mflux-cv 0.18.25__tar.gz

mflux_cv-0.18.25/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2026 Filip Strand
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

mflux_cv-0.18.25/PKG-INFO

@@ -0,0 +1,475 @@
+Metadata-Version: 2.3
+Name: mflux-cv
+Version: 0.18.25
+Summary: MLX native implementations of state-of-the-art generative image models.
+Keywords: flux,flux2,krea,krea-2,ideogram,ideogram-4,z-image,qwen,qwen-image,ernie,fibo,seedvr2,mlx,apple-silicon,metal,macos,mac,macbook,macbook-pro,on-device,local,m1,m2,m3,m4,m5,pro,max,image-generation,text-to-image,t2i,image-to-image,img2img,i2i,diffusion,stable-diffusion,generative-ai,inpainting,controlnet,lora,lokr,dora,fine-tuning,training,quantization,ai,ml,transformers,diffusers,huggingface
+Author: Filip Strand
+Author-email: Filip Strand <strand.filip@gmail.com>
+License: MIT License
+         
+         Copyright (c) 2026 Filip Strand
+         
+         Permission is hereby granted, free of charge, to any person obtaining a copy
+         of this software and associated documentation files (the "Software"), to deal
+         in the Software without restriction, including without limitation the rights
+         to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+         copies of the Software, and to permit persons to whom the Software is
+         furnished to do so, subject to the following conditions:
+         
+         The above copyright notice and this permission notice shall be included in all
+         copies or substantial portions of the Software.
+         
+         THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+         IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+         FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+         AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+         LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+         OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+         SOFTWARE.
+Classifier: Intended Audience :: Developers
+Classifier: Operating System :: MacOS
+Classifier: Operating System :: POSIX :: Linux
+Classifier: Programming Language :: Python :: 3
+Classifier: Programming Language :: Python :: 3.10
+Classifier: Programming Language :: Python :: 3.11
+Classifier: Programming Language :: Python :: 3.12
+Classifier: Programming Language :: Python :: 3.13
+Classifier: Programming Language :: Python :: 3.14
+Requires-Dist: filelock>=3.20.1
+Requires-Dist: fonttools>=4.60.2
+Requires-Dist: huggingface-hub>=1.1.6,<2.0
+Requires-Dist: hf-transfer>=0.1.9,<1.0
+Requires-Dist: matplotlib>=3.9.2,<4.0
+Requires-Dist: mlx>=0.27.0,<0.32.0 ; sys_platform == 'darwin'
+Requires-Dist: mlx[cuda13]>=0.30.3,<0.32.0 ; sys_platform == 'linux'
+Requires-Dist: numpy>=2.0.1,<3.0
+Requires-Dist: opencv-python>=4.10.0,<5.0
+Requires-Dist: piexif>=1.1.3,<2.0
+Requires-Dist: pillow>=12.1.1
+Requires-Dist: platformdirs>=4.0,<5.0
+Requires-Dist: regex>=2024.11.6
+Requires-Dist: requests>=2.32.4
+Requires-Dist: safetensors>=0.4.4,<1.0
+Requires-Dist: sentencepiece>=0.2.1,<1.0
+Requires-Dist: tokenizers>=0.20.3 ; python_full_version >= '3.13'
+Requires-Dist: toml>=0.10.2,<1.0
+Requires-Dist: torch>=2.7.1,<3.0
+Requires-Dist: torch>=2.8.0,<3.0 ; python_full_version >= '3.13'
+Requires-Dist: tqdm>=4.66.5,<5.0
+Requires-Dist: transformers>=5.0.0,<6.0
+Requires-Dist: twine>=6.1.0,<7.0
+Requires-Dist: urllib3>=2.6.0
+Requires-Dist: protobuf>=4.25.0,<8.0
+Requires-Dist: matplotlib>3.10,<4.0 ; extra == 'dev'
+Requires-Dist: pytest>=8.3.0,<9.0 ; extra == 'dev'
+Requires-Dist: pytest-timer>=1.0,<2.0 ; extra == 'dev'
+Requires-Dist: mlx==0.31.0 ; sys_platform == 'darwin' and extra == 'dev'
+Requires-Dist: mlx[cuda13]==0.30.3 ; sys_platform == 'linux' and extra == 'dev'
+Maintainer: Filip Strand
+Maintainer-email: Filip Strand <strand.filip@gmail.com>
+Requires-Python: >=3.10
+Project-URL: homepage, https://github.com/filipstrand/mflux
+Provides-Extra: dev
+Description-Content-Type: text/markdown
+
+# mflux-CV
+
+A drop-in community build of [mflux](https://github.com/filipstrand/mflux) by Filip Strand (MIT). It
+stays current with upstream `main` and adds a curated set of fixes, training features, and useful
+community PRs, with prebuilt wheels on GitHub Releases so improvements are easy to pick up between
+upstream releases. The Python package is still `mflux` and every CLI keeps its name,
+so this is a direct replacement in ComfyUI, image-studio, or any existing setup.
