MLX-Gen

MLX-Gen is a local image and video generation runtime for Apple Silicon and MLX. It exposes mlxgen for text-to-image, image-to-image, text-to-video, image-to-video, model download, model preparation, SeedVR2 image/video restoration and upscaling, optimized quantized model variants, and application progress callbacks.

Important

MLX-Gen started as a fork of mflux. Most credit for the current codebase goes to Filip Strand and the original mflux contributors. MLX-Gen keeps that foundation and focuses on making local generation predictable for real applications: users download or prepare models before running a job, then use one mlxgen command to generate, inspect supported modes, and compare generated examples and measured results. MLX-Gen contributes by adding tested T2I/I2I routes, Qwen Image Edit 2509/2511 routing and parity fixes, Bonsai Image support, Wan2.2 text-to-video, image-to-video, and video-to-video support, model-specific mixed quantization policies, published quantized Hugging Face repos optimized for MLX-Gen, and progress callbacks for apps. The fork exists so AbstractVision and related AbstractFramework projects can move quickly without losing the option to merge useful changes back upstream if that becomes valuable for the wider mflux community.

This screenshot shows AbstractFlow, a visual workflow authoring tool in the AbstractFramework ecosystem. AbstractFlow can orchestrate generative image/video capabilities exposed through AbstractVision and AbstractCore; the same MLX-Gen steps can also be run directly with the command-line examples below.

What It Does

MLX-Gen runs supported Hugging Face source models after you download them locally. It also runs quantized model variants that are published on Hugging Face for MLX-Gen. Some of those variants keep precision-sensitive layers at 8-bit or BF16 instead of blindly quantizing every layer, which is why they are published as MLX-Gen-specific repos. You explicitly download or prepare models first, then generation uses only files already on disk, which suits desktop apps, workflow engines, and long-running local jobs.

The main capabilities are:

• text-to-image generation with Qwen Image, FLUX.2 Klein, Z-Image, ERNIE Image Turbo, Bonsai Image, FIBO, and their optimized quantized variants where available;
• image-to-image modes, including latent img2img, instruction/reference edits, multi-reference edits, masked edit/inpaint where the selected model supports --mask-path, base-Qwen control-inpaint on the exact validated AbstractFramework/qwen-image-8bit row, Qwen structured control where the selected model supports --controlnet-image-path, Z-Image Turbo native inpaint on the exact validated AbstractFramework/z-image-turbo-8bit row, and route-specific reframe/outpaint workflows where the selected model supports them;
• Wan2.2 text-to-video, image-to-video, and prompt-guided video-to-video (plain or masked via --video-mask-path, which locks everything outside the mask to the source video), including TI2V-5B BF16/q8 packages plus A14B T2V/I2V BF16 and mixed q8/BF16 packages; video-to-video resamples the source onto the requested fps timeline (output keeps real-time speed) and copies source audio onto the output best-effort; the current public video-to-video route is limited to Wan2.2-T2V-A14B, uses one source video plus one prompt, requires --solver unipc, and does not include reference images or VACE-style learned conditioning; Wan I2V resolves output size from the source image aspect ratio so inputs are not stretched into a mismatched canvas;
• SeedVR2 image and video restoration through mlxgen upscale, with official 3B/7B source support including the dedicated seedvr2-7b-sharp route, published q8/q4 packages, shortest-edge target sizing, explicit scale factors such as 2x and 3x, streamed restore for longer clips, preserved source FPS, a conservative host-safe default video profile, and published five-second 1x and 2x validation bundles for the 3B/7B video path;
• explicit download and prepare workflows for local MLX-Gen model packages;
• JSON model capability inspection before starting a heavy run;
• strict route-specific LoRA routing and adapter application checks, with model-card compatibility preflight when cached adapter metadata is available and exact public proof rows for Qwen Image Edit original/2509/2511, base Qwen Image text and latent, Qwen Image 2512 text, the base Qwen Image q8 structured-control and control-inpaint rows, the exact AbstractFramework/z-image-turbo-8bit text and latent img2img rows, FLUX.2 Klein 9B edit and multi-reference, FLUX.2 Klein base 4B outpaint, ERNIE Image Turbo text and latent img2img, and all supported Wan q8 video routes including Lightning fast video-to-video; the LoRA guide documents the exact validated adapters per route, strict scale matching, copy-paste owner/repo:subdir/file adapter forms, and the validation evidence behind each proof row (same-seed A/B sheets for the image and T2V/I2V rows, a bounded multi-seed matrix for the video-to-video row). Those Lightning examples are documented against MLX-Gen q8 packages, not arbitrary external FP8 checkpoints. Bonsai LoRA stays fail-closed;
• shared progress events for applications embedding MLX-Gen.

