cmmd-pytorch

(Unofficial) PyTorch implementation of CLIP Maximum Mean Discrepancy (CMMD) for evaluating image generation models, proposed in Rethinking FID: Towards a Better Evaluation Metric for Image Generation. CMMD stands out to be a better metric than FID and tries to mitigate the longstanding issues of FID.

This implementation is a super simple PyTorch port of the original codebase. I have only focused on the JAX and TensorFlow specific bits and replaced them PyTorch. Some differences:

• The original codebase relies on scenic for computing CLIP embeddings. This repository uses transformers.
• For the data loading, the original codebase uses TensorFlow, this one uses PyTorch Dataset and DataLoader.

Setup

First, install PyTorch following instructions from the official website.

Then install the depdencies:

pip install -r requirements.txt

Running

python main.py /path/to/reference/images /path/to/eval/images --batch_size=32 --max_count=30000

A working example command:

python main.py reference_images generated_images --batch_size=1

It should output:

The CMMD value is:  7.696

This is the same as the original codebase, so, that confirms the implementation correctness 🤗

GPU execution is supported when a GPU is available.

Results

Below, we report the CMMD metric for some popular pipelines on the COCO-30k dataset, as commonly used by the community. CMMD, like FID, is better when it's lower.

Pipeline	Inference Steps	Resolution	CMMD
stabilityai/stable-diffusion-xl-base-1.0	30	1024x1024	0.696
segmind/SSD-1B	30	1024x1024	0.669
stabilityai/sdxl-turbo	1	512x512	0.548
runwayml/stable-diffusion-v1-5	50	512x512	0.582
PixArt-alpha/PixArt-XL-2-1024-MS	20	1024x1024	1.140
SPRIGHT-T2I/spright-t2i-sd2	50	768x768	0.512

Notes:

• For SDXL Turbo, guidance_scale is set to 0 following the official guide in diffusers.
• For all other pipelines, default guidace_scale was used. Refer to the official pipeline documentation pages here for more details.

As per the CMMD authors, with models producing high-quality/high-resolution images, COCO images don't seem to be a good reference set (they are of pretty small resolution). This might help explain why SD v1.5 has a better CMMD than SDXL.

Obtaining CMMD for your pipelines

One can refer to the generate_images.py script that generates images from the COCO-30k randomly sampled captions using diffusers.

Once the images are generated, run:

python main.py /path/to/reference/images /path/to/generated/images --batch_size=32 --max_count=30000

Reference images are COCO-30k images and can be downloaded from here.

Pre-computed embeddings for the COCO-30k images can be found here.

To use the pre-computed reference embeddings, run:

python main.py None /path/to/generated/images ref_embed_file=ref_embs.npy --batch_size=32 --max_count=30000

Acknowledgements

Thanks to Sadeep Jayasumana (first author of CMMD) for all the helpful discussions.