+
+**Install**
+
+```bash
+pip install mflux-cv
+```
+
+To track a specific tag instead, or to pick up work that has not been released yet:
+
+```bash
+pip install git+https://github.com/HowDidTheCatGetSoFat/mflux-cv.git@v.0.18.25-CV
+```
+
+Wheels are also attached to every [Release](https://github.com/HowDidTheCatGetSoFat/mflux-cv/releases).
+
+> **Install this or upstream `mflux`, never both.** The distribution is named `mflux-cv` so it can
+> live on PyPI alongside upstream, though both provide the same `mflux` import package. With both
+> installed, pip raises no error while the two overwrite each other's files, and uninstalling either
+> one leaves the other broken. Run `pip uninstall mflux` before installing this, and
+> `pip uninstall mflux-cv` before going back.
+
+**Relationship to upstream:** kept rebased on `filipstrand/mflux` so changes merge back cleanly. Every
+upstream commit is included and nothing is removed. Credit for the base project and every pulled PR
+goes to their authors.
+
+## Changelog (on top of upstream 0.18.0)
+
+### 0.18.25-CV
+
+- **Multi-ControlNet for FLUX.1.** Several controlnets can now be stacked, each with its own
+  checkpoint, control image and strength (for example depth + canny). Repeat `--controlnet-path`
+  and `--controlnet-image-path` (and optionally `--controlnet-strength`, or give one value for all):
+
+  ```bash
+  mflux-generate-controlnet -m dev --prompt "a modern living room" \
+      --controlnet-path org/depth --controlnet-image-path depth.png --controlnet-strength 0.7 \
+      --controlnet-path org/canny --controlnet-image-path canny.png --controlnet-strength 0.4
+  ```
+
+  The residual path is additive, so each net's residuals are spread over the transformer's blocks
+  with the rule the transformer already applies and then summed. Controlnets with different block
+  counts stack correctly, and a single controlnet renders exactly as before (the expansion
+  reproduces the transformer's own selection index for index, which is pinned by a test).
+- `--controlnet-path` is new: it selects a controlnet checkpoint (local path or HF repo) instead of
+  the one named by the model config, which also makes `Flux1Controlnet(controlnet_path=...)` work.
+  It was previously accepted by the constructor and silently ignored.
+- A single `--controlnet-image-path` / `--controlnet-strength` keeps its scalar shape, so existing
+  commands and the metadata round-trip are unchanged.
+- Canny preprocessing is decided per controlnet, from that net's own checkpoint name (the same match
+  `is_canny()` makes). A depth + canny stack therefore preprocesses only the canny image, and a
+  config-driven canny run behaves exactly as before.
+- Known limitation: the image metadata holds a single controlnet, so a stacked run records only its
+  first net there. Metadata-driven re-runs of a stack are not supported; pass the flags explicitly.
+  Single-controlnet metadata is unchanged.
