Mflux Cli

• Inference steps: When constructing a CLI call for any model, always check MODEL_INFERENCE_STEPS and use the model's default/recommended step count unless the user explicitly asks for a different value. This is especially important for distilled/non-base variants (for example fibo-lite, z-image-turbo, and other *-turbo/*-lite models), where using full/base-model step counts is usually counterproductive.
• Resource/inspection flags: Mention --low-ram to reduce memory usage and --stepwise-image-output-dir for stepwise outputs when useful.
• Python API requests: If a user asks for the Python API, treat the equivalent CLI script as the best starting reference for the underlying parameters and defaults.
• CLI implementation changes: When adding or changing CLI behavior, prefer extending shared parser methods and shared helpers before adding manual one-off arguments or save paths in a model CLI.
• Docs/examples drift: If CLI defaults, supported flags, or recommended usage changed, update the matching README examples in the same pass.

Capability inventory (current)

Core generation

• Text-to-image across Flux, Flux2, Qwen, Z-Image Turbo, FIBO.
• Image-to-image where supported (Flux, Qwen, Z-Image Turbo, FIBO).

Editing and conditioning

• Kontext image conditioning.
• In-context editing and reference-image workflows.
• CATVTON (virtual try-on).
• Redux multi-image conditioning.
• ControlNet (Canny).
• Depth conditioning.
• Fill / inpainting.
• Flux2 Edit and Qwen Edit (multi-image edit).

Upscaling

• SeedVR2 diffusion upscaler (preferred).
• Flux ControlNet upscaler (legacy).

Model management

• Quantized inference and saving quantized models.
• Local model path loading (with base-model hints when needed).

LoRA

• Load LoRAs, multi-LoRA, scale control.
• In-context style LoRA shortcuts.
• LoRA library lookup tool.

Metadata and reproducibility

• Export JSON metadata per image.
• Reuse prior parameters from metadata config files.
• Inspect metadata from existing images.

Prompt tooling

• Prompt files.
• Negative prompts where supported (not supported for Flux2).

Training

• LoRA finetuning.

Utilities

• DepthPro depth-map extraction.
• FIBO VLM prompt inspire/refine tools.