Hash (sha256): 0x1a77d21bebc4b4de78c474a90cb74dc0d2217caf4061971dbfa75ad406b75d81
  

• SDXL-base-0.9
• SDXL-refiner-0.9
• SD-2.1-512
• SD-2.1-768

Weights for SDXL:

SDXL-1.0: The weights of SDXL-1.0 are available (subject to a CreativeML Open RAIL++-M license) here:

• base model: https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0/
• refiner model: https://huggingface.co/stabilityai/stable-diffusion-xl-refiner-1.0/

SDXL-0.9: The weights of SDXL-0.9 are available and subject to a research license. If you would like to access these models for your research, please apply using one of the following links: SDXL-base-0.9 model, and SDXL-refiner-0.9. This means that you can apply for any of the two links - and if you are granted - you can access both. Please log in to your Hugging Face Account with your organization email to request access.

After obtaining the weights, place them into checkpoints/. Next, start the demo using

streamlit run scripts/demo/sampling.py --server.port <your_port>

Images generated with our code use the invisible-watermark library to embed an invisible watermark into the model output. We also provide a script to easily detect that watermark. Please note that this watermark is not the same as in previous Stable Diffusion 1.x/2.x versions.

To run the script you need to either have a working installation as above or try an experimental import using only a minimal amount of packages:

python -m venv .detect
source .detect/bin/activate

pip install "numpy>=1.17" "PyWavelets>=1.1.1" "opencv-python>=4.1.0.25"
pip install --no-deps invisible-watermark

To run the script you need to have a working installation as above. The script is then useable in the following ways (don't forget to activate your virtual environment beforehand, e.g. source .pt1/bin/activate):

# test a single file
python scripts/demo/detect.py <your filename here>
# test multiple files at once
python scripts/demo/detect.py <filename 1> <filename 2> ... <filename n>
# test all files in a specific folder
python scripts/demo/detect.py <your folder name here>/*

We are providing example training configs in configs/example_training. To launch a training, run

python main.py --base configs/<config1.yaml> configs/<config2.yaml>

where configs are merged from left to right (later configs overwrite the same values). This can be used to combine model, training and data configs. However, all of them can also be defined in a single config. For example, to run a class-conditional pixel-based diffusion model training on MNIST, run

python main.py --base configs/example_training/toy/mnist_cond.yaml

NOTE 1: Using the non-toy-dataset configs configs/example_training/imagenet-f8_cond.yaml, configs/example_training/txt2img-clipl.yaml and configs/example_training/txt2img-clipl-legacy-ucg-training.yaml for training will require edits depending on the used dataset (which is expected to stored in tar-file in the webdataset-format). To find the parts which have to be adapted, search for comments containing USER: in the respective config.

NOTE 2: This repository supports both pytorch1.13 and pytorch2for training generative models. However for autoencoder training as e.g. in configs/example_training/autoencoder/kl-f4/imagenet-attnfree-logvar.yaml, only pytorch1.13 is supported.

NOTE 3: Training latent generative models (as e.g. in configs/example_training/imagenet-f8_cond.yaml) requires retrieving the checkpoint from Hugging Face and replacing the CKPT_PATH placeholder in this line. The same is to be done for the provided text-to-image configs.

The GeneralConditioner is configured through the conditioner_config. Its only attribute is emb_models, a list of different embedders (all inherited from AbstractEmbModel) that are used to condition the generative model. All embedders should define whether or not they are trainable (is_trainable, default False), a classifier-free guidance dropout rate is used (ucg_rate, default 0), and an input key (input_key), for example, txt for text-conditioning or cls for class-conditioning. When computing conditionings, the embedder will get batch[input_key] as input. We currently support two to four dimensional conditionings and conditionings of different embedders are concatenated appropriately. Note that the order of the embedders in the conditioner_config is important.

The neural network is set through the network_config. This used to be called unet_config, which is not general enough as we plan to experiment with transformer-based diffusion backbones.

The loss is configured through loss_config. For standard diffusion model training, you will have to set sigma_sampler_config.

As discussed above, the sampler is independent of the model. In the sampler_config, we set the type of numerical solver, number of steps, type of discretization, as well as, for example, guidance wrappers for classifier-free guidance.

For large scale training we recommend using the data pipelines from our data pipelines project. The project is contained in the requirement and automatically included when following the steps from the Installation section. Small map-style datasets should be defined here in the repository (e.g., MNIST, CIFAR-10, ...), and return a dict of data keys/values, e.g.,

example = {"jpg": x,  # this is a tensor -1...1 chw
           "txt": "a beautiful image"}

where we expect images in -1...1, channel-first format.