Migrating to Qwen Image 2.1 without mixing pipeline generations

Preserve the working baseline

Keep the earlier workflow and its file provenance as a rollback baseline. The pinned ComfyUI 2.1 text-to-image template selects a separate diffusion model, Qwen3-VL text encoder and VAE, and ComfyUI v0.37.0 registers TextEncodeQwenImage21. Make a separate copy of that template, update ComfyUI as the template notes require, and verify those three files and node availability before changing prompts, references or adapters. First compare the base workflow without an adapter. This is a migration checklist, not a locally completed model run. Q04 Q05

A family name is not an architecture contract

A Qwen label does not establish interchangeability. Unified generation and editing describes this release; it does not imply that every historical checkpoint uses the same graph. Base compatibility is only one check alongside tensor layout, loader support and adapter provenance.

Separate runtime guides

The model author lists Diffusers, vLLM-Omni, SGLang and LightX2V as separate integrations. Its Diffusers example imports QwenImage21Pipeline and loads Qwen/Qwen-Image-2.1 with from_pretrained; the pinned ComfyUI template instead selects three Comfy-Org .safetensors files by node. These are distinct documented layouts, not interchangeable install commands or proof that a file converted for one runtime loads in another. Follow the chosen runtime's current instructions in a separate environment; do not install every framework to fix a ComfyUI node. Q01 Q05

Keep an experiment record

Record runtime, commit or release, template source, file revision, quantization, reference dimensions, output parameters, adapter use and actual output files. Third-party GGUF, INT4 and MLX conversions require independent checks of the publisher, loader and applicable terms. This page does not establish interchangeability between them. The Qwen Research License grants non-commercial research/evaluation use and requires a separate commercial license for commercial use; check the actual terms before deployment. Q02

Update old articles without rewriting history

Keep older guides scoped to their original release and link to 2.1. Do not replace version strings across old steps. Localized pages share identity and provenance; translation time is not a new benchmark date. This published guide records a source review, not a migration test or permission to update older articles automatically.

Sources and verification

Sources and English text rechecked 2026-09-26. No model download, ComfyUI or alternate-runtime execution, GPU run or quality benchmark. Procedures are editorial guidance, not a confirmed diagnosis of your environment.

• Q01 · Qwen-Image-2.1 official README
• Q04 · ComfyUI 0.37.0 Qwen nodes
• Q05 · Official Qwen 2.1 text-to-image template
• Q02 · Qwen Research License

Related reading

• Qwen Image 2.1 in ComfyUI: models, workflows and release boundaries
• Qwen Image 2.1 LoRAs: distinguish trainer support from downloadable adapters