Replace SixLabors.ImageSharp with a self-contained image decoder

[matt-edmondson]

Closes #354.

The answer to the question

Yes to the second option. ktsu.ImGui.App now decodes images itself and has no imaging dependency at all.

The dependency was only ever doing three things: decode a file to RGBA8, crop, and scale. ImageSharp moved to a split licence at 4.0, which broke the build for anyone letting the version float (#230) and left the package pinned to 3.1.x indefinitely — a pin nobody can lift. Swapping in another package (StbImageSharp, ImageMagick bindings) trades one licence and one supply-chain surface for another; writing the ~600 lines of decoding this library actually uses removes the question. The repo already has the precedent: ImGui.App.Testing/Bitmap32.cs hand-rolls PNG encoding over ZLibStream for exactly this reason.

What's here

ImGui.App/Images/ decodes to ImagePixels — a tightly packed, straight-alpha RGBA8 buffer with no row padding, which is exactly what UploadTextureRGBA wants.

Format	Covered	Not covered
PNG	All five colour types, bit depths 1/2/4/8/16, palettes, tRNS in all three forms, Adam7 interlacing, all five scanline filters	Nothing in the base format
JPEG	Baseline and extended sequential (SOF0/SOF1) and progressive (SOF2) Huffman, any sampling factors, restart intervals, greyscale and colour, the Adobe transform flag	Arithmetic coding, lossless and hierarchical modes, CMYK/YCCK, 12-bit samples
BMP	Core and info headers of every version, 1/4/8/16/24/32 bit, top-down and bottom-up, BI_RGB and BI_BITFIELDS	BI_RLE4/BI_RLE8, embedded PNG/JPEG payloads
TGA	Colour-mapped, true-colour and greyscale at 8/15/16/24/32 bit, raw and run-length encoded, the descriptor's origin and attribute-bit flags	Nothing in common use

The format comes from the file's own leading bytes, never its extension, so a mislabelled file still loads. Anything unrecognised or malformed raises InvalidImageDataException naming what was found. Inflation uses System.IO.Compression.ZLibStream from the BCL, so nothing new is pulled in.

ImageResampler replaces the Welch resize behind SetWindowIcon: a separable Lanczos-3 filter whose support widens by the reduction factor when downscaling, so shrinking averages rather than point-samples. It runs on premultiplied alpha and unpremultiplies afterwards, so the colour of fully transparent pixels does not bleed into their visible neighbours — which the old path did not do.

Verification

Decoder output was compared against libpng/libjpeg (through Pillow) over a matrix of encodings:

• PNG — RGBA8, RGB8, greyscale 1/8/16-bit, greyscale+alpha, palette, palette+tRNS, Adam7 interlaced: bit-exact, all of them.
• BMP — 24-bit, 32-bit, 8-bit palette, 1-bit: bit-exact. (The one intentional difference: a 32-bit BI_RGB BMP's alpha is honoured, with an all-zero channel treated as opaque, matching stb_image. Pillow discards it. Checked against the original RGBA source, our output is bit-exact and Pillow's is the lossy one.)
• TGA — 24/32-bit, raw and RLE, greyscale: bit-exact.
• JPEG — 4:4:4, 4:2:2, 4:2:0, greyscale, progressive, restart markers, odd dimensions, high-frequency noise: mean channel error 0.01–0.26, max 3. That is the range reference decoders differ by among themselves; chroma is upsampled by bilinear interpolation on half-offset sample centres, which for the usual 2x factors is the same triangle filter libjpeg applies.

New tests live in tests/ImGui.App.Tests/Images/ (76 cases). TestImageBuilder encodes PNG, BMP and TGA in memory so the decoders can be driven over their whole feature matrix — every colour type, every bit depth, every scanline filter, both row orders, both TGA packet kinds — without checking binary fixtures into the repo. The JPEG cases, which need a real encoder, are small base64 constants.

CI on 78dde73: 1563 tests, 0 failures, all 18 test projects green across Linux, Windows and macOS, plus a green iOS build. That run is the one that matters most here: the six demo UI suites decode the repository's real PNGs through the actual render path, so the new decoder is exercised against icon.png, ktsu.png and the trevor-*.png set rather than only against synthetic fixtures. SonarQube's quality gate passed at 83.9% coverage on new code and 0% duplication.

Performance

Measured on a 2048×2048 image:

	before	after
PNG decode	305 ms	75 ms
JPEG decode	—	396 ms
SetWindowIcon icon set	—	318 ms

PNG got faster than ImageSharp's path here because scanline expansion writes straight into the destination span with a memcpy fast path for 8-bit truecolour. JPEG is slower than libjpeg (floating-point IDCT, no SIMD) but well inside what a texture load can absorb; a DC-only block fast path covers the flat regions that dominate real photographs.

Breaking change

ImGuiApp.UseImageBytes(Image<Rgba32>, Action<byte[]>) is now UseImageBytes(ImagePixels, Action<byte[]>). That parameter was the only place the dependency reached the public surface, so removing the dependency cannot be done compatibly. Callers passing an Image<Rgba32> should load through ImageDecoder instead. Nothing else in the public API changed, and GetOrLoadTexture, TryGetTexture, SetWindowIcon and the texture cache are untouched from a caller's point of view.

Also worth stating plainly: GIF, WebP and TIFF are gone. ImageSharp decoded them and this does not. They do not appear in any icon or texture in this repo or its examples, and a file in one of those formats now fails with a message that names the format rather than silently — but if anyone is loading them, this is the regression to know about.

Docs

CLAUDE.md gains an "Image decoding" section with the support matrix and the three behaviours worth knowing before changing any of it (16-bit truncation, the all-zero-alpha heuristic, and JPEG's whole-image coefficient buffer). ImGui.App/README.md gains an "Image Loading" section with the same matrix and a standalone usage example. ImageSharp is removed from every dependency list.

🤖 Generated with Claude Code

https://claude.ai/code/session_01DCFaH3HSDMKHTmrQB7LcSP

[sonarqubecloud]

Quality Gate passed

Issues
37 New issues
0 Accepted issues

Measures
0 Security Hotspots
83.9% Coverage on New Code
0.0% Duplication on New Code

See analysis details on SonarQube Cloud