sairights/

sit-xl-2-256x256-sde-cfg

This repo contains PyTorch model definitions, pre-trained weights and training/sampling code for our paper exploring interpolant models with scalable transformers (SiTs).

We present Scalable Interpolant Transformers (SiT), a family of generative models built on the backbone of Diffusion Transformers (DiT). The interpolant framework, which allows for connecting two distributions in a more flexible way than standard diffusion models, makes possible a modular study of various design choices impacting generative models built on dynamical transport: using discrete vs. continuous time learning, deciding the model to learn, choosing the interpolant connecting the distributions, and deploying a deterministic or stochastic sampler. By carefully introducing the above ingredients, SiT surpasses DiT uniformly across model sizes on the conditional ImageNet 256x256 benchmark using the exact same backbone, number of parameters, and GFLOPs. By exploring various diffusion coefficients, which can be tuned separately from learning, SiT achieves an FID-50K score of 2.06.

This repository contains:

First, download and set up the repo:

We provide an environment.yml file that can be used to create a Conda environment. If you only want to run pre-trained models locally on CPU, you can remove the cudatoolkit and pytorch-cuda requirements from the file.

Pre-trained SiT checkpoints. You can sample from our pre-trained SiT models with sample.py. Weights for our pre-trained SiT model will be automatically downloaded depending on the model you use. The script has various arguments to adjust sampler configurations (ODE & SDE), sampling steps, change the classifier-free guidance scale, etc. For example, to sample from our 256x256 SiT-XL model with default ODE setting, you can use:

For convenience, our pre-trained SiT models can be downloaded directly here as well:

Custom SiT checkpoints. If you've trained a new SiT model with train.py (see below), you can add the --ckpt argument to use your own checkpoint instead. For example, to sample from the EMA weights of a custom 256x256 SiT-L/4 model with ODE sampler, run:

There are some more options; refer to train_utils.py for details.

We provide a training script for SiT in train.py. To launch SiT-XL/2 (256x256) training with N GPUs on one node:

Logging. To enable wandb, firstly set WANDB_KEY, ENTITY, and PROJECT as environment variables:

Then in training command add the --wandb flag:

Interpolant settings. We also support different choices of interpolant and model predictions. For example, to launch SiT-XL/2 (256x256) with Linear interpolant and noise prediction:

Resume training. To resume training from custom checkpoint:

Caution. Resuming training will automatically restore both model, EMA, and optimizer states and training configs to be the same as in the checkpoint.

We include a sample_ddp.py script which samples a large number of images from a SiT model in parallel. This script generates a folder of samples as well as a .npz file which can be directly used with ADM's TensorFlow evaluation suite to compute FID, Inception Score and other metrics. For example, to sample 50K images from our pre-trained SiT-XL/2 model over N GPUs under default ODE sampler settings, run:

The following command reproduces the SiT-XL/2 256x256 SDE + classifier-free guidance setting used for the verified ImageNet 256x256 evaluation:

On 50,000 generated samples, ADM's TensorFlow evaluator produced:

Likelihood. Likelihood evaluation is supported. To calculate likelihood, you can add the --likelihood flag to ODE sampler:

Notice that only under ODE sampler likelihood can be calculated; see sample_ddp.py for more details and settings.

Training (and sampling) could likely be speed-up significantly by:

Basic features that would be nice to add:

Precision in likelihood calculation could likely be improved by:

Our models were originally trained in JAX on TPUs. The weights in this repo are ported directly from the JAX models. There may be minor differences in results stemming from sampling on different platforms (TPU vs. GPU). We observed that sampling on TPU performs marginally worse than GPU (2.15 FID versus 2.06 in the paper).

This project is under the MIT license. See LICENSE for details.