Support for Bagel

[Jialuo-Li]

Hi! Thanks for the great work on this framework. I was wondering if you have plans to support Bagel (specifically the T2I stage) in the future? It would be a fantastic addition to the current suite. Thanks!

[Jayce-Ping]

Hi! Thank you so much for the great suggestion!

Supporting Bagel (specifically the T2I stage) is definitely a valuable direction. I have added this to our TODO list.

In the meantime, if you're looking to experiment with it right away, I recommend checking out the implementation in flow-grpo as a reference.

Once we integrate Bagel into Flow-Factory, it will unlock further possibilities for applying algorithms like NFT and AWM. I'll keep you posted on the progress!

[Jialuo-Li]

Thank you so much!

[Jayce-Ping]

@Jialuo-Li Hi! I’ve added support for Bagel in this branch.

Please note that since Bagel requires an older version of transformers (v0.49.0), I won't be merging this into the main branch unless Bagel is updated for compatibility with newer versions.

I have successfully tested the GRPO algorithm on Bagel for both text-to-image and image(s)-to-image tasks. Two example configurations are provided, though they haven't been fully fine-tuned yet. In my initial tests, I observed performance growth, but the model encountered a collapse after 300 steps (as shown in the training logs), which suggests a need for further hyperparameter tuning.

As for NFT and AWM, I am currently running experiments on NFT to see how much performance gain can be achieved.

[Jialuo-Li]

Thanks for the update and for setting up the dedicated branch! It’s a huge help.

[Jayce-Ping]

I have tested DiffusionNFT on Bagel and got the following performance increase:

Training Reward

Evaluation Reward

The config file is available at here.