Questions on applying to other T2I models like Qwen-Image

[JHLEE17]

I'm very impressed with the excellent performance of the DC-Gen project.
I have a few questions about applying this methodology to other Text-to-Image models.

1. Are there any official plans to support other T2I (or I2I) models besides FLUX, such as Qwen-Image?

2. (DC-AE Retraining) Qwen-Image uses a different VAE than FLUX. In this case, is it correct that we would need to train a new DC-AE from scratch to fit Qwen-Image's VAE?

3. (Application Guide) I would like to try applying this methodology to Qwen-Image myself. Could you provide a brief, high-level guide on which scripts to run and in what order, based on the process used for the FLUX model?

Thank you in advance for your response!

[julien-blanchon]

I don't think you need to train DC-AE at all for adaptation to another DiT such as Qwen Image.

Conceptually everythings should be on this image

For stage 1, it should be almost the same things, I think Qwen Image also have a 8x scale factor. So the same code should work.
Same for stage 2.

[JHLEE17]

Thanks for the comment!
• For the FLUX experiments, could you outline how you constructed the synthetic dataset (prompt sources, categories and resolution buckets)? If possible, would you be able to share the prompt set you used?
• On reproducibility, adding cmd-level examples with the key flags for Stage-1 and Stage-2 on FLUX would make it even easier to replicate your results

[julien-blanchon]

I'm not the author haha, from the paper only I don't have answer on these two, I'm also pretty much interested to know.
I hope the authors will come here and answer

[han-cai]

Thank you for your interest in our work!

1. Supporting I2I models is on our roadmap—our next focus will be Qwen-Image-Edit.

2. There’s no need to train a new DC-AE.

3. We are currently undergoing legal review, so we cannot share anything until this process is completed.

[JHLEE17]

@han-cai
Thank you for the great work and for open-sourcing DC-Gen!

• Could you share a rough timeline for the expected release of Qwen-Image-Edit (code/models) and the legal review?
• Do you plan to support dc-ae-f16 (x16)? From my perspective, x16 looks like a good speed–quality trade-off. If you chose not to experiment with x16, could you share the reason (e.g., resource constraints, quality issues, priorities)? Because my goal is to accelerate Qwen-Image without quality drop, and I’d like to train on dc-ae-f16.
• I’m happy to contribute runs with my own GPUs toward the Qwen-Image + dc-ae-f16 direction. If that would be helpful, please let me know the expected setup (data format, scripts, evaluation set, checkpoints) and any contribution guidelines.