Use mlxgen capabilities --model ... before long image-edit runs. Capability output describes the available route; validation reports and contact sheets describe whether an exact source handle or MLX-Gen optimized package passed a visual release gate. Release evidence should use true handles such as briaai/Fibo-Edit or AbstractFramework/flux.2-klein-9b-8bit, not short aliases.

Install

Install with uv:

uv tool install --upgrade mlx-gen

Or install into an environment:

python -m pip install -U mlx-gen

Check the command surface:

mlxgen --help

For application integration and shell-outs, use the mlxgen commands shown in mlxgen --help. The package still installs some mflux-generate-* compatibility entry points from the upstream codebase, but those are not the recommended integration surface for new tools.

First Commands

Download model files explicitly:

mlxgen download --model AbstractFramework/flux.2-klein-9b-8bit

Generate an image:

mlxgen generate \
  --model AbstractFramework/flux.2-klein-9b-8bit \
  --prompt "A cinematic wide shot of a compact sci-fi spaceship resting in deep snow on a frozen alien planet" \
  --width 768 \
  --height 432 \
  --steps 24 \
  --guidance 1.0 \
  --seed 6107 \
  --output spaceship.png

Generate several output images or videos in one run by passing more than one seed:

mlxgen generate \
  --model qwen-image \
  --prompt "A clean studio product photo" \
  --seed 101 202 303 \
  --output product.png

MLX-Gen saves one artifact per seed and appends the seed to the output stem automatically, for example product_seed_101.png, product_seed_202.png, and product_seed_303.png. mlxgen generate and mlxgen upscale also support --auto-seeds N when you want several random variations from one command. Duplicate explicit seeds are rejected, --auto-seeds must be greater than zero, --output supports {seed}, and SeedVR2 multi-source runs also support {input_name} with legacy {image_name} accepted as a compatibility alias.

Upscale an image with SeedVR2:

mlxgen download --model AbstractFramework/seedvr2-3b-8bit

mlxgen upscale \
  --model AbstractFramework/seedvr2-3b-8bit \
  --image-path input.png \
  --resolution 2x \
  --softness 0.25 \
  --metadata \
  --output input_2x.png

For SeedVR2, an integer --resolution is the target shorter edge while values such as 2x and 3x are scale factors. Both modes preserve the source aspect ratio. Use --softness 0.25 to 0.5 when the source has visible grain in smooth areas. Small image outputs use the untiled VAE path; large image outputs automatically use tiled VAE decode, and --vae-tiling forces tiled VAE encode/decode for image runs only. See docs/upscaling.md for a reproducible 5x SeedVR2 image comparison plus the accepted five-second Eiffel 1x and 2x 3B / 7B restore proof bundles.

Restore a real five-second validation clip with SeedVR2:

mlxgen upscale \
  --model ByteDance-Seed/SeedVR2-3B \
  --video-path input.mp4 \
  --start-seconds 70 \
  --max-frames 149 \
  --resolution 1x \
  --softness 0.0 \
  --color-correction wavelet \
  --temporal-chunk-size 29 \
  --temporal-chunk-overlap 8 \
  --low-ram \
  --mlx-cache-limit-gb 8 \
  --metadata \
  --output restored.mp4