+
+### 0.18.24-CV
+
+- FLUX.2-klein now exposes `flux2-klein-edit` / `flux2-edit` / `klein-edit` aliases (it does txt2img
+  and edit from the same weights), so the edit variant is selectable by name — used by the
+  ComfyUI-mflux-AnyModel node.
+- Qwen-Image-Edit: skip the unconditional pass at guidance 1.0 (it reduces to the conditional noise
+  there), halving per-step compute for CFG-distilled setups like the Lightning step-reduction LoRAs.
+
+### 0.18.23-CV
+
+- Updated the default Qwen models to the latest releases (based on #475's sibling, #474, credited
+  under [Community PRs pulled in](#community-prs-pulled-in)): `qwen-image` now loads Qwen-Image-2512 and
+  `qwen-image-edit` loads Qwen-Image-Edit-2511. The architecture is identical to the prior releases, so
+  they are drop-in; both were validated end-to-end on Apple Silicon. Unlike the upstream PR, the old
+  `qwen-edit-2509` alias still resolves to the actual 2509 weights (kept as its own entry) instead of
+  silently pointing at the new default.
+
+### 0.18.22-CV
+
+- Pulled in configurable VAE decode tiling (#475, credited under [Community PRs pulled in](#community-prs-pulled-in)):
+  `--vae-tiling` / `--vae-tile-size` decouple tiled decoding from full low-RAM mode, so you can cut peak
+  memory on large generations without the rest of the low-RAM penalty.
+
+### 0.18.21-CV
+
+- Krea 2 depth ControlNet: fix the estimated-depth convention (Depth Pro already outputs near = white,
+  like Depth-Anything-V2, so its map is now used un-inverted — an earlier inversion put the background in
+  front). `--save-depth-map` now writes the estimated map (`<output>_depth_map.png`).
+
+### 0.18.20-CV
+
+- Krea 2 depth ControlNet now supports quantization (`-q 8` / `-q 4`): the control deltas are baked
+  into the base weights before quantization, so the packed model keeps the depth control. Validated
+  end-to-end at int8.
+
+### 0.18.19-CV
+- New: **Krea 2 depth ControlNet** (`mflux-generate-krea2-controlnet`). Runs the community
+  [Krea-2-controlnet](https://github.com/Tanmaypatil123/Krea-2-controlnet) depth checkpoint natively in
+  MLX: the input projection is widened to take a depth latent concatenated on the channel axis, and the
+  attention/MLP deltas are merged into the base weights. Depth is taken from a supplied map
+  (`--depth-image-path`) or estimated with the native Depth Pro. See the Krea 2 depth ControlNet section
+  below.
+
+### 0.18.18-CV
+- Pulled in two new upstream models (credited under [Community PRs pulled in](#community-prs-pulled-in)):
+  Boogu-Image-0.1-Turbo (#446, `mflux-generate-boogu`) and Qwen-Image-Layered (#302,
+  `mflux-generate-qwen-layered`) for decomposing an image into RGBA layers.
+
+### 0.18.17-CV
+- Pulled in Ideogram 4 mlx-forge checkpoint loading (#445) and mixed-quant FLUX.2 inference (#436),
+  both credited under [Community PRs pulled in](#community-prs-pulled-in).
+
+### 0.18.16-CV
+- Pulled in two upstream bug fixes (credited under [Community PRs pulled in](#community-prs-pulled-in)):
+  fused-qkv LoRA loading (#459) and ERNIE / Krea 2 tiled img2img (#463).
+- Repo automation: CodeQL security scanning, structured issue forms, and PR / issue auto-labeling.
+
+### 0.18.15-CV
+First release under the `mflux-CV` name, in the community home at
+[HowDidTheCatGetSoFat/mflux-cv](https://github.com/HowDidTheCatGetSoFat/mflux-cv). Same codebase as the
+prior `+fxd0h` builds (0.18.1 through 0.18.5); this is the rebrand plus everything listed below.
+
+### 0.18.5
+- **Krea 2 `--krea2-uncensor <k>`**: scales the text-fusion projector's refusal layers (tapped Qwen3-VL
+  9/10/11) so explicit prompts render instead of being dodged. `k=1` is off, `~6` neutralises the filter.
