Changelog

Every release of Manfred, newest first. Earlier versions were developed privately; 0.19.0 is the first public one.

v1.0.1 — 2026-10-14

Added

Each release now carries docs-vX.Y.Z.tar.gz, Manfred’s section of the PITO MD docs site: an overview, install, seven guides (a first scene, building GLBs, the window, path-traced stills, camera clips, filming the window headless, an app’s own generators), the file formats, the command line, the recipe schema, this changelog and the API reference.

Fixed

manfred render --help now gives the draft’s real settings, 12 samples at most and 6 first; it still said 16 and 8 after 0.19.8 lowered them.

v1.0.0 — 2026-10-14

Manfred 1.0.0: the same scene tool as 0.19.8, now with a prebuilt Linux binary and a README that shows what it does.

Added

Each release now carries a ready-to-run Linux x86_64 binary, manfred-vX.Y.Z-linux-x86_64.tar.gz, with a SHA256SUMS file, so installing is one download instead of a build from source.

Changed

The README shows every feature with its own clip or chart: building recipes headless, editing in the window, path-traced renders and camera clips, with the speed and size measured: 39.8 MB against Blender 5.2.2’s 1,236.5 MB, a recipe to GLB in 29 ms, the window’s first frame in 0.88 s and a 4K path-traced final in 7.5 s.
The brand reads PITO MD in the README and the trademark notice.
The version is 1.0.0; scenes render and build exactly as in 0.19.8, and only the version written into each GLB changes.

v0.19.8 — 2026-10-12

Changed

Renders keep the Sobol sampler and spend its gain on speed: a final now adapts to 64 samples a pixel at most (from 128) and a draft to 12 (from 16, at least 6), where Sobol’s noise matches what the PCG sampler gave at 128 and 16. A --spp or recipe spp above 256 sizes the Sobol table to it. Measured on an RX 9060 XT, best of three: the showcase workshop’s 4K final 7.7 s, 6.2 s on the GPU (v0.19.6 12.1 s and 10.7 s); its 4K draft 2.6 s, 1.4 s on the GPU (the same); a 640x400 crop of the showcase hero at 4K 1.2 s, 0.9 s on the GPU (2.1 s and 1.8 s). Pixels differ from v0.19.7’s.
The pfx crates move to v0.28.1, whose Sobol sampler shifts its sequence per pixel, so a render’s mean matches the PCG sampler’s.

Fixed

v0.19.7’s Sobol sampler cost 10 to 20% more GPU time a sample at the old caps and stopped no sooner (118.3 samples a pixel for the workshop’s final under both samplers), so v0.19.7 rendered about 10% slower than v0.19.6; its note that adaptive renders stop sooner did not hold on these scenes.

v0.19.7 — 2026-10-12

Changed

The pfx crates move to v0.28.0, whose tracer samples with Cycles’ tabulated Sobol sequence instead of PCG: the same noise in about half the samples, so adaptive renders stop sooner. Pixels differ from v0.19.6’s; on scenes lit by small or narrow lights the mean sits up to 0.15% lower.
Manfred reports to no one: it sends no crash reports, writes no job records and gives the window no heartbeat. --report-test is gone, and pfx-report left the dependencies.

v0.19.6 — 2026-10-12

Changed

The pfx crates move to v0.27.0, on wgpu 30, and the window to egui 0.36.2. The tracer takes the GPU’s hardware ray queries when the adapter offers them (on Vulkan, VK_KHR_ray_query) and its own BVH otherwise. Measured on an RX 9060 XT, best of three, same samples a pixel as v0.19.5: the showcase workshop’s 4K final 12.1 s (from 39.2 s), 10.7 s on the GPU (from 37.4 s); its 4K draft 2.6 s (from 6.6 s); a 640x400 crop of the showcase hero at 4K 2.1 s (from 8.1 s). The first render on a machine waits a few seconds while the driver compiles the ray-query shader; its cache keeps it after that.

v0.19.5 — 2026-10-12

Added

manfred render --crop writes a .png or --master through a finish with bloom and vignette: it traces and finishes a margin around the rectangle (the bloom’s reach, aligned to its 2×2 blocks), lays the vignette by the full frame’s coordinates, then cuts the rectangle out. A frame whose sides are odd, overlay text, grain, a film dither and outlines are still refused for a .png or master crop.
--denoise on denoises an .exr or .hdr too, a crop included; without it they stay the raw trace, as before.

