Skip to content
CAI
Software that uses CAICheck a score

VincentGourbin/flux-2-swift-mlx

58.5

Adequate · 16 November 2026

43.7k

lines of production code

Swift

primary language

1

measurement over time

CAI band scale
CAI lens gauges

What this system is

Behavioural changes

1 commit (0 fixes) modifying Sources/FluxGemma4VLM, Sources/FluxTextEncoders/VLM

A change to existing behaviour in Sources/FluxGemma4VLM, Sources/FluxTextEncoders/VLM — 1 commit, 4 files: “feat(vlm): provider sélectionnable pour l'enrichissement — Gemma 4 E2B à côté de Qwen 3.5 (\#122)” (refs \#122).

Sources/FluxGemma4VLM, Sources/FluxTextEncoders/VLM · medium confidence · unverified

1 commit (0 fixes) modifying scripts

A change to existing behaviour in scripts — 1 commit, 1 file: “feat: Auto-enable DOP when trigger\_word is set, fix learning curve SVG”.

scripts · medium confidence · unverified

10 commits (1 fix) modifying Sources/Flux2Chains

A change to existing behaviour in Sources/Flux2Chains — 10 commits (1 fix), 7 files: “fix(pipeline): defer CoreGraphics draws to Default-QoS worker (\#92)”; “docs(chains): DocC comments on init/run, top-level README section”; “feat(chains): VLM-aware prompt enrichment + Vision-based subject mask (\#93)”; and 7 more (refs \#92, \#93, \#112, \#122, \#87, …).

Sources/Flux2Chains · medium confidence · unverified

11 commits (1 fix) modifying Sources/Flux2Core/Transformer

A change to existing behaviour in Sources/Flux2Core/Transformer — 11 commits (1 fix), 11 files: “feat: Quality fixes, performance profiling, and prompt upsampling (\#6)”; “refactor: Simplify LoRA training system and add learning curve visualization”; “feat(training): Implement layer-wise gradient checkpointing for transformer blocks”; and 8 more (refs \#6, \#29, \#62, \#17).

Sources/Flux2Core/Transformer · medium confidence · unverified

11 commits (1 fix) modifying Sources/Flux2Core/Utils

A change to existing behaviour in Sources/Flux2Core/Utils — 11 commits (1 fix), 5 files: “feat: Quality fixes, performance profiling, and prompt upsampling (\#6)”; “feat: Add profiling toolkit with Chrome Trace export and benchmarking (\#75)”; “refactor: Migrate profiling to swift-mlx-profiler package (\#77)”; and 8 more (refs \#6, \#75, \#77, \#118, \#114).

Sources/Flux2Core/Utils · medium confidence · unverified

11 commits (2 fixes) modifying Sources/Flux2App/ViewModels

A change to existing behaviour in Sources/Flux2App/ViewModels — 11 commits (2 fixs), 6 files: “feat: Expose Klein 4B/9B Base models in public API”; “feat: Add custom modelsDirectory for sandboxed macOS apps (\#54)”; “feat: validation image generation during LoRA training”; and 8 more (refs \#54, \#62, \#17, \#76).

Sources/Flux2App/ViewModels · medium confidence · unverified

11 commits (5 fixes) modifying Sources/Flux2App/Views

A change to existing behaviour in Sources/Flux2App/Views — 11 commits (5 fixs), 7 files: “refactor: Migrate profiling to swift-mlx-profiler package (\#77)”; “feat: Add maxReferenceImages property per model”; “fix: Eliminate I2I spatial shift and remove unused strength parameter”; and 8 more (refs \#77, \#17, \#94).

Sources/Flux2App/Views · medium confidence · unverified

12 commits (3 fixes) modifying Sources/FluxTextEncoders

A change to existing behaviour in Sources/FluxTextEncoders — 12 commits (3 fixs), 3 files: “feat: Switch VLM scoring to 0-100 scale, minimum rank 32, stricter criteria”; “feat: Merge FluxTextEncoders library into monorepo”; “refactor: Migrate profiling to swift-mlx-profiler package (\#77)”; and 9 more (refs \#77, \#103, \#122, \#104, \#66, …).