For video inputs, SeedVR2 preserves the source FPS by default and preserves the matching source audio segment by default as well. If MLX-Gen cannot prove that copied audio is still aligned safely, the run fails instead of publishing a silent output unexpectedly. Use --drop-audio only when you intentionally want a silent restored MP4. (This is deliberately stricter than Wan video-to-video's best-effort audio copy: restoration is a fidelity contract, while a failed mux on a generative run must not discard the finished video.) The safe public video profile defaults to 1x, enables --low-ram automatically, and rejects enlarged video output unless you explicitly pass --force-unsafe-video-memory. See docs/upscaling.md for the accepted five-second Eiffel quality proof bundle and the published Air France audio-copy proof bundle.

mlxgen upscale accepts one or more explicit seeds and --auto-seeds N on both image and video restore runs. When you process several seeds, MLX-Gen appends _seed_<seed> automatically. When one SeedVR2 invocation processes several source files, MLX-Gen also appends the source-file stem so each saved artifact gets its own path. If two source files share the same basename, keep --replace false or rename the inputs; overwrite-prone SeedVR2 batches are rejected when --replace true.

Inspect model capabilities before a run:

mlxgen capabilities --model AbstractFramework/flux.2-klein-9b-8bit

Capabilities are route contracts: they show which tasks, I2I modes, image counts, and options the selected model can dispatch. For release QA evidence on exact packages, use:

mlxgen validation --model AbstractFramework/qwen-image-edit-2509-8bit

LoRA support is route-specific. For LoRA work, inspect supports_lora, lora_status, and lora_validation_profile in mlxgen capabilities, download the adapter explicitly with mlxgen download, and use an adapter trained for the selected model family and route. Current exact proof rows cover original Qwen Image Edit, Qwen Image Edit 2509/2511, base Qwen Image text and latent, Qwen Image 2512 text, the exact AbstractFramework/z-image-turbo-8bit text and latent rows, FLUX.2 Klein 9B edit and multi-reference, FLUX.2 Klein base 4B outpaint, ERNIE Image Turbo text and latent, plus the current Wan q8 public video routes. The LoRA guide also includes the current LightX2V 4-step A14B timing comparison against the practical original Wan profiles, same-seed no-LoRA-versus-Lightning A/B sheets, recommended 4-step Qwen 2512 and Qwen Edit 2511 Lightning examples, a 240p-versus-480p T2V sweep, and copy-paste download and T2V/I2V commands for lightx2v/Wan2.2-Lightning, including the stable repo:subdir/file.safetensors form and equivalent absolute local-file usage after download. For example, a FLUX.2-dev LoRA is not accepted for FLUX.2 Klein. See docs/lora.md for the A/B validation method.

Create a local MLX-Gen model package, for example an 8-bit Qwen Image package:

mlxgen prepare \
  --model Qwen/Qwen-Image \
  --path ./models/qwen-image-8bit \
  --quantize 8

mlxgen generate does not download missing files. If something is not cached, MLX-Gen raises a clear DownloadRequiredError with the command to run. A complete local MLX-Gen package at ./models/<repo-name> can also satisfy a matching Hugging Face handle such as AbstractFramework/qwen-image-edit-2511-8bit.

Reproducible Example

The docs include a complete model-backed spaceship workflow:

• T2I: generate a spaceship in the snow.
• I2I edit: turn it into a pencil sketch.
• I2I edit: crash the same spaceship in the snow.
• I2I multi-reference: combine the crash layout and pencil-sketch style.
• T2V A14B: generate a spaceship taking off from a snow planet.
• I2V A14B: animate the generated spaceship taking off from the source image.

See docs/examples/spaceship-snow.md for the exact commands and included assets.

For image-edit contact sheets, command logs, and model/package status across Qwen Image Edit, Qwen Image Edit 2509/2511, FLUX.2 Klein, and latent I2I models, see docs/edit-capabilities.md. That guide also includes the exact base Qwen q8 structured-control proof route, the exact base Qwen q8 control-inpaint proof route, and the exact Z-Image Turbo q8 native-inpaint proof route. For the full Qwen route map, including how mlxgen capability ids line up with upstream Diffusers Qwen pipelines and where the published proof surfaces live, see docs/qwen-route-matrix.md. For a plain-language guide to latent img2img, instruction edit, masked edit/inpaint, multi-reference composition, Qwen structured control, generative reframe, and outpaint, see docs/image-edit-modes.md.