+
+### 0.18.4
+- **LoRA on quantized bases**: keep the adapter live at inference instead of baking it into the quantized
+  weights (baking re-quantized and badly diverged the output on the `--quantize` default).
+- **LoKr**: load LyCORIS LoKr adapters for Krea 2, Qwen, and Ideogram 4.
+- **Ideogram 4**: fix the stepwise-preview VAE decode crash on already-unpacked latents; guard the
+  injected-LoRA scan against transformers without `named_modules`.
+- **Krea 2 Raw**: download the diffusers transformer from HuggingFace.
+- **z-image**: shared `--saveinfo` filename builder; fix numeric-tag collisions.
+- **qwen-edit**: clearer error on empty `image_paths`; regenerated golden references.
+
+### 0.18.3
+- Review fixes: fp8-aware fused DoRA, training guards (LR, grad-accum reset on skip, qwen VAE flag),
+  route the CFG negative through an injected LoRA in training previews, surface LoRA bake failures on
+  save, EMA resume from live weights.
+
+### 0.18.2
+- **Krea 2 sigma schedule**: use the official dynamic exponential shift instead of a linear 1.15.
+
+### 0.18.1
+- **Training suite**: DoRA (weight-decomposed LoRA) for Krea 2, Ideogram 4, z-image, flux, flux2;
+  gradient accumulation; EMA of trained weights; caption dropout; masked loss; regularization /
+  prior-preservation images; continue training from an existing LoRA; non-finite-step guard; utf-8-safe
+  captions; free training-loss plot.
+- **Krea 2**: LoRA training, Raw variant, and diffusers-format loading.
+
+### Community PRs pulled in
+- **[filipstrand/mflux#459](https://github.com/filipstrand/mflux/pull/459) by Sahil Tanveer** — fix LoRA
+  loading for fused qkv layers: keep the shared rank/down projection whole and slice only the up
+  projection, so kohya/BFL FLUX LoRAs with a rank divisible by 3/4 load correctly.
+- **[filipstrand/mflux#463](https://github.com/filipstrand/mflux/pull/463) by Mike Wallio** — fix ERNIE
+  and Krea 2 img2img with tiled VAE latents: the 5D tiled-VAE pack path took the wrong slice; keep the
+  singleton temporal axis so tiled-decode img2img reconstructs correctly.
+- **[filipstrand/mflux#445](https://github.com/filipstrand/mflux/pull/445) by plz12345** — load
+  Ideogram 4 from mlx-forge converted checkpoints (bf16 / int8) by HF repo id, skipping the fp8-only
+  validation; plus a once-built boolean attention keep-mask. Merged with our gradient-checkpointing.
+- **[filipstrand/mflux#436](https://github.com/filipstrand/mflux/pull/436) by Ian Scrivener** — mixed-quant
+  inference for FLUX.2: quantize the transformer and text encoder to different levels, or load each from
+  its own path (`--model-transformer` / `--model-text-encoder`); the VAE stays bf16. Merged alongside our
+  LoKr flux2 changes.
+- **[filipstrand/mflux#446](https://github.com/filipstrand/mflux/pull/446) by plz12345** — new model:
+  Boogu-Image-0.1-Turbo (`mflux-generate-boogu`). Applied cleanly.
+- **[filipstrand/mflux#302](https://github.com/filipstrand/mflux/pull/302) by ZimengXiong** — new model:
+  Qwen-Image-Layered (`mflux-generate-qwen-layered`) for decomposing an image into RGBA layers, with a
+  low-memory chunked save path. We kept our README; the PR's stale old-structure README changes were dropped.
+- **[filipstrand/mflux#475](https://github.com/filipstrand/mflux/pull/475) by azrahello** — configurable
+  VAE decode tiling: `--vae-tiling` and `--vae-tile-size` (min 128, multiple of 16) enable tiled decoding
+  on its own, decoupled from `--low-ram`, to lower peak memory during the VAE decode phase. Applied cleanly.
