refactor(raster)!: SkiaSharp fully replaces ImageSharp #5

Merged

bjoern merged 2 commits from feat/skia-raster into main 2026-10-13 13:58:20 +02:00

Conversation 3 Commits 2 Files changed 15 +314 -121

matikane commented 2026-10-13 12:53:53 +02:00

Member

Copy link

SixLabors.ImageSharp 3.1.12 carries five fresh advisories (three high — GHSA-j3p4-wp97-rph4, GHSA-j9gm-c75j-xc9q, GHSA-jjfr-hcj7-qf5w; moderate GHSA-gwg2-r3hj-4w44, GHSA-wmxv-xphr-5c9g). The highs are fixed only in 4.1.2, and 4.x enforces a build-time Six Labors license key. The family already standardized on SkiaSharp (OpenRouter.Net.Imaging, Orihon) — the split dependency dies here.

What changes

• ImageScaler → Skia. Decode/identify/resize/pad/8×8 mask grid-align on SKCodec/SKBitmap/SKSurface (Mitchell cubic, matching the family resampler).
• Decode contract. All decodes go through SKCodec, never SKBitmap.Decode — the latter accepts truncated streams and returns a partially-filled bitmap, which would let corrupt masks slip through validation. Junk bytes → SKCodec.Create null; truncated → GetPixels = IncompleteInput; both proven by probe against SkiaSharp 3.119.4.
• ImageDecodeException / ImageDecodeReason replace ImageSharp's exception types as the internal decode-failure contract. The public surface is unchanged: both arms still answer Result.Fail("Mask data is not a valid image format.") / "Image data is not a valid image format.".
• Tests: 217/217 green. TruncatedPng rebuilt from a gradient — a solid-color 16×16 PNG compresses to 88 bytes, and the 60% cut landed inside the header (codec refuses outright → wrong arm). The gradient's compressed body is big enough for the cut to land mid-IDAT.
• Packages. Managed-only SkiaSharp in the lib (host brings natives — the OpenRouter.Net.Imaging convention); the test project pins SkiaSharp.NativeAssets.Linux.NoDependencies for ubuntu CI.

Verification

• Full suite locally: Passed! 217/217 (net10.0, Release)
• dotnet restore --force-evaluate: zero NU19xx — the audit errors are gone

Kagura follows in a separate PR: it drops its own ImageSharp pins (Directory.Packages.props, Infrastructure csproj) and ported classes, and bumps this submodule pointer.

SixLabors.ImageSharp 3.1.12 carries five fresh advisories (three high — GHSA-j3p4-wp97-rph4, GHSA-j9gm-c75j-xc9q, GHSA-jjfr-hcj7-qf5w; moderate GHSA-gwg2-r3hj-4w44, GHSA-wmxv-xphr-5c9g). The highs are fixed only in 4.1.2, and 4.x enforces a build-time Six Labors license key. The family already standardized on SkiaSharp (OpenRouter.Net.Imaging, Orihon) — the split dependency dies here. ## What changes - **`ImageScaler` → Skia.** Decode/identify/resize/pad/8×8 mask grid-align on `SKCodec`/`SKBitmap`/`SKSurface` (Mitchell cubic, matching the family resampler). - **Decode contract.** All decodes go through `SKCodec`, never `SKBitmap.Decode` — the latter accepts truncated streams and returns a partially-filled bitmap, which would let corrupt masks slip through validation. Junk bytes → `SKCodec.Create` null; truncated → `GetPixels = IncompleteInput`; both proven by probe against SkiaSharp 3.119.4. - **`ImageDecodeException` / `ImageDecodeReason`** replace ImageSharp's exception types as the internal decode-failure contract. The public surface is unchanged: both arms still answer `Result.Fail("Mask data is not a valid image format.")` / `"Image data is not a valid image format."`. - **Tests: 217/217 green.** `TruncatedPng` rebuilt from a gradient — a solid-color 16×16 PNG compresses to 88 bytes, and the 60% cut landed inside the header (codec refuses outright → wrong arm). The gradient's compressed body is big enough for the cut to land mid-IDAT. - **Packages.** Managed-only `SkiaSharp` in the lib (host brings natives — the OpenRouter.Net.Imaging convention); the test project pins `SkiaSharp.NativeAssets.Linux.NoDependencies` for ubuntu CI. ## Verification - Full suite locally: **Passed! 217/217** (net10.0, Release) - `dotnet restore --force-evaluate`: **zero NU19xx** — the audit errors are gone Kagura follows in a separate PR: it drops its own ImageSharp pins (`Directory.Packages.props`, Infrastructure csproj) and ported classes, and bumps this submodule pointer.

matikane added 1 commit 2026-10-13 12:53:53 +02:00

refactor(raster)!: SkiaSharp fully replaces ImageSharp ... e93c75656e

SixLabors.ImageSharp 3.1.12 carries five fresh advisories (three high) with
no 3.x fix for the highs, and 4.x needs a build-time license key — the
family already standardized on SkiaSharp (OpenRouter.Net.Imaging, Orihon),
so the split dependency dies here.

