Qwen Route Matrix¶
This page is the current route truth for Qwen Image and Qwen Image Edit in MLX-Gen. It maps the
public mlxgen surface to the upstream Diffusers Qwen pipeline families and points at the exact
proof assets that already exist.
Use this page for three questions:
- which Qwen route does MLX-Gen actually expose;
- what user inputs select that route;
- where the current public proof for that route lives.
For exact per-model option support, use:
mlxgen capabilities --model AbstractFramework/qwen-image-edit-2511-8bit
For exact release evidence on one row, use:
mlxgen validation --model AbstractFramework/qwen-image-edit-2511-8bit
Current Route Map¶
| Capability id | Upstream parity | Typical handles | Public inputs | Current proof surface |
|---|---|---|---|---|
qwen.text |
QwenImagePipeline |
Qwen/Qwen-Image, AbstractFramework/qwen-image-8bit, AbstractFramework/qwen-image-2512-8bit |
--prompt |
Route is shipped. Current exact LoRA-backed proof rows exist for AbstractFramework/qwen-image-8bit and AbstractFramework/qwen-image-2512-8bit in LoRA. |
qwen.latent |
QwenImageImg2ImgPipeline |
base Qwen Image handles | --image + --image-strength + --prompt |
Route is shipped. An exact LoRA-backed public proof now exists for AbstractFramework/qwen-image-8bit; see LoRA. |
qwen.edit |
QwenImageEditPipeline |
Qwen/Qwen-Image-Edit, Qwen/Qwen-Image-Edit-2509, Qwen/Qwen-Image-Edit-2511, prepared q8/q4 variants |
--image + --prompt |
Image edit capabilities covers the source/q8/q4 edit matrices. |
qwen.inpaint |
QwenImageEditInpaintPipeline |
edit-model rows, with the current exact proof on AbstractFramework/qwen-image-edit-2511-8bit |
--image + --mask-path + --prompt |
Qwen localized editing links the accepted masked-edit proof. |
qwen.base-inpaint |
QwenImageInpaintPipeline |
base Qwen rows other than the exact control-inpaint row: Qwen/Qwen-Image, AbstractFramework/qwen-image-4bit, AbstractFramework/qwen-image-2512-8bit |
--image + --mask-path + --prompt |
Masked editing links the current visual-smoke proof rows (4bit and 2512-8bit). |
qwen.multi-reference |
QwenImageEditPlusPipeline behavior |
Qwen/Qwen-Image-Edit-2509, Qwen/Qwen-Image-Edit-2511, prepared q8/q4 variants |
repeated --image + --prompt |
Image edit capabilities links the accepted 2509 and 2511 multi-reference matrices, and LoRA covers the exact 2511 q8 multi-angle proof row. |
qwen.reframe |
MLX-Gen canvas expansion on the Qwen edit route | original/2509/2511 edit families and prepared variants | --image + --reframe-padding + --prompt |
Reframe and outpaint links the accepted Qwen reframe matrices, and LoRA covers the exact 2511 q8 multi-angle proof row. |
qwen.outpaint |
MLX-Gen canvas expansion on the Qwen edit route | original/2509/2511 edit families and prepared variants | --image + --outpaint-padding + --prompt |
Reframe and outpaint links the accepted Qwen outpaint matrices, and LoRA covers the exact 2511 q8 multi-angle proof row. |
qwen.control |
QwenImageControlNetPipeline |
current exact public row: AbstractFramework/qwen-image-8bit |
--controlnet-image-path + --prompt |
Qwen localized editing links the accepted structured-control proof. |
qwen.control-inpaint |
QwenImageControlNetInpaintPipeline |
current exact public row: AbstractFramework/qwen-image-8bit |
--image + --mask-path + --prompt |
Qwen localized editing links the accepted control-inpaint proof. |
Important Public-Boundary Choices¶
MLX-Gen intentionally does not expose every upstream Qwen pipeline name as a separate public route id.
The current choices are:
QwenImageEditPlusPipelinebehavior is exposed throughqwen.multi-reference, not through a separate publicqwen.edit-plusid;- every base Qwen row exposes exactly one masked route: the exact validated
AbstractFramework/qwen-image-8bitrow keepsqwen.control-inpaint(its ControlNet sidecar and full-schedule behavior are unchanged), and every other trusted base row exposes nativeqwen.base-inpaintinstead. The same--image + --mask-path + --promptrequest selects whichever masked route the row carries; see Masked editing for the behavior differences; QwenImageLayeredPipelineis not a shipped public MLX-Gen route today.
That keeps the user-facing contract smaller while preserving honest route identity through
mlxgen capabilities.
Which Qwen Route To Choose¶
| Need | Route |
|---|---|
| Generate from text only | qwen.text |
Restyle one existing image with explicit --image-strength |
qwen.latent |
| Edit one source image with an instruction | qwen.edit |
| Edit only a masked part of one source image on the edit checkpoint | qwen.inpaint |
| Edit only a masked part of one source image on a base checkpoint | qwen.base-inpaint (or qwen.control-inpaint on the exact 8bit row) |
| Combine two or more reference images | qwen.multi-reference |
| Generate from text while fixing layout from a control image | qwen.control |
| Edit a masked source region on base Qwen with the stricter inpaint sidecar | qwen.control-inpaint (exact AbstractFramework/qwen-image-8bit row) |
| Expand the canvas outward | qwen.reframe or qwen.outpaint, depending on whether you want zoom-out or canvas extension |