Changed

A denoised crop’s traced window is aligned to 16 pixels of the full frame, the denoiser’s pooling grid. Measured on the showcase hero at 3840x2160, two 640x400 crops against the full frame’s .png: mean 0.04 and 0.03 levels, p99 1 level, at most 2 at the defaults; mean 0.07 at most with --draft. With --denoise off a finished crop equals the full frame. Full frames render exactly as in v0.19.4.

v0.19.4 — 2026-10-12

Added

A recipe’s [trace] takes spp, threshold and min_samples, the adaptive sampling manfred render uses for it; --spp, --threshold and --min-samples still override them, and --draft ignores them.

Changed

The pfx crates move to a revision whose tracer splits a dense band of the frame into column pieces instead of dropping to one lane, so 4K frames with large busy areas trace several times faster. The traced pixels are unchanged.

v0.19.3 — 2026-10-12

Changed

manfred render denoises on the GPU with pfx’s port of Intel Open Image Denoise, read from the tracer’s textures: about 2 s at 4K instead of about 5 s on the CPU. The pixels of a denoised .png change, so a plate and its crops come from one version.
A denoised crop grows by 96 pixels around the rectangle, the new denoiser’s reach; it matches the full frame closely rather than bit for bit, and --denoise off, .exr and .hdr crops stay exact.
manfred clip --trace keeps one denoiser across its frames.
The pfx crates move to pfx 0.26.0’s revision.
Measured on an RX 9060 XT, release build, the showcase workshop’s hero camera at 3840x2160 with the GPU to itself: a final at the defaults in 39.5 s end to end (37.4 s on the GPU, 118.3 samples a pixel on average), from 57.4 s in v0.19.2; a draft in 6.5 s (5.1 s on the GPU), from about 12 s.

v0.19.2 — 2026-10-12

Added

manfred render --crop x,y,w,h renders one rectangle of the frame with the same pixels the full frame has there, for variants of one part of a still.

Changed

manfred render --progress json gives a fraction and an ETA with adaptive sampling too, from the tracer’s rounds.
The pfx crates move to a newer revision whose tracer starts every trace at full width, several times faster on short renders, with the same pixels.

v0.19.1 — 2026-10-12

Added

manfred render samples adaptively: each pixel stops once its noise is under --threshold (0.01), after --min-samples (16), with --spp as the most any pixel takes; --threshold 0 runs every pixel to --spp.
manfred render --draft for a quick look: 16 samples at most, threshold 0.05.
manfred render --master <file.png> writes a 16-bit PNG master of the finished picture.
manfred render --progress json prints progress as JSON lines on stderr, with the fraction and an ETA when every pixel runs to --spp, and the samples a pixel took when it ends.

Changed

manfred render defaults to 128 samples at most, adaptive and denoised.
manfred render dithers its 8-bit PNG and writes its PNGs with fast compression.
The window’s render keeps a fixed sample count, so its progress still counts samples.
GPU tests run in pgpu’s clip class.

v0.19.0 — 2026-10-08

Added

The code is open source under MIT, with NOTICE, TRADEMARKS.md, a code of conduct and a security policy.
manfred gui --clip films the offscreen window playing a tape into an x264 master, a poster and a manifest.
bin/gate --public and CI run the gate that needs no GPU, a licence check included.

Changed

Command-line errors follow clap’s form: a short error on stderr and exit 2, and a bare manfred prints that error.
Every dependency comes from a public source over https, the pfx crates from pfx v0.25.0.
Scenes are written and opened through pfx-scene, and crash reports and job progress go through pfx-report.
The root package’s jobs and report features are gone.

Fixed

A recipe opened from a folder’s picker saves its edits.