What are the differences between DelineateAnything v2 and v1?
Hi, I noticed the new DelineateAnythingv2.pt model. Could you briefly explain its main improvements over v1, such as architecture, training data, performance, and usage? Is v2 a drop-in replacement for v1?
Thank you!
Hi @zt1y14 , thank you for the question! Here is a paper that was just accepted to ECCV 2026 for a detailed explanation: https://arxiv.org/pdf/2607.19069. In short, this version is much better for global usage (outside Europe, where v1 was trained).
Congratulations on the ECCV 2026 acceptance, and thank you for sharing the paper and clarification! This is very helpful and clears up several questions I had about v2.
I am currently encountering an issue with v1 where some large agricultural areas are not properly separated into individual fields and are instead merged into a single large polygon. Since my study area is outside Europe, the improved global generalization of v2 looks especially promising and may help address this issue. I’m looking forward to testing it on my data. Thank you again!
Amazing !!!
@zt1y14 , thank you! I'm thrilled to hear that v2 sounds like a great fit for your use case, resolving that exact under-segmentation issue outside Europe was our main motivation, which is why we were so eager to share it!
If it proves useful in your work, we'd really appreciate it if you could share the project with your colleagues and leave a ⭐️ on GitHub. Good luck with testing, and feel free to let us know how it goes!
@MykolaL I'm interested to know the timeline for releasing the v2 dataset, and whether it will include imagery at all resolutions? I'm also curious if you compared to a simple baseline, such as Unet used for Fields of the World?
Thanks
@robmarkcole thanks for interest. Yes we plan to release v2 whole dataset. Probably in 1-2 months, now is vacation period. Yes we tried the most advanced Unet from CVPR paper (please check DelAny paper v1). Also comparison with FTW, you could check last paper and DelAnyFlow for more details.