Repository navigation
Releases
Releases · Rabbit-Company/rabbit-denoise
v0.1.11
v0.1.10
Changes
analyze, and the first read ofdenoise --min-grain, are 1.6 to 2 times as fast, with the same results.
v0.1.9
Features
denoise --denoise-frames <RANGES>denoises only the frames given, such as0-299,515-1265, and copies the others unchanged. It is for programs that choose the scenes themselves, for example fromanalyze, so the video is read once and standard input works too. Every run of touching frames is denoised as if the video began and ended there, as with--min-grain.summarygivesdenoised_frames.
v0.1.8
Features
rabbit-denoise analyzesplits a video into scenes and measures the grain of each, without denoising. The grain is the deviation of the luma grain in 8 bit code values, measured on single frames and between pairs of frames, so that neither texture nor motion reads as grain. With--progress json, asceneevent gives each scene's frames, times, luma and chroma grain, and luma grain by brightness.denoise --min-grain <X>denoises only the scenes whose grain, asanalyzemeasures it, reaches X, and copies the others unchanged. For anime, 2 denoises grainy scenes such as a grainy opening and keeps clean ones as they are. It reads the video twice, so the input must be a file. With--progress json, ascenesevent reports what will be denoised, the first read'sprogressevents carrystageanalyze, andsummarygivesdenoised_frames.
Changes
- A closed standard output, as with
| head, no longer ends rabbit-denoise with a panic.denoisegoes on writing its video.analyzestops.
v0.1.7
Fixes
- Video compressed with H.264, such as a Blu-ray in AVC or a web release, came out with squares of fine checkerboards and stripes in flat areas, where the grain was left in parts. The codec quantizes the grain in transforms on the 8x8 grid, so on that grid most high frequencies of a block are zero and a few are large. rabbit-denoise measured the grain on the same grid, and the median it takes put those frequencies 2.5 to 4.5 times too weak in variance, so the filter kept them wherever they cleared the low threshold. The grain is now measured on blocks 3 samples off the grid. On a 1080p film scan in H.264, the grain left at the finest frequencies in flat areas fell from 36 to 1 per cent of the source's, and the 8x8 grid in the output from 26 per cent stronger edges than elsewhere to none. Such sources lose more grain and also more texture as faint as the grain, so use
--sigma-scale 0.8to keep more of it. On known grain compressed with x264 the result is 0.9 dB closer to the clean picture on average and up to 2.4 dB, with x265 0.1 dB. Lossless grain and the HEVC anime and film clips tested are unchanged within 0.05 dB.
v0.1.6
Changes
- About 9 per cent faster on the GPU.
filter_groupsadds up its patches in workgroup memory only within 2 pixels of its references instead of over the whole searched region, so the GPU runs more workgroups at once.
v0.1.5
Changes
- 1 to 2 per cent faster on the GPU. Each entry point runs for all three planes in one pass, and
find_groupsonly reads the motion fields, so the planes run at the same time instead of one after another. - The GPU filter needs 8 storage buffers per shader stage instead of 9.
v0.1.3
Changes
- The grain between frames and in what the first stage removed is measured on the GPU, in the same submission as the first stage. Only the medians are left to the CPU, so a slow CPU no longer holds the GPU back. At 1080p on a Radeon RX 9060 XT with an i7-8700K, the x86-64-v2 build goes from 10.6 to 19.1 fps on video without grain, and grainy video from about 18 to 20 fps. The GPU finds the same blocks as the CPU, to within 0.1 per cent.
- The first stage's result goes into the GPU's ring of frames on the GPU, so the second stage no longer uploads it.
- Without a GPU, what the first stage removed is measured on all threads.
v0.1.2
Changes
- About twice as fast on video with little or no grain, from 7.0 to 14.6 fps at 1080p on a Radeon RX 9060 XT with an i7-8700K. Such video measures no grain between frames, so its frames waited to be redone and the GPU and CPU took turns. The second stage is now queued behind the first stage then too, and the output is the same.
- The block search that measures grain between frames is about 40 per cent faster on the CPU.
v0.1.1
Changes
- About twice as fast on GPUs with much more compute than float atomic throughput, from 8.8 to 17.5 fps at 1080p on a Radeon RX 9060 XT.
filter_groupsadds up the patches of each 4x4 block of references in workgroup memory, so each pixel goes to the accumulators once. Adding every patch with float atomics took 75 per cent of the time there. - The grain between frames and in the residuals is measured on all planes at once, and chroma scales the luma basic estimates once for each size.