Sources/FluxTextEncoders · medium confidence · unverified

17 commits (2 fixes) modifying Sources/Flux2Core/Training/Configuration

A change to existing behaviour in Sources/Flux2Core/Training/Configuration — 17 commits (2 fixs), 2 files: “fix: Fix VLM scoring during training — isVLMGenerated filter, final step eval (\#74)”; “feat: Add Ostris content/style timestep sampling modes”; “feat: Expose Klein 4B/9B Base models in public API”; and 14 more (refs \#74, \#53, \#62).

Sources/Flux2Core/Training/Configuration · medium confidence · unverified

2 commits (1 fix) modifying Sources/FluxTextEncoders/Model/Qwen3VL

A change to existing behaviour in Sources/FluxTextEncoders/Model/Qwen3VL — 2 commits (1 fix), 5 files: “fix(loading): sixth review pass — load only the verified series, protect in-place relocation”; “feat: Qwen3.5-4B VLM service + LoRA evaluation pipeline (\#66)” (refs \#66).

Sources/FluxTextEncoders/Model/Qwen3VL · medium confidence · unverified

22 commits (6 fixes) modifying Sources/Flux2Core/Training

A change to existing behaviour in Sources/Flux2Core/Training — 22 commits (6 fixs), 4 files: “feat(training): Add maxResolution parameter to prevent OOM during training”; “fix: Fix VLM scoring during training — isVLMGenerated filter, final step eval (\#74)”; “feat: Switch VLM scoring to 0-100 scale, minimum rank 32, stricter criteria”; and 19 more (refs \#74, \#79, \#62, \#122, \#71, …).

Sources/Flux2Core/Training · medium confidence · unverified

27 commits (13 fixes) modifying Sources/Flux2Core/Training/Loop

A change to existing behaviour in Sources/Flux2Core/Training/Loop — 27 commits (13 fixs), 1 file: “feat: Add gradient\_checkpointing config param (not yet implemented)”; “fix: Fix VLM scoring during training — isVLMGenerated filter, final step eval (\#74)”; “fix: Memory cleanup for validation + Klein 9B config fixes”; and 24 more (refs \#74, \#53, \#122, \#118, \#104, …).

Sources/Flux2Core/Training/Loop · medium confidence · unverified

3 commits (0 fixes) modifying Sources/FluxTextEncoders/Embeddings

A change to existing behaviour in Sources/FluxTextEncoders/Embeddings — 3 commits, 5 files: “feat: Merge FluxTextEncoders library into monorepo”; “refactor: Simplify LoRA training system and add learning curve visualization”; “feat: Qwen3.5-4B VLM service + LoRA evaluation pipeline (\#66)” (refs \#66).

Sources/FluxTextEncoders/Embeddings · medium confidence · unverified

3 commits (0 fixes) modifying Sources/FluxTextEncoders/Utils

A change to existing behaviour in Sources/FluxTextEncoders/Utils — 3 commits, 2 files: “feat: Merge FluxTextEncoders library into monorepo”; “refactor: Migrate profiling to swift-mlx-profiler package (\#77)”; “refactor: Quiet logs by default, opt-in verbosity via --verbose” (refs \#77).

Sources/FluxTextEncoders/Utils · medium confidence · unverified

3 commits (1 fix) modifying Sources/Flux2Core/Training/Optimizer

A change to existing behaviour in Sources/Flux2Core/Training/Optimizer — 3 commits (1 fix), 3 files: “feat: Add Lion optimizer as alternative to AdamW for LoRA training (\#53)”; “feat: Add training session control with pause/resume/checkpoint”; “fix(training): update ResumableAdamW for MLX optimizer API (\#95)” (refs \#53, \#95).

Sources/Flux2Core/Training/Optimizer · medium confidence · unverified

32 commits (9 fixes) modifying Sources/Flux2Core/Configuration

A change to existing behaviour in Sources/Flux2Core/Configuration — 32 commits (9 fixs), 6 files: “fix(loading): relocated model reports its real size, add path override hook”; “feat: Add isForInference/isForTraining + fix VAE repo”; “feat: Expose Klein 4B/9B Base models in public API”; and 29 more (refs \#23, \#54, \#29, \#21, \#62, …).

Sources/Flux2Core/Configuration · medium confidence · unverified

38 commits (17 fixes) modifying Sources/Flux2Core/Loading

A change to existing behaviour in Sources/Flux2Core/Loading — 38 commits (17 fixs), 7 files: “fix(api): Keep applyTransformerWeights public, make mapTransformerWeights internal”; “feat: Quality fixes, performance profiling, and prompt upsampling (\#6)”; “fix(loading): relocated model reports its real size, add path override hook”; and 35 more (refs \#6, \#19, \#54, \#107, \#104, …).

Sources/Flux2Core/Loading · medium confidence · unverified

4 commits (1 fix) modifying Sources/FluxTextEncoders/Model

A change to existing behaviour in Sources/FluxTextEncoders/Model — 4 commits (1 fix), 4 files: “feat: Merge FluxTextEncoders library into monorepo”; “feat: On-the-fly transformer quantization with int4 support”; “feat: MLX memory optimizations from LTX-2-MLX research (\#29)”; and 1 more (refs \#29).

Sources/FluxTextEncoders/Model · medium confidence · unverified

4 commits (1 fix) modifying Sources/FluxTextEncoders/Model/Qwen3

A change to existing behaviour in Sources/FluxTextEncoders/Model/Qwen3 — 4 commits (1 fix), 4 files: “feat: Merge FluxTextEncoders library into monorepo”; “refactor: Simplify LoRA training system and add learning curve visualization”; “feat: MLX memory optimizations from LTX-2-MLX research (\#29)”; and 1 more (refs \#29).

Sources/FluxTextEncoders/Model/Qwen3 · medium confidence · unverified

5 commits (0 fixes) modifying Sources/FluxTextEncoders/Configuration

A change to existing behaviour in Sources/FluxTextEncoders/Configuration — 5 commits, 9 files: “feat: Merge FluxTextEncoders library into monorepo”; “feat: Expose default parameters and license info in model registry API (\#23)”; “feat: MLX memory optimizations from LTX-2-MLX research (\#29)”; and 2 more (refs \#23, \#29, \#21, \#66).

Sources/FluxTextEncoders/Configuration · medium confidence · unverified

5 commits (1 fix) modifying Sources/Flux2Core/Scheduler

A change to existing behaviour in Sources/Flux2Core/Scheduler — 5 commits (1 fix), 1 file: “feat: Add JSON config support for LoRA scheduler overrides”; “Initial commit: Flux.2 Swift MLX implementation”; “feat: Image-to-Image support with strength parameter (\#3)”; and 2 more (refs \#3).

Sources/Flux2Core/Scheduler · medium confidence · unverified

5 commits (2 fixes) modifying Sources/Flux2Core

A change to existing behaviour in Sources/Flux2Core — 5 commits (2 fixs), 4 files: “fix: prevent crash when cancelling generation mid-process”; “Initial commit: Flux.2 Swift MLX implementation”; “fix(chains): third review pass on Image Preparation — real crop scale, process-area validation”; and 2 more.

Sources/Flux2Core · medium confidence · unverified

52 commits (13 fixes) modifying Sources/Flux2Core/Pipeline

A change to existing behaviour in Sources/Flux2Core/Pipeline — 52 commits (13 fixs), 3 files: “feat: Quality fixes, performance profiling, and prompt upsampling (\#6)”; “fix(pipeline): defer CoreGraphics draws to Default-QoS worker (\#92)”; “feat: Add profiling toolkit with Chrome Trace export and benchmarking (\#75)”; and 49 more (refs \#6, \#92, \#75, \#19, \#29, …).

Sources/Flux2Core/Pipeline · medium confidence · unverified

6 commits (2 fixes) modifying Sources/FluxTextEncoders/Vision

A change to existing behaviour in Sources/FluxTextEncoders/Vision — 6 commits (2 fixs), 5 files: “feat: Merge FluxTextEncoders library into monorepo”; “refactor: Migrate profiling to swift-mlx-profiler package (\#77)”; “feat: MLX memory optimizations from LTX-2-MLX research (\#29)”; and 3 more (refs \#77, \#29, \#66).

Sources/FluxTextEncoders/Vision · medium confidence · unverified

7 commits (2 fixes) modifying Sources/Flux2Core/VAE

A change to existing behaviour in Sources/Flux2Core/VAE — 7 commits (2 fixs), 3 files: “feat: Quality fixes, performance profiling, and prompt upsampling (\#6)”; “feat: MLX memory optimizations from LTX-2-MLX research (\#29)”; “Initial commit: Flux.2 Swift MLX implementation”; and 4 more (refs \#6, \#29, \#3, \#106, \#76).

Sources/Flux2Core/VAE · medium confidence · unverified

71 commits (14 fixes) modifying Sources/Flux2CLI

A change to existing behaviour in Sources/Flux2CLI — 71 commits (14 fixs), 20 files: “fix(cli): Use model.defaultSteps instead of hardcoded fallback”; “feat: Add gradient\_checkpointing config param (not yet implemented)”; “feat: Quality fixes, performance profiling, and prompt upsampling (\#6)”; and 68 more (refs \#6, \#74, \#75, \#77, \#53, …).

Sources/Flux2CLI · medium confidence · unverified

8 commits (2 fixes) modifying Sources/Flux2Core/Training/Data

A change to existing behaviour in Sources/Flux2Core/Training/Data — 8 commits (2 fixs), 4 files: “refactor: Simplify LoRA training system and add learning curve visualization”; “fix: Resolution-grouped latent caching and non-gated VAE download”; “feat: LoRA training support for Flux.2 models”; and 5 more (refs \#106).

Sources/Flux2Core/Training/Data · medium confidence · unverified

9 commits (3 fixes) modifying Sources/Flux2Core/LoRA

A change to existing behaviour in Sources/Flux2Core/LoRA — 9 commits (3 fixs), 3 files: “refactor: Simplify LoRA training system and add learning curve visualization”; “feat: Add LoRA adapter support for Klein 4B”; “feat: MLX memory optimizations from LTX-2-MLX research (\#29)”; and 6 more (refs \#29).

Sources/Flux2Core/LoRA · medium confidence · unverified

Fixes

10 commits (7 fixes) fixing Sources/FluxTextEncoders/Loading

A fix in Sources/FluxTextEncoders/Loading — 10 commits (7 fixs), 3 files: “feat: Merge FluxTextEncoders library into monorepo”; “feat: Add custom modelsDirectory for sandboxed macOS apps (\#54)”; “fix(loading): third review pass — relocation guards everywhere, no resident garbage”; and 7 more (refs \#54, \#104, \#66).

Sources/FluxTextEncoders/Loading · medium confidence · unverified

6 commits (4 fixes) fixing Sources/FluxTextEncoders/Model/Qwen35

A fix in Sources/FluxTextEncoders/Model/Qwen35 — 6 commits (4 fixs), 8 files: “fix: Robust VLM JSON extraction when score is inside thinking block (\#68)”; “fix: Reduce log verbosity — remove debug prints, use FluxDebug (\#73)”; “fix(loading): sixth review pass — load only the verified series, protect in-place relocation”; and 3 more (refs \#68, \#73, \#66, \#69, \#70).

Sources/FluxTextEncoders/Model/Qwen35 · medium confidence · unverified

7 commits (4 fixes) fixing Sources/Flux2Core/Training/Control

A fix in Sources/Flux2Core/Training/Control — 7 commits (4 fixs), 3 files: “fix: Fix VLM scoring during training — isVLMGenerated filter, final step eval (\#74)”; “feat: Auto-cleanup control files when starting training”; “feat: Add training session control with pause/resume/checkpoint”; and 4 more (refs \#74).

Sources/Flux2Core/Training/Control · medium confidence · unverified

Test coverage

43 commits adding/updating tests in Tests/Flux2CoreTests; 7 commits adding/updating tests in Tests/Flux2ChainsTests; 9 commits adding/updating tests in Tests/FluxTextEncodersTests.

Dependencies

17 commits updating dependencies (1 manifest)

A dependency / build maintenance change in (dependencies) — 17 commits, 1 file: “deps: Update mlx-swift to 0.30.2 and swift-transformers to 1.1.6”; “test: Comprehensive profiling test coverage (\#78)”; “feat: Add vision-based prompt upsampling for multi-image I2I”; and 14 more (refs \#78, \#77, \#93, \#112, \#122, …).

(dependencies) · medium confidence · unverified

Written by cca.canine.dev from the codebase's own history, inside the signed delivery this page is composed from.

How this codebase got here

Baseline

  • First survey — no prior run to compare against. CAI 58.5.

Lenses

  • Code Health 79.0
  • Architecture 92.8
  • Maturity 87.0
  • Readiness 27.0
  • Security 100.0
  • Performance 100.0

Changes since last survey

  • 250 commits — 180 feature/other, 70 fixes

By area

  • Sources/Flux2Core — 90 commits
  • docs/examples — 45 commits
  • (repo) — 34 commits
  • Sources/Flux2CLI — 22 commits
  • (root) — 19 commits
  • Sources/FluxTextEncoders — 8 commits
  • Tests/Flux2CoreTests — 7 commits
  • Sources/Flux2App — 5 commits
  • examples/cat-toy — 5 commits
  • Sources/Flux2Chains — 4 commits
  • examples/tarot-style — 4 commits
  • Tests/Flux2ChainsTests — 2 commits
  • .serena/memories — 1 commit
  • docs/CLI.md — 1 commit
  • docs/INPAINTING_GUIDE.md — 1 commit
  • docs/knowledge — 1 commit
  • docs/screenshots — 1 commit

Notable commits

  • fix: Merge branch 'fix/vae-download-klein4b'
  • fix: Merge fix/integration-issue-39: Expose Base models + fix gated detection
  • fix: Merge hotfix/cancel-generation-crash: fix crash on generation cancellation
  • fix: Merge pull request #123 from VincentGourbin/fix/model-downloader-relocated-model-storage
  • fix: Merge pull request #25 from VincentGourbin/fix/memory-cache-cleanup
  • fix: Merge pull request #44 from VincentGourbin/fix/stale-gated-tests
  • fix: Merge pull request #49 from VincentGourbin/fix/cli-default-steps
  • fix: Merge pull request #58 from VincentGourbin/fix/i2i-spatial-shift-and-strength
  • fix: Merge pull request #80 from VincentGourbin/fix/lora-evaluator-vae-variant
  • fix: Merge pull request #81 from VincentGourbin/fix/vae-variant-remaining-callsites
  • fix: Merge pull request #83 from VincentGourbin/fix/klein-base-transformer-subfolder
  • fix: fix(api): Keep applyTransformerWeights public, make mapTransformerWeights internal
  • fix: fix(app): constrain I2I layout overflow (#94)
  • fix: fix(chains): third review pass on Image Preparation — real crop scale, process-area validation
  • fix: fix(cli): Use model.defaultSteps instead of hardcoded fallback
  • fix: fix(loading): fourth review pass — one shared weight-directory inspector
  • fix: fix(loading): harden path overrides after review — no regression for symlinked layouts
  • fix: fix(loading): relocated model reports its real size, add path override hook
  • fix: fix(loading): seventh review pass — deterministic series pick, MistralVLM missed
  • fix: fix(loading): sixth review pass — load only the verified series, protect in-place relocation
  • …and 230 more

Written by cca.canine.dev from the codebase's own history, inside the signed delivery this page is composed from.

Survey your own repository

VincentGourbin/flux-2-swift-mlx was measured the same way every project in this corpus was: the same rubric, at a pinned commit, with the result published in full. Point a surveyor at a repository you know and see whether you agree with it.

About this page

  • The score is its most recent published measurement, taken on 16 November 2026 at a pinned commit. It is not a live figure and does not change until the project is measured again.
  • Measured at commit 315c27e83187909bc5e1f880b4850918342d891b — the exact code this score is about.
  • Scored under rubric-2026.11.4 — the same rubric and the same method as every other entry in this index.
  • Measured by Canine Code Analyzer (cca.canine.dev) using cca measure-2026.11.83+preprod-a489942cf94b.