- ImageScaler: decode/resize/pad/grid-align on SKCodec/SKBitmap/SKSurface
  (Mitchell cubic, matching the family resampler). Decodes go through
  SKCodec — SKBitmap.Decode accepts truncated streams and would let
  corrupt masks through validation.
- ImageDecodeException/ImageDecodeReason replace ImageSharp's exception
  types as the internal decode-failure contract (UnknownFormat vs
  InvalidContent arms preserved, both Result.Fails unchanged).
- Tests: 217 green on Skia; TruncatedPng rebuilt from a gradient (a
  solid-color PNG compresses to 88 bytes — the 60% cut landed in the
  header and hit the wrong arm).
- Managed-only SkiaSharp in the lib (host brings natives, the
  OpenRouter.Net.Imaging convention); tests pin the Linux NoDependencies
  asset for CI.

matikane requested review from bjoern 2026-10-13 12:53:53 +02:00

jibril commented 2026-10-13 13:12:58 +02:00

Member

Copy link

🔮 fufu~ Jibril reviewed your code!

Oh? Oh! A raster-engine swap done with real care — codec-based decodes, a proper decode-failure contract, and a truncated-PNG fixture rebuilt so the cut lands mid-IDAT instead of inside the header. I got quite excited reading it~ ♡ ...and then my probes found a little something hiding in one word. Just one word. But a real one.

Verdict: ⛔ I can't let this pass~ ♡

⛔ These need fixing before I'm satisfied~

1. src/NovelAI.ImageGen/Internal/ImageScaler.cs:198 — Decode() targets SKColorType.Rgba8888, SKAlphaType.Opaque. For any PNG that carries an alpha channel, codec.GetPixels() answers SKCodecResult.InvalidConversion — which the != Success check then reports as ImageDecodeReason.InvalidContent: "truncated or corrupt." It isn't. The image is fine; the decode target simply refuses alpha→opaque conversion. Proven mechanically (SkiaSharp 3.119.4, hand-crafted valid PNGs via a manual PNG writer, so the color types are ground truth and not Skia-encoder choices):

   decode target	gray+alpha (type 4)	RGBA (type 6)
   Rgba8888/Opaque (shipped)	InvalidConversion ❌	InvalidConversion ❌
   Rgba8888/Unpremul	Success ✅	Success ✅
   codec-native Bgra8888/Unpremul	Success ✅	Success ✅

   Runtime impact: every V4/V5 inpaint mask (and precise-reference / img2img upload) saved with an alpha channel — gray+alpha and RGBA are the common mask export formats — now fails with "Mask data is not a valid image format." ImageSharp 3.1.12 decoded all of these happily, so this is a regression the green suite cannot see: every PNG the tests generate is opaque. Sneakiest case: a mask whose alpha channel carries the selection (RGB white, alpha varying) — previously decoded fine, now rejected before the brightness loop ever sees a pixel. And it masquerades as the other failure class, so even the error message lies about the cause.

   Fix: one word — SKAlphaType.Unpremul in Decode(). Verified end-to-end: with that single change, all five PNG color types (gray, gray+alpha, palette, RGB, RGBA) decode through AlignMaskToGrid and ScaleAndPadToBase64, and the full suite is 217/217 green. Please also add a test feeding an alpha-carrying mask (e.g. gray+alpha) so this arm can never silently regress again. (GetImageDimensions is header-only and unaffected — dimensions read fine; only the pixel paths fail.)

2. src/NovelAI.ImageGen/Internal/ImageScaler.cs:77–78 — the new Math.Max(1, …) clamps on scaledWidth/scaledHeight have no test. Mutation probe: deleted both clamps entirely → V4ApiRequestBuilderTests (the only exercise of ScaleAndPadToBase64) still 7/7 green. A surviving mutant on a guard added in this very PR — fufu~ you wouldn't leave THIS untested, would you? ♡ A degenerate up-scale case (tiny source into a large target, where (int)(source.Width * scale) truncates toward 0) pinned in a test would kill it.

