mflux-community/mflux v.0.19.1

Release 0.19.1

on GitHub

🔒 Security

• Locked dependencies updated beyond all reported vulnerable ranges: raises the minimum supported versions of Pillow, Requests, PyTorch, Transformers, and urllib3, refreshes affected transitive dependencies, and removes the now-unused cryptography dependency from the lock. This addresses 35 Dependabot findings: 19 high, 12 medium, and 4 low. (#655)

🐛 Bug Fixes

• --instruction exclusivity is declared, and visible to mflux-capabilities: in-context-edit's prompt/instruction mutual exclusion was hand-checked after parsing, so the capabilities dump published --instruction as an independent flag and --prompt-file --instruction slipped through entirely. --instruction now joins the --prompt/--prompt-file argparse group; the "at least one" requirement stays post-parse. Also fixes --width's help interpolating the HEIGHT default, and documents why fibo-edit's matte writes no sidecar. (#578, #669)
• Baking a LoRA into a sub-8-bit quantized model silently dropped the adapter: the fold requantized the merged weight at the base layer's own precision, and below 8 bits the quantization step is coarser than a typical LoRA delta (on Krea 2 at q4 the group step measures ~12x the delta rms), so -q 4 --lora X with baking on, which is the default, generated the base image while the logs reported success. Sub-8-bit layers now requantize at q8 when a delta is folded in, the same escape the fp8 path takes, with a console line reporting how many layers moved; the per-layer loader already reconstructs such mixed saves from stored shapes. Applies to runtime baking and to mflux-save of a quantized model with --lora. (#665, #668)
• Foreign --model values are rejected on the boogu, z-image, flux2 and ideogram4 CLIs: all of them silently ignored a builtin name from another family and ran their own model instead (mflux-generate-boogu --model dev ran Boogu; mflux-generate-ideogram4 --model qwen-image ran Ideogram). They now resolve through the same restriction as the krea2/lens/ernie/z-image-turbo CLIs: aliases of the command's own model (or, for flux2, any registry flux2- entry, and for z-image, its distilled sibling) are accepted, anything else errors during CLI startup before any weights move, and paths or repo ids keep loading through model_path as before. flux2-edit additionally enforces the same distilled-checkpoint --guidance 1.0 rule as flux2-generate. (#578, #650)
• Shell completions cover every installed command: the generator's hand-maintained list had drifted six commands behind pyproject (mflux-generate-boogu, mflux-generate-lens, both mflux-generate-ernie-image commands, mflux-generate-ideogram4, mflux-capabilities never completed). Commands are now discovered from the installed console scripts and, where a CLI exposes build_parser(), completions are generated from the CLI's real parser instead of a hand-copied recipe of it. (#578, #651)
• mflux-capabilities publishes wire types, not Python converter names: options validated by a named converter leaked the function name as their type (--vae-tile-size claimed type vae_tile_size, --mlx-cache-limit-gb claimed positive_float, every Path option claimed PosixPath). Named converters now map to what they yield: int, float, path, and int-or-scale for values that accept a pixel count, a 2x factor or auto. The dump's type field is pinned to that closed vocabulary by a test. (#578, #652)
• The metadata sidecar of an in-context result can reproduce the run: get_right_half() dropped negative_prompt, image_paths and the redux fields when cloning metadata onto the cropped half, and the --save-full-image composite was saved without a sidecar even under --metadata. redux_image_paths was also serialized as a Python repr ("['a.png', 'b.png']") instead of a list, so --config-from-metadata could never read it back. The unreachable restore branches for controlnet_save_canny and image_outpaint_padding, keys no released writer has ever emitted, are removed along with a stray debug print() in the outpaint parse block. (#578, #649)
• mflux-generate-redux --config-from-metadata can replay a sidecar: #649 made redux_image_paths a JSON list, but the redux CLI still declared the flag argparse-required and supports_metadata_config=False, so a sidecar-only rerun exited 2 before restore ran. -C now restores the paths and strengths, --redux-image-paths is required after restore rather than during parsing, and a missing sidecar path errors at parse time naming the sidecar. Old repr-string sidecars are not treated as path lists. (#578, #663)

📝 Documentation

• Lens and Boogu Image documented: both models join the README's supported-models table, each with a per-model README covering usage and options. (#659)

🧰 DX & Maintenance

• Machine-readable model registry extract: a new scripts/ci_extract_models.py (with a ci-extract workflow and just recipe) dumps the supported-model registry as JSON, needed by CI to auto-build MFlux models for Hugging Face. (#658)
• PyPI publish gated behind a protected pypi environment: the release workflow now runs under the pypi deployment environment, so its reviewer-approval and branch-protection rules apply before anything reaches PyPI, and the trusted publisher is pinned to that environment name. A fast-fail step also rejects dispatches from any ref other than main. (#646)
• Lens DiT loads through the shared WeightDefinition seam: replaces its bespoke weight-loading path with the mechanism the other models use. (#654)
• Flaky Gemma 2 causality test fixed: the test could pick an out-of-vocab token and fail spuriously. (#653)