Release 0.18.1

🎨 New Model Support

• Krea 2: Add text-to-image support for krea/Krea-2-Turbo — a single-stream MMDiT built on the Qwen-Image stack (Qwen-Image VAE + a 12-layer Qwen3-VL-4B text-encoder tap). Includes the mflux-generate-krea2 CLI (live progress, --metadata, stepwise output), er_sde and Euler samplers, and mflux-save quantization caching.

✨ Improvements

• Atomic --lora and --image flags: Pair each path with its value on a single, repeatable flag — --lora A.safetensors 0.7 --lora B.safetensors (scale defaults to 1.0) and --image photo.jpg 0.6 (strength defaults to the model default). This removes the positional-alignment footgun of the parallel --lora-paths/--lora-scales and --image-path/--image-strength lists, which remain fully supported and are marked deprecated in --help. (#438)

🐛 Bug Fixes

• Ideogram 4 stepwise output: Fix --stepwise-image-output-dir for Ideogram 4 by routing unpacked latents to vae.decode instead of decode_packed_latents when channels are already VAE-ready. (#444)

📝 Documentation

• Related projects: Add mflux-paint to the Related projects list. (#471)

🧰 DX & Maintenance

• PyPI trusted publishing: Publish releases via GitHub Actions OIDC (pypa/gh-action-pypi-publish) instead of a long-lived PYPI_API_TOKEN secret.

👩‍💻 Contributors

• @Amo643
• @anthonywu
• @filipstrand
• @plz12345

---

Release 0.18.0

🎨 New Model Support

• ERNIE-Image & ERNIE-Image-Turbo: Port Baidu's ERNIE-Image models with text-to-image CLI entrypoints, LoRA inference and training, and mflux-save support.
• Ideogram 4 FP8: Add Ideogram 4 FP8 text-to-image support with JSON caption handling, FP8 safetensors loading, and a dedicated CLI entrypoint.

✨ Improvements

• FLUX.2 Klein 9B KV-cache: Add KV-cache support for flux2-klein-9b-kv with roughly 2.4× speedup on multi-reference edit workloads.

🐛 Bug Fixes

• FLUX.2 Klein Edit guidance: Allow --guidance > 1.0 for FLUX.2 Klein edits by checking the resolved FLUX.2 model config instead of requiring a base model name; defaults remain unchanged.
• FLUX.2 Klein mflux-generate: Fix FileNotFoundError: text_encoder_2 by routing Klein models through Flux2Klein and skipping the unused T5/text_encoder_2 weight path.
• Memory management: Evict the text encoder after encoding and clear the MLX cache between seeds on multi-seed runs to prevent OOM on large models such as FLUX.2 Klein 9B.

📝 Documentation

• Related projects: Add mlx-taef and mlx-teacache to the Related projects list.

👩‍💻 Contributors

• @azrahello
• @c2p-cmd
• @IonDen
• @lpalbou
• @michaeltrefry
• @omercelik
• @plz12345

---

Release 0.17.5

🐛 Bug Fixes

• Qwen Image Edit mflux-save: Route Qwen edit model names to QwenImageEdit and save through the same path as inference so VisionTransformer (encoder.visual) weights are written. Saving with QwenImage previously omitted those weights and led to random vision encoders after reload.
• Battery saver callback: Harden Apple Silicon battery detection when system_profiler is missing and resolve the helper script via absolute paths.

📝 Documentation

• Related projects: Clarify that MindCraft Studio is a macOS app built on mflux.

🧰 DX & Maintenance

• Dependencies: Relax the protobuf upper bound to allow current 7.x releases while keeping a safe ceiling below 8.0.

👩‍💻 Contributors

• @anthonywu
• @f-gibellini
• @JiwaniZakir

---

Release 0.17.4

🐛 Bug Fixes

• Z-Image PEFT/ModelScope LoRA keys: Extend the Z-Image LoRA mapping with .default tensor name variants so adapters in PEFT/ModelScope layouts (for example Tongyi-MAI exports) resolve and apply correctly instead of matching zero weights.

👩‍💻 Contributors

• @filipstrand

---

Release 0.17.3

🐛 Bug Fixes

• FLUX.2 edit guidance metadata: Preserve the requested guidance value for FLUX.2 Klein base image-edit runs so mflux-info and saved metadata report the actual guidance used instead of always showing 1.0.

👩‍💻 Contributors

• @filipstrand

---

Release 0.17.2

🐛 Bug Fixes

• Shared tokenizer cache resolution: Fix Hugging Face tokenizer resolution when a repo is only partially cached locally, preserving offline-first behavior for valid cached layouts while retrying ambiguous cached primaries once before surfacing real load errors.

🧰 DX & Maintenance

• Tokenizer resolution coverage: Expand shared tokenizer-resolution regression tests to cover root-layout tokenizers, fallback edge cases, and refresh failure handling.

👩‍💻 Contributors

• @filipstrand

---

Release 0.17.1

🐛 Bug Fixes

• Hugging Face tokenizer dependencies: Declare protobuf so minimal installs (including uv tool install mflux) include packages Transformers may require when loading tokenizers, fixing failures such as mflux-generate-fibo when the tokenizer falls back off the fast path.

👩‍💻 Contributors

• @filipstrand

---

Release 0.17.0

🎨 New Model Support

• FIBO Edit: Add image-editing support for the FIBO model family.
• FIBO Edit remove-background workflow: Support the dedicated remove-background edit path for FIBO.

✨ Improvements

• Training image scaling: Scale training images by area rather than longest side for more consistent preprocessing.
• MLX 0.31.x: Allow MLX 0.31.x in dependency ranges.
• FLUX.2 LoRA mapping: Expand LoRA key mapping coverage for FLUX.2.

🐛 Bug Fixes

• Training optimizer state: Evaluate optimizer state after each training step as intended.
• Local tokenizer loading: Fix loading tokenizers from local paths.
• Dynamic-resolution image edit: Restore correct behavior for image edit when using dynamic resolution.

👩‍💻 Contributors

• @filipstrand
• @icelaglace
• @TheOrsa
• @waldheinz

---

Release 0.16.9

✨ Improvements

• Broader LoRA compatibility for FLUX.2 and Z-Image: Expand LoRA mapping coverage so more adapter key layouts resolve cleanly for FLUX.2 and Z-Image models.

👩‍💻 Contributors

• @filipstrand

---

Release 0.16.8

✨ Improvements

• Local-model LoRA training: Allow LoRA training to work when the base model is supplied from a local path, including the FLUX.2 and Z-Image training adapters.

📝 Documentation

• Distilled-model step defaults: Clarify CLI guidance so examples prefer model default inference steps unless the user intentionally overrides them.

👩‍💻 Contributors

• @waldheinz

---