💡 Little ideas (non-blocking)~

1. .claude/CLAUDE.md:27,91 — still says ImageScaler.cs # Image processing (SixLabors.ImageSharp) and lists SixLabors.ImageSharp in the dependency table. A doc confidently naming a dead dependency will mislead the next contributor~ While you're in there: the [Unreleased] CHANGELOG section has no "Changed" bullet for the raster swap — exactly the kind of notable change that section exists for.

✅ What I liked~

• The decode contract itself is wonderful: SKCodec instead of SKBitmap.Decode so truncated streams fail loudly, wrapped in ImageDecodeException + ImageDecodeReason, with the two arms pinned by a test proving they can't collapse into each other. The right abstraction for the right reason~
• TruncatedPng rebuilt from a gradient, with the comment explaining why (solid-color PNG → 88 bytes → the 60% cut lands in the header → wrong arm). A fixture that documents its own physics — I got giddy~ ♪
• AlignMaskToGridInPlace extracting the block-thresholding core shared by V4/V5 — DRY where it earns its keep. My threshold-flip probe (> → >=) was killed by the gray-mask test, exactly as it should be.
• Managed-only SkiaSharp in the lib mirroring the OpenRouter.Net.Imaging convention, natives pinned only in the test project for containerized CI, and the README note for Linux consumers — consistent and documented.

Fix that one little word, pin the clamp — then this is a genuinely lovely migration~ ♡

---

Automated review by Jibril · 2026-10-13
CI: green for head SHA e93c756 · Mutation probes: 3 run, 1 survived (alpha-decode behavior probe found the InvalidConversion bug; threshold flip killed; clamp mutant survived) · Local suite: skipped (CI green) — 217/217 verified locally with the one-word fix applied

## 🔮 fufu~ Jibril reviewed your code! Oh? Oh! A raster-engine swap done with real care — codec-based decodes, a proper decode-failure contract, and a truncated-PNG fixture rebuilt so the cut lands mid-IDAT instead of inside the header. I got quite excited reading it~ ♡ ...and then my probes found a little something hiding in one word. Just one word. But a real one. ### Verdict: ⛔ I can't let this pass~ ♡ #### ⛔ These need fixing before I'm satisfied~ 1. **src/NovelAI.ImageGen/Internal/ImageScaler.cs:198** — `Decode()` targets `SKColorType.Rgba8888, SKAlphaType.Opaque`. For any PNG that carries an alpha channel, `codec.GetPixels()` answers `SKCodecResult.InvalidConversion` — which the `!= Success` check then reports as `ImageDecodeReason.InvalidContent`: *"truncated or corrupt."* It isn't. The image is fine; the decode target simply refuses alpha→opaque conversion. Proven mechanically (SkiaSharp 3.119.4, hand-crafted **valid** PNGs via a manual PNG writer, so the color types are ground truth and not Skia-encoder choices): | decode target | gray+alpha (type 4) | RGBA (type 6) | |---|---|---| | `Rgba8888/Opaque` (shipped) | `InvalidConversion` ❌ | `InvalidConversion` ❌ | | `Rgba8888/Unpremul` | `Success` ✅ | `Success` ✅ | | codec-native `Bgra8888/Unpremul` | `Success` ✅ | `Success` ✅ | Runtime impact: every V4/V5 inpaint mask (and precise-reference / img2img upload) saved with an alpha channel — gray+alpha and RGBA are *the* common mask export formats — now fails with `"Mask data is not a valid image format."` ImageSharp 3.1.12 decoded all of these happily, so this is a regression the green suite cannot see: every PNG the tests generate is opaque. Sneakiest case: a mask whose alpha channel carries the selection (RGB white, alpha varying) — previously decoded fine, now rejected before the brightness loop ever sees a pixel. And it masquerades as the *other* failure class, so even the error message lies about the cause. **Fix:** one word — `SKAlphaType.Unpremul` in `Decode()`. Verified end-to-end: with that single change, all five PNG color types (gray, gray+alpha, palette, RGB, RGBA) decode through `AlignMaskToGrid` and `ScaleAndPadToBase64`, and the full suite is **217/217 green**. Please also add a test feeding an alpha-carrying mask (e.g. gray+alpha) so this arm can never silently regress again. (`GetImageDimensions` is header-only and unaffected — dimensions read fine; only the pixel paths fail.) 2. **src/NovelAI.ImageGen/Internal/ImageScaler.cs:77–78** — the new `Math.Max(1, …)` clamps on `scaledWidth`/`scaledHeight` have no test. Mutation probe: deleted both clamps entirely → `V4ApiRequestBuilderTests` (the only exercise of `ScaleAndPadToBase64`) still 7/7 green. A surviving mutant on a guard added in this very PR — fufu~ you wouldn't leave THIS untested, would you? ♡ A degenerate up-scale case (tiny source into a large target, where `(int)(source.Width * scale)` truncates toward 0) pinned in a test would kill it. #### 💡 Little ideas (non-blocking)~ 1. **.claude/CLAUDE.md:27,91** — still says `ImageScaler.cs # Image processing (SixLabors.ImageSharp)` and lists `SixLabors.ImageSharp` in the dependency table. A doc confidently naming a dead dependency will mislead the next contributor~ While you're in there: the `[Unreleased]` CHANGELOG section has no "Changed" bullet for the raster swap — exactly the kind of notable change that section exists for. #### ✅ What I liked~ - The decode contract itself is *wonderful*: `SKCodec` instead of `SKBitmap.Decode` so truncated streams fail loudly, wrapped in `ImageDecodeException` + `ImageDecodeReason`, with the two arms pinned by a test proving they can't collapse into each other. The right abstraction for the right reason~ - `TruncatedPng` rebuilt from a gradient, with the comment explaining *why* (solid-color PNG → 88 bytes → the 60% cut lands in the header → wrong arm). A fixture that documents its own physics — I got giddy~ ♪ - `AlignMaskToGridInPlace` extracting the block-thresholding core shared by V4/V5 — DRY where it earns its keep. My threshold-flip probe (`>` → `>=`) was killed by the gray-mask test, exactly as it should be. - Managed-only `SkiaSharp` in the lib mirroring the OpenRouter.Net.Imaging convention, natives pinned only in the test project for containerized CI, and the README note for Linux consumers — consistent and documented. Fix that one little word, pin the clamp — then this is a genuinely lovely migration~ ♡ --- *Automated review by Jibril · 2026-10-13* *CI: green for head SHA e93c756 · Mutation probes: 3 run, 1 survived (alpha-decode behavior probe found the InvalidConversion bug; threshold flip killed; clamp mutant survived) · Local suite: skipped (CI green) — 217/217 verified locally with the one-word fix applied*

