From c2cfacd7d3bad560ccbc0509e33d3862439bedf1 Mon Sep 17 00:00:00 2001 From: Duncan Tourolle Date: Wed, 7 Oct 2026 07:14:52 -0400 Subject: [PATCH] Record the new Best in the spec and the manual MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit denoise.md §15: why Best became one network, how it compares with the mixture and Medium on real photographs and the chart, the candidates that fell short, the file names, the saved-edit numbering, and the timings on the 3050 -- 0.51-0.54 s whole-frame, 0.95 s in tiles, against 2.60 s for the mixture in tiles. The manual lists Bilinear, Fast and Best, says an edit made with Medium opens with Best, and gives the new time; Medium's close-up goes. --- docs/dev/denoise.md | 56 ++++++++++++++++++++ docs/manual/README.md | 28 +++++----- docs/manual/index.html | 27 +++++----- docs/manual/media/develop-denoise-medium.png | 3 -- 4 files changed, 83 insertions(+), 31 deletions(-) delete mode 100644 docs/manual/media/develop-denoise-medium.png diff --git a/docs/dev/denoise.md b/docs/dev/denoise.md index 4cb0985..88992e4 100644 --- a/docs/dev/denoise.md +++ b/docs/dev/denoise.md @@ -571,3 +571,59 @@ In PyTorch fp16 the same network over the whole 20 MP frame in one call took 3.5 tiles, so a card that holds a whole frame gains more than the 6 GB one does; `WHOLE_FRAME_MAX` is a constant sized for 6 GB until the limit follows the card's memory. +## 15. Best becomes one network (0.24) + +The photographer's goal for 0.24 was Best's quality in under a second on the laptop. Whole frames +(§14) took the mixture from 2.60 s to 1.37 s and no further on a 6 GB card, so the other half was +a single network that holds the mixture's quality at a third of its work. Methods are now +`Bilinear`, `Fast` and `Best`; Medium and the mixture are retired. + +**The network** is `fb-combo` (darkroom-denoise, 2026-10-07): §11's shape (32-64-128-192, blocks +1-1-2-2, 3.2 M parameters, 48 GMAC/MP, halo 192), 20 000 steps from `fb-edges2` ← `student-m`, +taught by the mixture at a half share, with 10 % drawn scenes and 25 % crops from the edge-rich +cells of the training frames (branch `edge-sampling`). Scored on real photographs — the chart +overstated the mixture's lead (a chart-sharp network was softer than Medium on real edges) — on the +validation crops in the top quarter for sharp detail: + +| | Edge PSNR, ISO 1600 / 6400 / 25600 | Sharpness kept | Smooth areas | Held-out PSNR, ISO 400 / 1600 / 6400 / 25600 | Chart edge | +|---|---|---|---|---|---| +| Mixture (Best to 0.23) | 30.71 / 29.93 / 28.55 | 0.899 / 0.868 / 0.782 | 42.61 / 41.71 / 40.12 | 40.69 / 39.84 / 38.47 / 36.71 | 0.82 | +| `fb-combo` (Best from 0.24) | 30.67 / 29.87 / 28.49 | 0.902 / 0.874 / 0.792 | 42.54 / 41.57 / 39.85 | 40.64 / 39.78 / 38.38 / 36.54 | 0.89 | +| Medium (to 0.23) | 30.43 / 29.68 / 28.38 | 0.896 / 0.862 / 0.776 | 42.59 / 41.69 / 40.08 | 40.59 / 39.75 / 38.40 / 36.65 | 1.30 | + +Edges within 0.04–0.06 dB and more sharpness kept at every ISO; the known shortfall is smooth areas +at ISO 25600, 0.27 dB. The photographer took it as it stood at 20 000 of a planned 30 000 steps. +Others tried on the way, each short of the mixture on real photographs: `fb-sharp` (drawn scenes, +chart-sharp but Medium's real edges), `fb-edges` (half edge-rich crops: edges close, ISO 25600 +flats −0.24 dB), `fb-edges2` (a quarter: 0.03–0.14 dB short everywhere, chart 1.33–1.47), and a +from-scratch 24-48-96-128 between Fast and Medium. + +**Files.** `mosaic-hq-1408.onnx`, `mosaic-hq.onnx` (any size) and `mosaic-hq-1408.a16w16.onnx` for +the Hexagon. A new name, not Medium's or Best's: the result cache keys a model by name and size, +and this one is byte for byte Medium's size. The tablet form lost 0.00 dB in simulated QDQ at every +ISO and at most 0.09 dB across the ×0.5–×4 noise bracket (A16W8 0.08 / 0.26 dB; int8 −10.6 dB); +not yet confirmed on the tablet itself. + +**Saved edits** keep their numbers: 2, which was Medium, is now Best; 3, which was Best, is past the +end and reads as the default, Best. Both land on the new network with no migration. + +**Measured** 2026-10-07, `_MG_8862`, RTX 3050 Laptop, TensorRT fp16, P3 / 5001 MHz, nothing else on +the card: + +| Best | Network time | Peak GPU memory | +|---|---|---| +| mixture, 1408² tiles (0.23) | 2.60 s | — | +| mixture, whole frame (§14) | 1.37 s | — | +| `fb-combo`, 1408² tiles | 0.95 s | 0.55 GB | +| `fb-combo`, whole frame (two 4160 × 3248) | **0.51–0.54 s** | 1.75 GB | + +Decode and the hot-pixel pass add 0.4–0.5 s, so a photograph is about a second end to end. Whole +frame against tiles: max |Δ| 0.0029, 90 dB apart — fp16's spread. The whole-frame engine's first +build took 12 minutes (the mixture's 28); from the cache it loads in about a second, so the session +keeps the engine's ordinary 30 s idle decay rather than unloading after each photograph: at 1.75 GB +it fits beside the develop view on a 6 GB card, and an unload would cost the next photograph a +second. + +The manual's close-up for Best is still the mixture's render, which this network matches to within +the table above; it is re-recorded with the next pass of `tools/manual/record.sh`. + diff --git a/docs/manual/README.md b/docs/manual/README.md index f406b87..ab19b74 100644 --- a/docs/manual/README.md +++ b/docs/manual/README.md @@ -215,22 +215,22 @@ colour that come with it, while keeping the fine detail. Look at it at `Method` chooses how: -- `Best`, the default: two networks, one for smooth areas and one for - edges, blended where each is better. The cleanest skies and the sharpest - lettering, and the slowest. -- `Medium`: one network taught by `Best`. Nearly as clean in smooth areas, - a little softer on hard edges, in about a third of the time. -- `Fast`: a smaller one, taught the same way. Visibly noisier at very high - ISO than the other two, but still far cleaner than none, and quick. +- `Best`, the default: clean skies and sharp lettering, edges kept as + crisp as the camera recorded them. +- `Fast`: a smaller network, taught the same way. Visibly noisier at very + high ISO than `Best`, but still far cleaner than none, and quicker. - `Bilinear`: the camera's ordinary conversion, noise and all. +A photograph last edited with `Medium`, which earlier versions offered, +opens with `Best`. + The photograph shows the camera's ordinary conversion while the network works, with its progress in the bar at the top, and changes when it is -done — on a laptop's graphics card, about two and a half seconds for a -20-megapixel photograph with `Best` and under one with the other two; -longer on a processor alone or on the tablet. The first photograph after -installing waits a few minutes more while the graphics card prepares each -network, once. The result is kept, so a photograph opened again, +done — on a laptop's graphics card, about a second for a 20-megapixel +photograph with `Best`, reading the file included; longer on a processor +alone or on the tablet. After installing, the graphics card spends up to a +quarter of an hour preparing each network, once, in the background; the +photographs developed meanwhile take a little longer. The result is kept, so a photograph opened again, or exported, does not wait a second time, and switching back to a method already used is quick. `Strength` eases it off: below 100 % it puts back some of what was removed, @@ -241,8 +241,8 @@ The lamp and railing of a night frame at ISO 8000, at 1:1, by each method: | Bilinear | Fast | |---|---| | ![The railing and the lamp at ISO 8000, as the camera recorded them](media/develop-denoise-bilinear.png) | ![The same, with the Fast network](media/develop-denoise-fast.png) | -| **Medium** | **Best** | -| ![The same, with the Medium network](media/develop-denoise-medium.png) | ![The same, with the Best network](media/develop-denoise-best.png) | +| **Best** | | +| ![The same, with the Best network](media/develop-denoise-best.png) | | It works on raw files from any camera with the usual colour pattern of red, green and blue squares — not on JPEGs, and not yet on Fujifilm's diff --git a/docs/manual/index.html b/docs/manual/index.html index 2aeec12..6f89d4d 100644 --- a/docs/manual/index.html +++ b/docs/manual/index.html @@ -296,22 +296,21 @@ colour that come with it, while keeping the fine detail. Look at it at 1:1, where noise lives.

