Qwen-Edit-2509-Upscale-LoRA — community photo restoration adapter for Qwen-Image-Edit-2509

TL;DR

A community LoRA for Qwen/Qwen-Image-Edit-2509 that retargets the image-edit backbone to photo restoration / upscaling — recovering from low resolution (up to 16×), JPEG artifacts (quality as low as 5%), Gaussian blur (up to 3 px radius), motion blur (up to 64 px), color banding (3-bit), pixelation, and noise (up to 50%). Distributed as a two-stage A→B adapter pair via Hugging Face; usable with a 8-step Lightning LoRA at inference. 12.6k downloads in the first month and 71 dependent HF Spaces — another data point that the “tiny adapter on a frozen Qwen-Image-Edit-2509 DiT” recipe survives in the wild for very different task axes from the original camera-control LoRA.

Key claims

Trained on a filtered subset of Unsplash-Lite and UltraHR-100K at 2.4 MP resolution, restricted to photography subjects (landscapes, architecture, interiors, portraits, plants, vehicles, abstract photos, man-made objects, food) — explicitly not trained on 2D, illustrations, artworks, text, graphics, or CGI [model card §training data].
The adapter is shipped as a two-LoRA stack — “LoRA A and LoRA B are meant to be wired one after another” — currently qwen-edit-enhance_64-v3_000001000 (A) and qwen-edit-enhance_000004250 (B), with a V4 candidate pair also published [model card §current best, §V4 candidates].
Trained to recover from a wide range of degradations simultaneously: 16× upscaling, oversharpening, 50% noise, 3 px Gaussian blur, 5%-quality JPEG, 64 px motion blur, 16× pixelation, 3-bit color banding, and the outputs of other upscale models up to 16× [model card §degradation envelope].
Recommended inference recipe: ModelSamplingAuraFlow with shift kept below 0.3 (as low as 0.02 at higher resolutions); LCM sampler best, then Euler-Ancestral / Euler; intended to be used with the 8-step Lighting LoRA [model card §inference recipe].
Distribution channel: 12,592 downloads in the first month, 71 dependent HF Spaces, base model Qwen/Qwen-Image-Edit-2509; no published training procedure beyond the model card prose [HF page metadata].
Adoption signal: official amplification by @Ali_TongyiLab (Alibaba’s own foundation-model account, owner of the underlying Qwen-Image-Edit-2509) reaching 26K views — a community LoRA on a Qwen model being promoted by the Qwen team.

Method

A LoRA adapter on Qwen/Qwen-Image-Edit-2509 loaded through the standard Diffusers pipe.load_lora_weights(...) API. The card discloses no architectural detail (rank, target modules, hypernetwork, optimizer) and instead leans entirely on the degradation envelope, training data, and inference-recipe documentation as the artifact’s contract.

The novel structural decision relative to the camera-control LoRA on the same base (Qwen-Edit-2509-Multiple-angles LoRA — community camera-control adapter for Qwen-Image-Edit-2509) is the two-LoRA wiring: LoRA A is loaded first, then LoRA B is loaded “after” it — i.e. the adapter is intentionally factored into two stages that compose at inference. The card does not motivate this design choice, but the V4 candidates have asymmetric ranks (16 for A, 64 for B), suggesting A is a lighter “structural” adapter and B is a higher-capacity “detail” adapter. The user-facing prompt template is "Enhance image quality, followed by textual description of the scene" — i.e. the prompt slot is still used for content description, the adapter rewrites the noise-injection / sampling behavior to favor restoration.

This composes on top of the same stack pattern documented for the Multiple-angles LoRA: a community LoRA + an inference-acceleration LoRA (here: 8-step Lighting LoRA) loaded on the same frozen Qwen-Image-Edit-2509 backbone.

Results

No quantitative evaluation in the model card. The only empirical evidence on the page is qualitative — sample images claimed to be from 8-step Lighting LoRA inference — plus the adoption statistics (12.6k downloads, 71 dependent Spaces, official Tongyi Lab amplification). The card explicitly carves out the work as “still work in progress, checkpoints are being uploaded on the go,” with V3 currently labeled the best and V4 candidates in evaluation.

Why it’s interesting

This is the second community LoRA on Qwen/Qwen-Image-Edit-2509 that the wiki has tracked, after Qwen-Edit-2509-Multiple-angles LoRA — community camera-control adapter for Qwen-Image-Edit-2509. Both ship the same shape of artifact — a small adapter on a frozen image-edit DiT, intended to stack with an independently-trained few-step distillation LoRA, distributed without a paper — but for completely different task axes: dx8152 targets camera-axis control (a geometry-level transform), vafipas663 targets restoration / quality recovery (a degradation-inverse transform). The fact that both have crossed five-digit download counts and 70-100 dependent Spaces is concrete field evidence that Qwen-Image-Edit-2509 has become the community’s preferred frozen substrate for “tiny adapter repurposes the model for a structured task” — the same recipe pattern Parameter-Efficient Finetuning documents from the research side via DiffHDR (DiffHDR: Re-Exposing LDR Videos with Video Diffusion Models), LumiVid (HDR Video Generation via Latent Alignment with Logarithmic Encoding (LumiVid)), and Fit-VTO (FIT: A Large-Scale Dataset for Fit-Aware Virtual Try-On).

The two-LoRA A→B factorization is the new data point versus the camera-control adapter, which shipped as a single LoRA: vafipas663 is empirically claiming that intentional decomposition of the adapter into two sequentially-applied stages beats a single capacity-matched LoRA on the restoration task. This is adjacent to LoRWeB’s “basis-of-LoRAs beats single LoRA at matched capacity” finding (Spanning the Visual Analogy Space with a Weight Basis of LoRAs), although the LoRWeB result is a learned basis with a router, not a hand-wired pipeline.