Published Models

Quantized and BF16 model variants optimized for MLX-Gen are published under the AbstractFramework organization on Hugging Face. Current published packages include:

FLUX.2 Klein:

• AbstractFramework/flux.2-klein-4b-4bit
• AbstractFramework/flux.2-klein-4b-8bit
• AbstractFramework/flux.2-klein-9b-4bit
• AbstractFramework/flux.2-klein-9b-8bit
• AbstractFramework/flux.2-klein-base-4b-4bit
• AbstractFramework/flux.2-klein-base-4b-8bit
• AbstractFramework/flux.2-klein-base-9b-4bit
• AbstractFramework/flux.2-klein-base-9b-8bit

Qwen Image and Qwen Image Edit:

• AbstractFramework/qwen-image-4bit
• AbstractFramework/qwen-image-8bit
• AbstractFramework/qwen-image-2512-4bit
• AbstractFramework/qwen-image-2512-8bit
• AbstractFramework/qwen-image-edit-4bit
• AbstractFramework/qwen-image-edit-8bit
• AbstractFramework/qwen-image-edit-2509-4bit
• AbstractFramework/qwen-image-edit-2509-8bit
• AbstractFramework/qwen-image-edit-2511-4bit
• AbstractFramework/qwen-image-edit-2511-8bit

Z-Image, ERNIE, and FIBO:

• AbstractFramework/z-image-4bit
• AbstractFramework/z-image-8bit
• AbstractFramework/z-image-turbo-4bit
• AbstractFramework/z-image-turbo-8bit
• AbstractFramework/ernie-image-turbo-4bit
• AbstractFramework/ernie-image-turbo-8bit
• AbstractFramework/fibo-4bit
• AbstractFramework/fibo-8bit

SeedVR2 upscaling:

• AbstractFramework/seedvr2-3b-4bit
• AbstractFramework/seedvr2-3b-8bit
• AbstractFramework/seedvr2-7b-4bit
• AbstractFramework/seedvr2-7b-8bit

SeedVR2 7B can also run from the official ByteDance-Seed/SeedVR2-7B source model, including the seedvr2-7b-sharp checkpoint alias, or from the published q8/q4 package handles above. See docs/upscaling.md and docs/quantization.md for the validated 7B source/q8/q4 profile.

Wan2.2 video:

• AbstractFramework/wan2.2-ti2v-5b-diffusers-bf16
• AbstractFramework/wan2.2-ti2v-5b-diffusers-8bit
• AbstractFramework/wan2.2-t2v-a14b-diffusers-bf16
• AbstractFramework/wan2.2-t2v-a14b-diffusers-8bit
• AbstractFramework/wan2.2-i2v-a14b-diffusers-bf16
• AbstractFramework/wan2.2-i2v-a14b-diffusers-8bit

Use mlxgen download --model <repo-id> to cache a published model, then pass the repository id to the relevant command: mlxgen generate for image/video generation or mlxgen upscale for SeedVR2 upscaling. See docs/quantization.md for the complete current package matrix with source sizes, optimized package sizes, task coverage, and quantization notes.

For Wan2.2 TI2V-5B, the published BF16 MLX-Gen package is 21.2 GiB versus 31.9 GiB for the upstream source snapshot. It is mainly a smaller reusable source-equivalent package because MLX-Gen already loads Wan transformer/VAE weights at BF16 runtime precision. The published q8 package is 16.9 GiB. Since the 2026-06-12 runtime-precision fix, Wan q8 packages dequantize all transformer-block linears to BF16 at load, so q8 is a storage/download saving only and runtime memory matches the BF16 package. Wan TI2V-5B q4 or mixed q4/q8 is not published as a supported package. See the exact benchmark profile and the dated correction in docs/quantization.md.

Wan A14B Measurements

Wan A14B was measured on an Apple M5 Max with 128 GB unified memory. The published-card benchmark uses small, repeatable low-RAM runs and records full-process Darwin physical footprint, RSS, MLX allocator peak, and generation time. Use these values for the listed profiles; memory and runtime scale with resolution, frame count, step count, cache settings, and image-to-video conditioning.

