R5

Changes:

• Add TTA mode.

Notes:

• The model is same as r0.1. You need to download model files from here.

R4.1

Changes:

• Update ncnn to 20220420. Close #11
• Clamp input pixel values. Close #8

Notes:

• The model is same as r0.1. You need to download model files from here.

R4

New features:

• Automatically choose tile_size according to available VRAM
• New tile_size_w / tile_size_h options for overriding width and/or height of tile_size
• Lower VRAM usage in fp16 precision

Fix:

• Fixed crash. #6

Notes:

• The model is same as r0.1. You need to download model files from here.

R3.2

Changes:

• Build with the latest version of NCNN

Notes:

• The model is same as r0.1. You need to download model files from here.

R3.1

Changes:

• Build with the latest version of NCNN

Notes:

• The model is same as r0.1. You need to download model files from here.

R3

Changes:

• Improve the build process
• Build with the latest version of NCNN

Fixes:

• Fixed incorrect padding when the input clip width or height is odd.
• Fixed incorrect result for particular locales.

Notes:

• The model is same as r0.1. You need to download model files from here.

R2

Changes:

• Support 32-bit float input and deprecate 8-bit integer input.
• New precision option.

Notes:

• The model is same as r0.1. You need to download model files from here.

R1

Minor bug fix.

The model is same as r0.1. You need to download model files from here.

r0.1

First release