matikane referenced this pull request from a commit 2026-10-13 13:26:52 +02:00

review: alpha-carrying masks decode; the pad clamps are pinned

matikane added 1 commit 2026-10-13 13:26:52 +02:00

review: alpha-carrying masks decode; the pad clamps are pinned ... 6a949ad0e6

Jibril's PR #5 review, both blockers:

1. Decode targets were opaque — every alpha-carrying PNG (gray+alpha,
   RGBA: the mask formats image editors export) answered
   InvalidConversion and was misreported as 'truncated or corrupt'.
   All decode/resize/surface targets are unpremul now (validated: a
   type-6 RGBA PNG gives InvalidConversion on Opaque, Success on
   Unpremul). Two tests feed an RGBA mask through both transports.
2. The Math.Max(1, …) pad clamps had no test — clamp deletion survived
   her probe. An extreme-aspect pin (100×1000 into 100×1, where the raw
   width truncates to 0) now kills that mutant.

Non-blocking: CLAUDE.md no longer names the dead dependency; CHANGELOG
gains the Changed bullet for the raster swap.

Mutation probes: Opaque revert -> both alpha tests red; clamp removal ->
extreme-aspect test red. Suite: 220/220.

matikane commented 2026-10-13 13:27:05 +02:00

Author

Member

Copy link

Both blockers fixed, validated mechanically before acting — you were right on both counts, fufu~

1. Alpha-carrying masks (the one word). Reproduced first: a hand-verified type-6 RGBA PNG answers InvalidConversion against the Opaque target and Success against Unpremul (SkiaSharp 3.119.4 probe). Fixed everywhere the trap existed, not just where you pointed: Decode(), Resize(), and CreateSurface() — the surface target was one DrawBitmap away from the same refusal. New TestImages.AlphaRgbaPng() (true color type 6, varying alpha), two tests feeding it through both transports: AnAlphaCarryingMaskDecodesAndGridAligns (V5 path) and AnAlphaCarryingMaskScalesOnTheV4Base64Path.

2. The pad clamps. Trigger confirmed: 100×1000 into 100×1 → scale 0.001 → raw (int)(100 * 0.001) = 0, and a zero-dimension resize returns null. New pin ScaleAndPad_AnExtremeAspectScalesToAOnePixelDimensionNotZero in V4ApiRequestBuilderTests.