Correction (2026-07-05): the q8 rows below predate the 2026-06-12 runtime-precision fix. Since that fix, Wan q8 packages dequantize all transformer-block linears to BF16 at load, so runtime memory matches the BF16 rows (re-measured: 27.8 GiB MLX peak at the same T2V profile). Plan memory as if running the BF16 package; q8 saves disk and download only.

Model	Package	Disk	Physical Peak	Max RSS	MLX Peak	Time	Profile
Wan2.2 T2V-A14B	BF16	64.1 GiB	33.0 GiB	31.8 GiB	27.7 GiB	152.7 s	384x224, 33 frames, 12 steps, 8 fps
Wan2.2 T2V-A14B	mixed q8/BF16	39.5 GiB	20.7 GiB	19.5 GiB	15.5 GiB	154.8 s	384x224, 33 frames, 12 steps, 8 fps
Wan2.2 I2V-A14B	BF16	64.1 GiB	33.7 GiB	31.8 GiB	28.2 GiB	228.2 s	384x384, 33 frames, 12 steps, 8 fps
Wan2.2 I2V-A14B	mixed q8/BF16	39.5 GiB	21.5 GiB	19.6 GiB	15.9 GiB	242.2 s	384x384, 33 frames, 12 steps, 8 fps

In these runs, mixed q8/BF16 reduces disk usage by about 38% versus BF16 MLX-Gen packages; since 2026-06-12 it is not a runtime-memory or speed improvement. See docs/quantization.md for model-family quantization details and metrics JSON. The 0.18.11 release also validates the published A14B q8 T2V/I2V handles on a 41-frame, 15-step, 480x240-target profile with saved MP4/contact-sheet evidence in the quantization docs.

Ecosystem

MLX-Gen is used as the local Apple Silicon generation backend for:

• AbstractVision, the direct integration layer for abstracting generative image and video capabilities across local and hosted providers;
• AbstractCore, which can expose OpenAI-compatible endpoints backed by AbstractVision providers, including image and video capabilities;
• AbstractFlow, a visual workflow authoring layer that can compose generative image/video nodes alongside other media, text, and agent workflows.

MLX-Gen remains useful as a standalone CLI package, and it is also designed for applications: jobs run from model files already on disk, apps can inspect supported modes before loading weights, and progress callbacks make long runs observable.

Documentation

• Getting started: installation, first runs, SeedVR2 upscaling, and Wan video.
• API and CLI: command surface, router behavior, image-to-image modes, generative reframe, backend-specific outpaint, SeedVR2 sizing, Wan video sizes, capabilities, and Python entry points.
• Image edit modes: what latent img2img, edit-reference, multi-reference, generative reframe, and outpaint mean in practice, with examples.
• Wan video: practical Wan2.2 T2V/I2V sizing, plain and masked prompt-guided A14B video-to-video with included proof artifacts, a measured motion-fidelity ladder (strength vs gesture preservation), broader A14B target size families, and 5-second M5 Max comparison clips.
• Example workflow: reproducible image and video commands.
• Image upscaling: SeedVR2 sizing, published 3B/7B q8/q4 package usage, the host-safe video restore profile, published five-second Eiffel 1x and 2x 3B/7B validation bundles, readable tone-correction labels, and 5x source/output comparisons.
• Image edit capabilities: image-edit contact sheets, exact model/package status, and command logs.
• Reframe and outpaint: --reframe-padding and --outpaint-padding routes with the mixed June 8 profile, FLUX.2 Klein base source-model validation, and links to the exact Qwen and FLUX route proofs.
• Model management: download, prepare, and run from local model files.
• Quantization: q8/q4/BF16 policies and measurements.
• Python integration: route-resolved runtime loading, serial multi-output reuse for unified mlxgen generate families, SeedVR2's direct-model boundary, progress callbacks, and AbstractVision/AbstractCore notes.
• FAQ: recurring questions, image-to-image mode selection, SeedVR2 sizing, Qwen edit variants, negative prompts, reframe/outpaint, Wan resolutions, and usage limits.
• Troubleshooting: common setup and runtime failures.
• Acknowledgements: upstream mflux and model-community credits.

License

MLX-Gen is MIT licensed. Model weights remain governed by their original licenses and access terms.