Several Improvements

[dnl13]

Proposed Changes:

• Added: ComfyUI Device Manager to handle device management for Comfy.
• Moved: Model location to /models/LLM.
• Introduced: max_new_tokens, num_beams, do_sample, and fill_mask parameters to Florence2 node.
• Updated: Logic to accurately draw masks from multiple polygons within the same image in the Florence2PostprocessAll Node.
• Modified: apply method to handle polygons in the Florence2PostprocessAll Node when using REFERRING_EXPRESSION_SEGMENTATION.

Example:

These changes were made to improve performance and functionality based on recent discoveries:

• Performance Optimization: Experimentation revealed that adjusting max_tokens not only improves processing speed at lower values but also allows for better mask quality when appropriately adjusted.
• Enhanced Masking: Previously, only one eye was masked when searching for "both eyes". Now, in Florence2PostprocessAll, both eyes are accurately captured within a single mask.
• Selective Masking: In Florence2Postprocess, users can now select either eye (left/right) using the index parameter.

Please review the changes and feel free to test them to ensure they meet your expectations. I believe these modifications align well with the intended improvements, and I welcome any feedback or further adjustments if needed.

A screenshot of my changes:

Thank you!

[skeyic]

I like your version, it works very well, and I hope it can be merged. The only regret is that it made me download the model again. lol