Dear authors,
We recently introduced a 10M scale image editing dataset UnicEdit-10M together with a more comprehensive and systematic benchmark, UnicBench, designed to evaluate both Basic Edit and Complex Edit tasks (including spatial transformations, compositional edits, and reasoning-based edits). UnicBench adopts fine-grained evaluation metrics, aiming to more faithfully assess model behavior in realistic and challenging image editing scenarios.
On UnicBench, we conducted a systematic evaluation of FLUX.1-Kontext.
Our results show that FLUX.1-Kontext exhibits particularly strong performance on Non-edit Consistency, demonstrating excellent preservation of irrelevant regions and strong edit consistency. In addition, its overall score ranks among the top tier of open-source image editing models.
Beyond evaluation, UnicEdit-10M is currently undergoing final checks and staged release. We believe this dataset could be valuable for further training or analysis of models such as FLUX.1-Kontext, especially for improving consistency-aware and context-preserving image editing.
We welcome discussion and collaboration, and sincerely appreciate your contribution to the open-source image editing community.
Relevant links:
Dear authors,
We recently introduced a 10M scale image editing dataset UnicEdit-10M together with a more comprehensive and systematic benchmark, UnicBench, designed to evaluate both Basic Edit and Complex Edit tasks (including spatial transformations, compositional edits, and reasoning-based edits). UnicBench adopts fine-grained evaluation metrics, aiming to more faithfully assess model behavior in realistic and challenging image editing scenarios.
On UnicBench, we conducted a systematic evaluation of FLUX.1-Kontext.
Our results show that FLUX.1-Kontext exhibits particularly strong performance on Non-edit Consistency, demonstrating excellent preservation of irrelevant regions and strong edit consistency. In addition, its overall score ranks among the top tier of open-source image editing models.
Beyond evaluation, UnicEdit-10M is currently undergoing final checks and staged release. We believe this dataset could be valuable for further training or analysis of models such as FLUX.1-Kontext, especially for improving consistency-aware and context-preserving image editing.
We welcome discussion and collaboration, and sincerely appreciate your contribution to the open-source image editing community.
Relevant links: