Possible Qwen-Image sampling bug in DiffSynth-Studio (Issue #823)

TL;DR

A community-reported GitHub issue on modelscope/DiffSynth-Studio flagging a possible bug in the Qwen-Image sampling pipeline. The issue body could not be retrieved at filing time (fetcher resolved to the repo README), so the exact failure mode is unknown — filed here as a pointer to track if reproduction details surface. Adjacent issues on the same repo (#832, #1048) discuss CFG sensitivity and scheduler-config mismatches for Qwen-Image and Qwen-Image-Edit when running few-step distilled / lightning LoRA variants, which is the kind of bug class this report likely sits in.

Key claims

DiffSynth-Studio is the ModelScope-maintained open-source Diffusion engine that hosts the Qwen-Image inference + training pipelines [README §Introduction].
The repo’s Qwen-Image pipeline is diffsynth.pipelines.qwen_image.QwenImagePipeline, used for both text-to-image and the Qwen-Image-Edit / 2509 / 2511 variants [README update history].
Adjacent issues on the same repo flag CFG-related quality regressions (#832: “qwen image lora may get bad result when cfg is 4”) and scheduler-config mismatches when porting few-step / lightning LoRAs (#1048: scheduler shift / dynamic-shifting config differs from official) [search].

Method

Not applicable — this is a community bug report, not a method. The reporter (per Jiaming’s Slack note) suspects the Qwen-Image sampling path in DiffSynth-Studio diverges from the intended behavior. Without the issue body, the specific divergence (scheduler config, CFG handling, timestep schedule, latent normalization, etc.) is undetermined. Best next step for anyone reproducing internally: compare DiffSynth’s QwenImagePipeline sampling loop against the reference implementation in the Qwen-Image technical report.

Results

No quantitative results. The issue is a reproduction request / bug filing rather than a finding.

Why it’s interesting

Qwen-Image is one of the more heavily-used open-source image-generation backbones internally (filed as Qwen-Image Technical Report) and DiffSynth-Studio is the path through which most Qwen-Image LoRAs and ControlNets in the wider community are trained and served. A sampling-side bug there would silently degrade every downstream evaluation that runs Qwen-Image through this engine — including community LoRA ablations like Qwen-Image-Layered-Control: text-guided layer extraction and Qwen-Edit-2509-Multiple-angles LoRA — community camera-control adapter for Qwen-Image-Edit-2509. Worth confirming before trusting any Qwen-Image benchmark numbers that came out of DiffSynth.