Releases: Jamy-L/Handheld-Multi-Frame-Super-Resolution
Release list
1.3.2: Configurable flow upscaling + bug fixs
1.3.1 : Optimised alignment + Minor fix
This improves the alignment speed with patches of size 64, by using a more efficient data layout.
1.3.0
What's Changed
New yaml configuration
The program now features a default yaml config that contains all the default hyperparameters. Users can provide their own yaml, or override some yaml arguments using command line arguments. This comes with a code cleanup related to arguments
Improved alignment module
The alignment step has been refactored and improved to be closer to the original implementation. The block matching step on the L2 norm now uses the FFT trick explained in the HDR+ article and is now much faster. The block matching step on the L1 has also be reimplemented to better use the GPU capacities. The flow upsampling step now uses ica iterations between each scale, which is faster and more precise than the old code. The ica refinement has also be reworked to be faster by using proper reduction tricks.
New dng exporter
The dng export feature is more robust to input tags, and will now use the full 16 bits range. The install instructions are also more detailed, and no longer windows-locked.
v1.2.2 : Noise Profile argument + Cleaner install
- The noise profile can noise be specified at execution time using the
alphaandbetaarguments, to skip the metadata check on the raw image. - A new argument allows the accumulated robustness mask to be saved next to output.
- The requirements have been cleaned and facilitated, by providing a conda environment featuring proper torch-cuda an numba versions.
V1.2.1 : Enabled postprocessing by default
This release updates the default prostprocessing settings of run_handheld.py. Gamma correction, tonemaping and color correction will now be applied by default (as was the case in example.py)