Qwen-Image 2.1 in ComfyUI: Native 2K, RGBA Output, Multi-Image Editing and Prompt-Following Tests

Qwen-Image 2.1 is an open-source image model that supports text-to-image, image editing and RGBA output with an alpha channel. This episode deploys it as a one-click mirror and runs a set of practical tests. The notes below come from the episode subtitles and the published mirror page; no workflow JSON was shipped or parsed, so no node-level parameters are claimed beyond what the episode states.

What the model supports

• Text-to-image generation.
• Image editing with up to 10 input images (multi-image editing).
• RGBA / alpha-channel output.
• Native 2K generation, up to roughly 4 megapixels.
• ComfyUI added official support quickly.

The episode describes community reception as mixed: some users find the editing capability very strong, others see an obvious "AI look" and breakdowns on complex detail. The author's view is that it is a good model that still needs another optimization pass.

The two quick-start workflows

Text-to-image. The official route uses CFG 1; the ComfyUI official graph sets 25 steps. In the episode the author tested 40 steps, euler + simple, denoise 1 — close to the official defaults. Resolution can go to 2K, i.e. about 4 megapixels.

Image editing. Supports up to 10 input images. There is a latent-size toggle: when true, the size comes from the resolution selector; when false, the first image's size is used. In the prompt you can refer to reference images as image1, image2, and so on. Raising CFG makes the negative prompt take effect.

Test results from the episode

• Chinese text: a long-standing Qwen strength; the Chinese strings listed in the prompt rendered correctly, and two extra groups stayed stable.
• Prompt following: a red apple, yellow lemon, perfume bottle and blue ceramic cup test passed on count, white background and photographic feel. Occlusion was not perfect — the lemon had some overlap — which the author suggests could be improved with higher CFG and a negative prompt. A harder variant passed 100%.
• Realistic detail: an extreme close-up of skin showed slight texture transitions; acceptable, but the author notes portrait aesthetics may not beat older Creature 2 or Z1-style models, and using Qwen-Image 2.1 for portraits would need fine-tuning.
• Multi-person group shots were also tested in the episode.

Deploying the mirror

1. Open the mirror page: https://www.compshare.cn/images/ZJKkR4K2CSMI?referral_code=A7fTQUYY9lCDFJwA6MMvoq
2. Deploy an instance and wait for the image to come up.
3. Open ComfyUI, pick the quick-start workflow, set the prompt and run.

Links

• Mirror (with referral code): https://www.compshare.cn/images/ZJKkR4K2CSMI?referral_code=A7fTQUYY9lCDFJwA6MMvoq
• Video tutorial (YouTube): https://www.youtube.com/watch?v=UA0ZfVsvGFI
• Chinese original (Bilibili): https://www.bilibili.com/video/BV1pJhe6jEvN

Limits

• The CFG/step values above are the episode's settings and official defaults, not a universal optimum.
• Test results are single-run observations, not a controlled benchmark.
• The mirror is a deployment convenience; quality depends on prompt, seed and sampler settings.