Add strong CWB refinement for Kroma v0.2 base-to-turbo
#3
by heyheywoah - opened
Adds a second-stage CWB refinement produced from:
- kroma-v0.2-base-to-turbo.safetensors
- kroma-v0.2-base-to-turbo-cwb.safetensors
Settings: ModelUtils Baseline CWB preset, median consensus, similarity alignment with threshold 0.4, alpha 2.0, mismatch mode skip, CUDA processing, and source dtype preservation.
The output contains 430 tensors and preserves the source dtype layout: 256 BF16 tensors and 174 FP32 tensors.
silveroxides changed pull request status to merged