Method chooses how:

+

A photograph last edited with Medium, which earlier versions offered, +opens with Best.

The photograph shows the camera's ordinary conversion while the network works, with its progress in the bar at the top, and changes when it is -done — on a laptop's graphics card, about two and a half seconds for a -20-megapixel photograph with Best and under one with the other two; -longer on a processor alone or on the tablet. The first photograph after -installing waits a few minutes more while the graphics card prepares each -network, once. The result is kept, so a photograph opened again, +done — on a laptop's graphics card, about a second for a 20-megapixel +photograph with Best, reading the file included; longer on a processor +alone or on the tablet. After installing, the graphics card spends up to a +quarter of an hour preparing each network, once, in the background; the +photographs developed meanwhile take a little longer. The result is kept, so a photograph opened again, or exported, does not wait a second time, and switching back to a method already used is quick. Strength eases it off: below 100 % it puts back some of what was removed, @@ -319,8 +318,8 @@ as grain without colour, for a picture that does not look too smooth.

The lamp and railing of a night frame at ISO 8000, at 1:1, by each method:

- - + +
BilinearFast
The railing and the lamp at ISO 8000, as the camera recorded themThe same, with the Fast network
MediumBest
The same, with the Medium networkThe same, with the Best network
Best
The same, with the Best network

It works on raw files from any camera with the usual colour pattern of red, green and blue squares — not on JPEGs, and not yet on Fujifilm's diff --git a/docs/manual/media/develop-denoise-medium.png b/docs/manual/media/develop-denoise-medium.png deleted file mode 100644 index bae7a69..0000000 --- a/docs/manual/media/develop-denoise-medium.png +++ /dev/null @@ -1,3 +0,0 @@ -version https://git-lfs.github.com/spec/v1 -oid sha256:1dc3a44ddfb440704d749e96495908fc7834c940e18067efdbbf8123a3f96c77 -size 251267