Thank you for your excellent work and for open-sourcing LeapAlign. We truly appreciate your contribution to the community.
We are currently evaluating the released models using the official pretrained weights and inference code without any modifications, following the instructions here: https://github.com/RockeyCoss/LeapAlign_Code#2-run-inference
However, we observed that the generation quality of flux_hpdv2_hpsv2 is not as expected. Specifically, there is minor noise along with slightly unclear object edges in the detailed regions of the generated images. In contrast, flux_geneval_hpsv2 and flux_mjhq30k_hpsv3 produce much cleaner and more acceptable results under the same inference setup:
Questions:
- Is this level of noise in flux_hpdv2_hpsv2 expected behavior with the current official weights?
- If not, could you point out what might be causing this issue? Are there any specific post-processing operations that we might have missed for this particular variant?
Any guidance would be greatly appreciated. Thanks again for your time!
Thank you for your excellent work and for open-sourcing LeapAlign. We truly appreciate your contribution to the community.
We are currently evaluating the released models using the official pretrained weights and inference code without any modifications, following the instructions here: https://github.com/RockeyCoss/LeapAlign_Code#2-run-inference
However, we observed that the generation quality of flux_hpdv2_hpsv2 is not as expected. Specifically, there is minor noise along with slightly unclear object edges in the detailed regions of the generated images. In contrast, flux_geneval_hpsv2 and flux_mjhq30k_hpsv3 produce much cleaner and more acceptable results under the same inference setup:
Questions:
Any guidance would be greatly appreciated. Thanks again for your time!