Qwen-Image 2.1 ControlNet Union Fun in ComfyUI vs reference images

The ComfyUI changelog for September 29, 2026 lists "Qwen-Image 2.1 ControlNet" as Union Fun ControlNet support for Qwen-Image 2.1 editing, plus a tiny VAE for faster previews and lower memory use. Sume's image docs describe no ControlNet input; guided edits go through reference images, and for ChatGPT Image 2.5 a mask.

ComfyUI facts are from its changelog, read 2026-10-01. Sume facts are from the Image API, read 2026-10-01.

What did ComfyUI add?

Under New Open-Source Model Support the changelog lists Qwen-Image 2.1 ControlNet and Qwen-Image 2.1 tiny VAE on September 29, 2026, alongside MiniMax-H3 Fun ControlNet Union 2.0. Qwen-Image-2.1 itself arrived in v0.37.0 on September 21, with native 2K generation and editing, an alpha channel and multi-image references. The changelog gives no workflow details for the ControlNet entry beyond that one line.

How do guided edits work on Sume?

Image models take input_references; the catalog example shows it as a range from 0 to 10 for one model, and each model lists its own range. ChatGPT Image 2.5 takes up to 16 image references, an optional mask_url and background: auto|transparent|opaque. These are reference and mask inputs, not edge, depth or pose control maps.

Can I get ControlNet-style layout control without ControlNet?

Partly, by describing the layout in the prompt and passing a reference image that shows it. The approach is described in pose and layout guidance without ControlNet. It gives looser control than a control map, and there is no control_context_scale equivalent in the cited docs.

Which Qwen ids does Sume list?

The image contract accepts qwen/qwen-image and qwen/qwen-image-max. A request that sets a parameter the selected model does not list is rejected with 400 unsupported_parameter, so read the catalog before sending references or masks.