hdkai / plasma Goto Github PK

Conversion

• rgb_to_yuv
• yuv_to_rgb
• rgb_to_luminance

Filters

• contrast
• exposure
• saturation
• temperature
• tint
• selective_color
• clarity
• highlights
• shadows
• sharpen

Sampling

• bilateral_filter_2d
• splat_bilateral_grid
• slice_bilateral_grid
• gaussian_blur_2d
• gaussian_blur_3d
• laplacian_of_gaussian_2d
• color_sample_1d
• color_sample_3d

See how Lightroom does it.

• Try training with transfer learning

Add a last channel for weights then perform a homogenous divide.

And use torch.hub to retrieve it.

• Guided Image Filtering by He et al.
• Slides
• Reference implementation

Like Lightroom's "Texture" slider. Makes a significant difference.

Gradient

• radial_gradient
• top_bottom_gradient
• bottom_top_gradient
• left_right_gradient
• right_left_gradient

Linear

• contrast
• saturation
• temperature
• tint
• selective_color
• clarity
• highlights
• shadows
• sharpen
• texture
• exposure
• tone_curve

A very weird issue:

• Randomly positioned boxes with color tints
• For a given image, the position of the box is constant
• Does not happen on my MacBook Pro, only on AWS
• Isolated to clarity filter

This just might fix the enhancement problem entirely.

Caused by sqrt(0.) in _blend_overlay.

Using LUT's run into precision issues at extrema, unsurprisingly.

• Convert linear filters to broadcastable spatial filters.
• Merge with spatial.

See 1 Captain.

Shadows

• Replace bilateral filter with extremely strong Gaussian filter (use down-upsample trick)
• Replace overlay blending with soft light blending

Highlights

• Remove bilateral filter
• Contrast scale inverted mask

See test/media/filter/7.jpg.

Use naive exposure formulation.

As plasma.io sub-package.

Use gentle bilateral filter on chroma channels.

It would be interesting to explore the performance of this approach on extreme exposure pairs.