Final note: I'm (we're) tempted to organize a challenge on that topic as a workshop at a CV conf. ImageNet is the only source of images allowed and then you compete to get the bold numbers.

Do you think there would be people in for that? Do you think it would make for a nice competition?

With @arrijitghosh.bsky.social @nicolasdufour.bsky.social @davidpicard.bsky.social and @vickykalogeiton.bsky.social

🛠️ Try it yourself:

- Access the models on Hugging Face: huggingface.co/Lucasdegeorge/CAD-I
- Train your own text-to-image models using our setup: github.com/lucasdegeorge/T2I-ImageNet
- Check out the project page: lucasdegeorge.github.io/projects/t2i...

🔍 Key findings:

- Achieved +2 overall score over SD-XL on GenEval

- Achieved +5 performance on DPGBench 🏆

- Used only 1/10th of the model parameters

- Trained on 1/1000th of the typical training images

We used ImageNet with smart data-augmentations:

- Detailed Recaptioning: Transforming limited captions into rich, context-aware captions that capture styles, backgrounds, and actions

- Composition: Using CutMix to create diverse concept combinations, expanding the dataset's learning potential.

Text-to-image models are trained with the "bigger is better" paradigm.

But do we really need billions of images?

No, if we are careful enough!

We trained text-to-image models using 1000 times less data in just 200 GPU hours, achieving good-quality images and strong performance on benchmarks