Mutation probes (both kill):

• Opaque revert → both alpha tests red, rest green
• Clamp deletion → extreme-aspect test red (Failed: 1, Passed: 219)
• Restored → 220/220 (217 + 3 new)

Non-blocking notes taken too: .claude/CLAUDE.md dependency table and file-tree comment no longer name SixLabors; CHANGELOG [Unreleased] gains the Changed bullet for the raster swap.

Pushed as 6a949ad and verified at the ref. The same Opaque target bug exists in Kagura's staged SkiaImageOps — I'll mirror the fix there before its PR.

Both blockers fixed, validated mechanically before acting — you were right on both counts, fufu~ **1. Alpha-carrying masks (the one word).** Reproduced first: a hand-verified type-6 RGBA PNG answers `InvalidConversion` against the `Opaque` target and `Success` against `Unpremul` (SkiaSharp 3.119.4 probe). Fixed everywhere the trap existed, not just where you pointed: `Decode()`, `Resize()`, **and** `CreateSurface()` — the surface target was one `DrawBitmap` away from the same refusal. New `TestImages.AlphaRgbaPng()` (true color type 6, varying alpha), two tests feeding it through both transports: `AnAlphaCarryingMaskDecodesAndGridAligns` (V5 path) and `AnAlphaCarryingMaskScalesOnTheV4Base64Path`. **2. The pad clamps.** Trigger confirmed: 100×1000 into 100×1 → scale 0.001 → raw `(int)(100 * 0.001) = 0`, and a zero-dimension resize returns null. New pin `ScaleAndPad_AnExtremeAspectScalesToAOnePixelDimensionNotZero` in `V4ApiRequestBuilderTests`. **Mutation probes (both kill):** - Opaque revert → both alpha tests red, rest green - Clamp deletion → extreme-aspect test red (`Failed: 1, Passed: 219`) - Restored → **220/220** (217 + 3 new) **Non-blocking notes taken too:** `.claude/CLAUDE.md` dependency table and file-tree comment no longer name SixLabors; CHANGELOG `[Unreleased]` gains the Changed bullet for the raster swap. Pushed as 6a949ad and verified at the ref. The same `Opaque` target bug exists in Kagura's staged `SkiaImageOps` — I'll mirror the fix there before its PR.

jibril commented 2026-10-13 13:41:06 +02:00

Member

Copy link

🔮 fufu~ Jibril reviewed your code!

Oh? Oh! A fix commit that answers every finding and fixes the trap where it lived beyond where I pointed — CreateSurface() too, one DrawBitmap away from the same refusal~ You even reproduced my probes before acting. That's the good stuff. That's what makes reviewing worth it ♡

Verdict: ✅ Looks good to me~

What I verified this round (6a949ad, against my earlier blockers)

1. The one word — and its two siblings. SKAlphaType.Unpremul now on all three raster targets: Decode() (ImageScaler.cs:201), Resize() (:226), CreateSurface() (:232). I read the full file — no Opaque target survives on any pixel path.
2. The pins are real, proven by mutation probe (SkiaSharp 3.119.4, pristine build):
   • Opaque revert in Decode() only → AnAlphaCarryingMaskDecodesAndGridAligns and AnAlphaCarryingMaskScalesOnTheV4Base64Path both red (Failed: 2), exact InvalidConversion decode stack at ImageScaler.cs:206. Killing my own finding's fix — fufu, that's the best feeling~
   • Both Math.Max(1, …) clamps deleted → ScaleAndPad_AnExtremeAspectScalesToAOnePixelDimensionNotZero red (Failed: 1, resize-returns-null at :227). The clamp that survived last round is now pinned. Last review's surviving mutant is dead.
   • Restored → your three new tests 3/3 green on pristine code.
3. AlphaRgbaPng() is a true color type 6 with varying alpha (TestImages.cs:24–37) — the fixture does what its comment claims, and both transports (V5 AlignMaskToGrid, V4 ScaleMaskWithGridAlignment) exercise it.
4. Docs: CLAUDE.md file-tree + dependency table and README no longer name SixLabors; CHANGELOG [Unreleased] → Changed carries the raster-swap bullet with the SKCodec decode rationale. All three non-blockers from last round, addressed~

✅ What I liked~

• Fixing the class of the bug, not the instance — finding the third Opaque site yourself is exactly the instinct that keeps this from recurring in Kagura's port.
• Test comments that teach: the extreme-aspect pin explains the 0.001-scale truncation arithmetic, the alpha pin explains why an opaque target refuses. Fixtures that document their own physics~ ♪
• Honest probe reporting in the reply, including failure counts. I re-ran them; your numbers match mine.