+- **[filipstrand/mflux#474](https://github.com/filipstrand/mflux/pull/474) by imbible** — bump the default
+  Qwen models to Qwen-Image-2512 and Qwen-Image-Edit-2511. Both validated generating end-to-end on Apple
+  Silicon (identical architecture, drop-in). We diverged in one place: kept a separate `qwen-image-edit-2509`
+  entry so `qwen-edit-2509` still resolves to the real 2509 weights instead of the new default.
+
+### Krea 2 depth ControlNet
+
+Steer a Krea 2 generation with the depth of a reference image, running natively in MLX. Uses the
+community [Krea-2-controlnet](https://github.com/Tanmaypatil123/Krea-2-controlnet) depth checkpoint by
+Tanmay Patil (base weights: `krea/Krea-2-Raw` / `krea/Krea-2-Turbo`).
+
+```bash
+mflux-generate-krea2-controlnet \
+  --model krea2 \
+  --controlnet-path /path/to/depth-control-lora.safetensors \
+  --image-path reference.png \
+  --prompt "a glowing crystal orb on a wooden table, studio photo" \
+  --steps 8 --seed 42 --height 1024 --width 1024 \
+  --output out.png
+```
+
+- `--image-path` estimates depth with the native Depth Pro (near = white, used as-is, the same
+  convention the checkpoint was trained on). Add `--save-depth-map` to also write the estimated map. For
+  the closest match to the training data, pass a Depth-Anything-V2 map directly with `--depth-image-path`.
+- `--controlnet-strength` scales how strongly the control deltas are merged (default `1.0`).
+- `--krea2-uncensor` is supported here too.
+- `-q 8` / `-q 4` quantize the variant: the control deltas are baked into the base weights before
+  quantization, so the packed model keeps the depth control.
+
+---
+
+> The rest of this file is the upstream mflux documentation.
+
+![image](src/mflux/assets/logo.jpg)
+
+[![MFLUX](https://img.shields.io/pypi/v/mflux?label=MFLUX&logo=pypi&logoColor=white)](https://pypi.org/project/mflux/)
+[![MLX](https://img.shields.io/pypi/v/mlx?label=MLX&logo=pypi&logoColor=white)](https://pypi.org/project/mlx/)
+[![CI](https://github.com/filipstrand/mflux/actions/workflows/tests.yml/badge.svg)](https://github.com/filipstrand/mflux/actions/workflows/tests.yml)
+
+### About
+
+Run the latest state-of-the-art generative image models locally on your Mac in native MLX!
+
+### Table of contents
+
+- [💡 Philosophy](#-philosophy)
+- [💿 Installation](#-installation)
+- [🎨 Models](#-models)
+- [✨ Features](#-features)
+- [🌱 Related projects](#related-projects)
+- [🙏 Acknowledgements](#-acknowledgements)
+- [⚖️ License](#%EF%B8%8F-license)
+
+---
+
+### 💡 Philosophy
+
+MFLUX is a line-by-line MLX port of several state-of-the-art generative image models from the [Huggingface Diffusers](https://github.com/huggingface/diffusers) and [Huggingface Transformers](https://github.com/huggingface/transformers) libraries. All models are implemented from scratch in MLX, using only tokenizers from the [Huggingface Transformers](https://github.com/huggingface/transformers) library. MFLUX is purposefully kept minimal and explicit, [@karpathy](https://gist.github.com/awni/a67d16d50f0f492d94a10418e0592bde?permalink_comment_id=5153531#gistcomment-5153531) style.
+
+---
+
+### 💿 Installation
+If you haven't already, [install `uv`](https://github.com/astral-sh/uv?tab=readme-ov-file#installation), then run:
+
+```sh
+uv tool install --upgrade mflux
+```
+
+After installation, the following command shows all available MFLUX CLI commands: 
+
+```sh
+uv tool list 
+```
+
+To generate your first image using, for example, the z-image-turbo model, run
+
+```
+mflux-generate-z-image-turbo \
+  --prompt "A puffin standing on a cliff" \
+  --width 1280 \
+  --height 500 \
+  --seed 42 \
+  --steps 9 \
+  -q 8
+```
+
+![Puffin](src/mflux/assets/puffin.png)
+
+The first time you run this, the model will automatically download which can take some time. See the [model section](#-models) for the different options and features, and the [common README](src/mflux/models/common/README.md) for shared CLI patterns and examples.
