home | about | git

vkdt

an experimental image processing graph

the vkdt project addresses several shortcomings of the core architecture of darktable. these points are the main motivation for a new core architecture:

developer sanity

• single code-path GPU processing only (not i386, SSE/AVX, OpenCL mixed)
• single pipeline with full res full region only (no need to sync gui, preview, cropped)
• bloat free
• no gtk (also no qt)
• introspection code not needed: configs are human readable
• most data stays on the GPU from initial processing to display
• strict separation of gui (c++) and core code (c, glsl)
• module reordering/multi instancing transparently solved with a node graph

core processing features

• vkdt does roughly 20x more work and is roughly 20x faster at the same time (don't ask me how this happens, classic dt just loses speed between the cracks)
• more speed means we can enable more advanced modules by default, pain free (see below)
• processing is a node graph: vkdt supports multiple inputs (merge/align frames) and multiple outputs (variants, histograms, ..) at the same time

user facing features

• full res raw processing usually faster than vsync
• supports 10-bit displays out of the box (try this with gtk/cairo..)
• highlight inpainting
• interactive brush strokes with guided smoothing
• local laplacian pyramids for local contrast
• deconvolution sharpening
• thumbnail processing is disk i/o limited (be sure to use fast SSD where you can) and rawspeed limited: some raws can be decoded quicker (RAF) than others (CR2).
• animations with optional feedback loops
• full screen colour management that supports dual monitors even when the window is split between the two

screenshots

images shown here are taken from the playraw category on pixls.us.

lighttable mode:

darkroom mode, showing the configurable favourites tab:

darkroom mode, showing the node graph connections: