mflux-manual-testing

Some regressions (especially in CLIs and image IO) are easiest to catch by running the commands and visually inspecting outputs. This skill provides a lightweight, change-driven manual test checklist.

When to Use

• You changed any CLI entrypoint(s) under src/mflux/models/**/cli/.
• You touched callbacks (e.g. stepwise output, memory saver) or metadata/image saving.
• Tests are green but you want confidence in real command usage.

Strategy (change-driven)

• Identify what changed on your branch (new flags, default behavior changes, new callbacks, new models).
• Only run manual checks for the touched areas; don’t try to exercise every CLI.
• Prefer 1–2 seeds and a small step count (e.g. 4) for fast iteration, unless the change affects convergence/quality.
• Before manual CLI testing, reinstall the local tool executables so you’re testing the latest code:

uv tool install --force --editable --reinstall .

Core CLI checks (pick what’s relevant)

• Basic generation: run the CLI once with a representative prompt and confirm the output is not “all noise”.
• Model saving (if relevant): if you touched weight loading/saving or model definitions, run mflux-save for the affected model(s) and verify:
  • the output directory is created
  • the command completes without missing-file errors
• Run from disk (if relevant): if you touched save/load paths or model resolution, generate from a locally saved model directory by passing --model /full/path/to/saved-model and confirm it runs and produces a sane image.
• Stepwise outputs (if relevant): run with --stepwise-image-output-dir and confirm:
  • step images are written for each step
  • the final step image matches the final output image qualitatively
  • the composite image is created
• Low-RAM path (if relevant): run with --low-ram and confirm:
  • generation completes
  • output quality is sane (no unexpected all-noise output)
• Metadata (if relevant): run with --metadata and confirm the .metadata.json sidecar is emitted and looks consistent.

Output review (human-in-the-loop)

• Always point the human reviewer at:
  • the final output image path
  • any stepwise directory / composites
  • any metadata JSON files
• Ask the human to visually confirm “looks correct” rather than attempting pixel-perfect parity manually.

diffusers reference comparison (new model ports)

mflux does not install diffusers; use a sibling clone (commonly ../diffusers on Desktop).

When: validating a new MLX port before merge, or when golden tests / visuals look wrong.

Match settings: same prompt, width, height, seed, steps, guidance. For fair speed comparisons, match precision (mflux bf16 vs diffusers bf16, not -q 8 vs bf16).

diffusers setup tips (read the reference pipeline first):

• Compare model_index.json / from_pretrained kwargs with mflux get_download_patterns() — disable or pass None for components mflux does not load
• local_files_only=True / HF_HUB_OFFLINE=1 to use cache and surface missing files early
• Distilled vs base variants are often the same pipeline class with different checkpoint + steps/guidance

Run both:

# mflux
uv run mflux-generate-<model> --prompt "..." --width 640 --height 368 --seed 7 --steps 8 --guidance 1.0 --output /tmp/mflux.png

# diffusers (from diffusers repo)
cd ../diffusers && HF_HUB_OFFLINE=1 uv run python -c "..."  # inline script; see mflux-debugging

If visuals differ: do not assume the port is broken. Run the latent injection workflow in mflux-debugging to separate (a) transformer/VAE quality from (b) RNG/scheduler differences.

Save comparison PNGs to explicit paths; report /usr/bin/time -p totals. Do not commit comparison artifacts.

Notes

• If the installed uv tool executable behaves differently from uv run python -m ..., prefer the local module run to isolate environment/tooling issues.
• If you need to reinstall the local tool executables, see the repo rules for the current recommended command.