+
+<details>
+<summary>Python API</summary>
+
+Create a standalone `generate.py` script with inline `uv` dependencies:
+
+```python
+#!/usr/bin/env -S uv run --script
+# /// script
+# requires-python = ">=3.10"
+# dependencies = [
+#   "mflux",
+# ]
+# ///
+from mflux.models.z_image import ZImageTurbo
+
+model = ZImageTurbo(quantize=8)
+image = model.generate_image(
+    prompt="A puffin standing on a cliff",
+    seed=42,
+    num_inference_steps=9,
+    width=1280,
+    height=500,
+)
+image.save("puffin.png")
+```
+
+Run it with:
+
+```sh
+uv run generate.py
+```
+
+For more Python API inspiration, look at the [CLI entry points](src/mflux/models/z_image/cli/z_image_turbo_generate.py) for the respective models.
+</details>
+
+<details>
+<summary>⚠️ Troubleshooting: hf_transfer error</summary>
+
+If you encounter a `ValueError: Fast download using 'hf_transfer' is enabled (HF_HUB_ENABLE_HF_TRANSFER=1) but 'hf_transfer' package is not available`, you can install MFLUX with the `hf_transfer` package included:
+
+```sh
+uv tool install --upgrade mflux --with hf_transfer
+```
+
+This will enable faster model downloads from Hugging Face.
+
+</details>
+
+<details>
+<summary>DGX / NVIDIA (uv tool install)</summary>
+
+```sh
+uv tool install --python 3.13 mflux
+```
+</details>
+
+---
+
+### 🎨 Models
+
+MFLUX supports the following model families. They have different strengths and weaknesses; see each model’s README for full usage details.
+
+| Model | Release date | Size | Type | Training | Description |
+| --- | --- | --- | --- | --- | --- |
+|[Z-Image](src/mflux/models/z_image/README.md) | Nov 2025 | 6B | Distilled & Base | Yes | Fast, small, very good quality and realism. |
+|[Krea 2](src/mflux/models/krea2/README.md) | Jun 2026 | 12B | Turbo (distilled) | No | Very good quality with a wide range of styles; good for creative exploration. |
+|[FLUX.2](src/mflux/models/flux2/README.md) | Jan 2026 | 4B & 9B | Distilled & Base | Yes | Fastest + smallest with very good qaility and edit capabilities. |
+|[Ideogram 4](src/mflux/models/ideogram4/README.md) | Jun 2026 | 9B | Base | No | JSON-caption-native, typography-focused text-to-image generation. |
+|[ERNIE-Image](src/mflux/models/ernie_image/README.md) | Apr 2026 | 8B | Distilled & Base | No | Single-stream DiT from Baidu. Vivid, high-contrast output. |
+|[FIBO](src/mflux/models/fibo/README.md) | Oct 2025+ | 8B | Distilled & Base | No | Very good JSON-based prompt understanding. Has edit capabilities. |
+|[SeedVR2](src/mflux/models/seedvr2/README.md) | Jun 2025 | 3B & 7B | — | No | Best upscaling model. |
+|[Qwen Image](src/mflux/models/qwen/README.md) | Aug 2025+ | 20B | Base | No | Large model (slower); strong prompt understanding and world knowledge. Has edit capabilities |
+|[Depth Pro](src/mflux/models/depth_pro/README.md) | Oct 2024 | — | — | No | Very fast and accurate depth estimation model from Apple. |
+|[FLUX.1](src/mflux/models/flux/README.md) | Aug 2024 | 12B | Distilled & Base | No (legacy) | Legacy option with decent quality. Has edit capabilities with 'Kontext' model and upscaling support via ControlNet |
+
+---
+
+### ✨ Features
+
+**General**
+- Quantization and local model loading
+- LoRA support (multi-LoRA, scales, library lookup), including LyCORIS LoKr on FLUX.1 and FLUX.2
+- Metadata export + reuse, plus prompt file support
+
+**Model-specific highlights**
+- Text-to-image and image-to-image generation.
