Completed: Qwen edit parity expansion¶

Metadata¶

Created: 2026-05-28
Status: Completed
Completed: 2026-06-22

ADR status¶

Governing ADRs: ADR 0001, ADR 0002, ADR 0003
ADR impact: No new ADR was required. The shipped Qwen surface stays explicit, route-specific, and fail-closed under the existing generation contract.

Outcome¶

Completed as the current production-grade Qwen route expansion.

MLX-Gen now ships:

the first-class Qwen route matrix;
exact masked edit on qwen.inpaint;
exact structured control on qwen.control;
exact base-Qwen control-inpaint on qwen.control-inpaint;
exact public route proofs for Qwen 2511 multi-reference, reframe, and outpaint under the current LoRA contract.

The important design choice was not to invent another public qwen.edit-plus route id. MLX-Gen already exposes the user-visible multi-image behavior through qwen.multi-reference, so the route-matrix and capability contract document that mapping directly.

Closing code reality¶

mlxgen generate now exposes the Qwen route family through explicit capability rows: qwen.text, qwen.latent, qwen.edit, qwen.inpaint, qwen.multi-reference, qwen.reframe, qwen.outpaint, qwen.control, and qwen.control-inpaint.
The unified router injects the exact validated Qwen ControlNet sidecars and rejects conflicting --controlnet-model values.
The shipped guidance=1 Lightning proof path now skips inactive negative-prompt work and records only the effective negative prompt in metadata.
File-backed source/mask condition caches now invalidate when the input files change in place.
Qwen edit-family LoRA proof now covers the exact public routes users actually reach through mlxgen generate, not only the original single-image edit row.

Validation¶

Accepted public proof assets:

Representative proof sheets:

Focused tests now cover:

Qwen route selection and fail-closed option handling;
mask/control route separation;
exact sidecar enforcement;
Qwen prompt-path behavior at guidance=1;
current public validation-profile resolution.

Remaining boundary¶

Broader Qwen control families, layered composition, or new edit pipelines should land as new bounded items if they become important.
The current public Qwen claim stays exact-row oriented. Future expansion should only promote another Qwen route after it has its own accepted proof bundle.