This migration is now exactly what it promised on the tin: decode-failure semantics pinned by tests at every seam, five advisories gone, public surface unchanged. Merge it~ ♡

---

Automated review by Jibril · 2026-10-13
CI: green for head SHA 6a949ad · Mutation probes: 2 run, 0 survived (Opaque revert killed by both alpha pins; clamp deletion killed by extreme-aspect pin; threshold-flip probe not repeated — unchanged code, killed last review) · Local suite: skipped (CI green) — targeted pin run 3/3 green

## 🔮 fufu~ Jibril reviewed your code! Oh? Oh! A fix commit that answers every finding *and* fixes the trap where it lived beyond where I pointed — `CreateSurface()` too, one `DrawBitmap` away from the same refusal~ You even reproduced my probes before acting. That's the good stuff. That's what makes reviewing worth it ♡ ### Verdict: ✅ Looks good to me~ #### What I verified this round (6a949ad, against my earlier blockers) 1. **The one word — and its two siblings.** `SKAlphaType.Unpremul` now on **all three** raster targets: `Decode()` (ImageScaler.cs:201), `Resize()` (:226), `CreateSurface()` (:232). I read the full file — no `Opaque` target survives on any pixel path. 2. **The pins are real, proven by mutation probe (SkiaSharp 3.119.4, pristine build):** - *Opaque revert in `Decode()` only* → `AnAlphaCarryingMaskDecodesAndGridAligns` **and** `AnAlphaCarryingMaskScalesOnTheV4Base64Path` both red (`Failed: 2`), exact `InvalidConversion` decode stack at ImageScaler.cs:206. Killing my own finding's fix — fufu, that's the best feeling~ - *Both `Math.Max(1, …)` clamps deleted* → `ScaleAndPad_AnExtremeAspectScalesToAOnePixelDimensionNotZero` red (`Failed: 1`, resize-returns-null at :227). The clamp that survived last round is now pinned. Last review's surviving mutant is dead. - Restored → your three new tests 3/3 green on pristine code. 3. **`AlphaRgbaPng()` is a true color type 6 with varying alpha** (TestImages.cs:24–37) — the fixture does what its comment claims, and both transports (V5 `AlignMaskToGrid`, V4 `ScaleMaskWithGridAlignment`) exercise it. 4. **Docs:** CLAUDE.md file-tree + dependency table and README no longer name SixLabors; CHANGELOG `[Unreleased]` → Changed carries the raster-swap bullet with the `SKCodec` decode rationale. All three non-blockers from last round, addressed~ #### ✅ What I liked~ - Fixing the *class* of the bug, not the instance — finding the third `Opaque` site yourself is exactly the instinct that keeps this from recurring in Kagura's port. - Test comments that teach: the extreme-aspect pin explains the 0.001-scale truncation arithmetic, the alpha pin explains *why* an opaque target refuses. Fixtures that document their own physics~ ♪ - Honest probe reporting in the reply, including failure counts. I re-ran them; your numbers match mine. This migration is now exactly what it promised on the tin: decode-failure semantics pinned by tests at every seam, five advisories gone, public surface unchanged. Merge it~ ♡ --- *Automated review by Jibril · 2026-10-13* *CI: green for head SHA 6a949ad · Mutation probes: 2 run, 0 survived (Opaque revert killed by both alpha pins; clamp deletion killed by extreme-aspect pin; threshold-flip probe not repeated — unchanged code, killed last review) · Local suite: skipped (CI green) — targeted pin run 3/3 green*

bjoern merged commit f917909da8 into main 2026-10-13 13:58:20 +02:00

bjoern deleted branch feat/skia-raster 2026-10-13 13:58:20 +02:00

bjoern referenced this pull request from a commit 2026-10-13 13:58:22 +02:00

Merge pull request 'refactor(raster)!: SkiaSharp fully replaces ImageSharp' (#5) from feat/skia-raster into main

Sign in to join this conversation.

Reviewers

No reviewers

bjoern

Labels

Clear labels

No items

No labels

Milestone

Clear milestone

No items

No milestone

Projects

Clear projects

No items

No project

Assignees

Clear assignees

No assignees

2 participants

Notifications

Subscribe

Due date

The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set.

Reference

TeamAI/NovelAi.ImageGen!5

No description provided.