+- LoRA finetuning
+- In-context editing, multi-image editing, and virtual try-on
+- ControlNet (Canny), depth conditioning, fill/inpainting, and Redux
+- Upscaling (SeedVR2 and Flux ControlNet)
+- Depth map extraction and FIBO prompt tooling (VLM inspire/refine)
+
+See the [common README](src/mflux/models/common/README.md) for detailed usage and examples, and use the model section above to browse specific models and capabilities.
+
+> [!NOTE]
+> As MFLUX supports a wide variety of CLI tools and options, the easiest way to navigate the CLI in 2026 is to use a coding agent (like [Cursor](https://cursor.com), [Claude Code](https://www.anthropic.com/claude-code), or similar). Ask questions like: “Can you help me generate an image using z-image?”
+
+
+---
+
+<a id="related-projects"></a>
+
+### 🌱 Related projects
+
+- [MindCraft Studio](https://themindstudio.cc/mindcraft#models) — macOS app built on mflux by [@shaoju](https://github.com/shaoju)
+- [Mflux-ComfyUI](https://github.com/raysers/Mflux-ComfyUI) by [@raysers](https://github.com/raysers)
+- [MFLUX-WEBUI](https://github.com/CharafChnioune/MFLUX-WEBUI) by [@CharafChnioune](https://github.com/CharafChnioune)
+- [mflux-fasthtml](https://github.com/anthonywu/mflux-fasthtml) by [@anthonywu](https://github.com/anthonywu)
+- [mflux-streamlit](https://github.com/elitexp/mflux-streamlit) by [@elitexp](https://github.com/elitexp)
+- [mlx-taef](https://github.com/IonDen/mlx-taef) — TAESD/TAEF tiny-autoencoder live previews and low-memory FLUX decode for mflux, by [@IonDen](https://github.com/IonDen)
+- [mlx-teacache](https://github.com/IonDen/mlx-teacache) — TeaCache step-skipping to speed up FLUX generation in mflux, by [@IonDen](https://github.com/IonDen)
+
+---
+
+### 🙏 Acknowledgements
+
+MFLUX would not be possible without the great work of:
+
+- The MLX Team for [MLX](https://github.com/ml-explore/mlx) and [MLX examples](https://github.com/ml-explore/mlx-examples)
+- Black Forest Labs for the [FLUX project](https://github.com/black-forest-labs/flux)
+- Bria for the [FIBO project](https://huggingface.co/briaai/FIBO)
+- Tongyi Lab for the [Z-Image project](https://tongyi-mai.github.io/Z-Image-blog/)
+- Baidu for the [ERNIE-Image project](https://huggingface.co/baidu/ERNIE-Image)
+- Ideogram for the [Ideogram 4 project](https://huggingface.co/ideogram-ai/ideogram-4-fp8)
+- Krea.ai for the [Krea 2 project](https://www.krea.ai/blog/krea-2-technical-report)
+- Qwen Team for the [Qwen Image project](https://qwen.ai/blog?id=a6f483777144685d33cd3d2af95136fcbeb57652&from=research.research-list)
+- ByteDance, @numz and @adrientoupet for the [SeedVR2 project](https://github.com/numz/ComfyUI-SeedVR2_VideoUpscaler)
+- Hugging Face for the [Diffusers library implementations](https://github.com/huggingface/diffusers) 
+- Depth Pro authors for the [Depth Pro model](https://github.com/apple/ml-depth-pro?tab=readme-ov-file#citation)
+- The MLX community and all [contributors and testers](https://github.com/filipstrand/mflux/graphs/contributors)
+
+---
+
+### ⚖️ License
+
+This project is licensed under the [MIT License](LICENSE).