qwen-image-flash-loras

Qwen-Image LoRA gallery (MLX)

The style-LoRA gallery used by the Qwen Image demo app for nvidia/Qwen-Image-Flash, mirrored from the qwen-image-flash-lora-dlc Space so the gallery has a stable home.

Why mirror at all: the Space's loras.json references ~25 repos under an account that has already been renamed once (lichorosario → furaidosu) — every one of those references resolves only through a redirect today. A gallery that breaks when an author renames or deletes a repo is not a gallery you can ship in an app.

These adapters were trained on Qwen/Qwen-Image. Flash is a DMD2 student of that same teacher with the architecture intact (60 layers, 24 heads, joint_attention_dim 3584), so they load by shape. Every adapter here was verified to bind against the Flash DiT through the MLX port's dialect remapper — 24 bind all 14 target key types, the rest bind 12 (they simply don't train the modulation projections). None bind zero, which is the silent failure mode worth checking: an unrecognized dialect matches nothing, renders the base model, and raises no error.

Manifest

loras.json — {schemaVersion, base, adapters[]}. Each adapter carries id, title, repo, file, trigger, defaultStrength, aspect, optional versions[], and provenance (upstream repo + original weight filename + license).

Consumers resolve {repo, file} per adapter, so a mirrored entry and an upstream-referenced one are handled identically.

Mirrored here (23)

Weights redistributed under the upstream license shown.

Referenced upstream, not mirrored (4)

These declare no usable redistribution grant — license: other with no license name or link, or no license field at all. They stay selectable in the gallery (their entries point at the original repo) but their weights are deliberately not redistributed here. If an author states a license, they can move into the mirror.

License

Mixed, per adapter — see the tables above and PROVENANCE.md; the repo-level tag is other because a single SPDX id would misrepresent the set. Adapters are redistributed under their own upstream terms, which travel with them:

• CreativeML OpenRAIL-M — permits redistribution and commercial use; the Section II use restrictions must be passed on to every downstream recipient, and they apply to you. Canonical text: https://huggingface.co/spaces/CompVis/stable-diffusion-license
• Apache-2.0 — permissive; retain the notice.

The base model is separate: nvidia/Qwen-Image-Flash is under the NVIDIA Open Model License, not Apache-2.0 like Qwen-Image.

Nothing here is authored by us. Credit belongs to the original creators listed in the Source column; this repo only re-hosts and documents.