Thanks for your paper and work. I'm conducting research related to your work, but encountered some issues.
While using your figcodifier-8B model, I couldn't replicate the nearly perfect image-to-code correspondence seen in your paper, despite my images being significantly simpler. Could this be due to incorrect sampling params settings?
The paper only specifies setting the temperature parameter to 0.7, but doesn't mention configurations for top_k, top_p etc. My current setup with temperature at 0.7 shows unsatisfactory results. I also noticed that the default params json in FigCodifier's documentation perform even worse.
Thanks for your paper and work. I'm conducting research related to your work, but encountered some issues.
While using your figcodifier-8B model, I couldn't replicate the nearly perfect image-to-code correspondence seen in your paper, despite my images being significantly simpler. Could this be due to incorrect sampling params settings?
The paper only specifies setting the temperature parameter to 0.7, but doesn't mention configurations for top_k, top_p etc. My current setup with temperature at 0.7 shows unsatisfactory results. I also noticed that the default params json in FigCodifier's documentation perform even worse.