visual grounding problem

[go-ahead-maker]

Hi, thank you for contributing such powerful model.

I'm trying to use skywork-38B to do REC task. But I found that the coordinates mismatch the given ref too much.

Is is because the raw results from model need to do post-process like resize or alignment?

[fakerbaby]

Hi @go-ahead-maker,

To help us diagnose the issue, could you please describe the specific failure case you experienced? Understanding the details will be very helpful.

It's important to note that the Skywork-R1V series models are multi-modal reasoning-enhanced models. We don't recommend them for recommendation (REC) tasks because they prioritize complex thinking over rapid perception, which can lead to longer processing times. If you're working on grounding or other perception-focused tasks, you might find other instruction-following vision-language models more suitable.

Please feel free to reach out if anything else comes up.