- Replaced the flawed resolution calculation with a 2D search algorithm.
- The new logic correctly treats the user-provided 'length' as a fixed constraint.
- It now finds a compatible (width, height) pair that satisfies radial attention divisibility requirements for the given length.
- This resolves all previous and tensor mismatch errors.
- Replaced the previous resolution calculation with a more robust heuristic search to find compatible dimensions for radial attention.
- The new logic correctly ensures the total number of video tokens is a multiple of the block size, resolving the persistent tensor mismatch errors.
- Improved handling for both square and non-square resolutions to find the closest valid dimensions.
- Added a constraint to the function to ensure the total number of video tokens is a power of two.
- This resolves a caused by tensor shape mismatches in the node's mask generation logic.
- Overhauled the resolution calculation to be compatible with video radial attention by incorporating video length and patch divisor.
- Added UI controls for , , and to the ResolutionSelector node.
- The node now outputs for easier integration with video pipelines.
- Updated presets and added an example workflow for the new functionality.
Resolved an issue where Gemini API would not return images in 'generate' mode when an input image was present. The logic now only passes images in 'edit' or 'variation' modes. Fixed a bug where the ResolutionSelector's 'size' output was being overridden by a default 'aspect_ratio' in the Gemini config node. Updated presets and example workflows to align with these changes.
ImpactFrames