TextEncodeQwenImage21: reference inputs, resolution and latent output

Document the core class and its input/output contract without confusing UI capacity with model support.

Updated 2026-09-25Published revision: 7Source baseline: not recorded

Provenance

TextEncodeQwenImage21 is defined in the ComfyUI v0.37.0 core Qwen module, not a third-party package that must be located under the same name in Manager. Q04 Search the technical class identifier when translated display names differ.

Contract

Inputs include clip, positive and negative prompt text, optional VAE, resolution and dynamic reference-image inputs. Outputs are positive conditioning, negative conditioning and latent. The implementation uses an approximate resolution-squared pixel budget with 32-pixel alignment. With resolution=0, each reference keeps approximately its own dimensions after 32-pixel rounding. When a valid reference exists, the first one determines the empty output latent canvas; without any reference, the code uses resolution or 1024 if it is zero. Q04

Diagnose connections by their types

A conditioning output is not a latent input. Nor is an empty latent an already encoded finished edit. Check the port type and the template's intended role together. The existence of a latent output does not establish that another model architecture accepts its layout.

Separate a batch from separate references

The reviewed implementation takes the first batch item from each reference input. A ten-image batch on one port is not equivalent to ten individually connected references. The v0.37.0 node schema generates names through image_16, but ComfyUI's Qwen-Image-2.1 announcement describes support for up to 10 input images; the extra schema slots are not evidence of a higher supported model count. Q04 Q05 Establish whether the task is batch generation or multi-reference editing before changing the connections.

Catalog identity

The site records class identifier, owning package, observed version and source independently. A renamed menu label should not produce a new content identity. The version identifies the reviewed snapshot; it is not a perpetual compatibility guarantee.

Sources and verification

Source, code and language review on 2026-09-26. No GPU run, weight download or quality benchmark. Procedures are editorial guidance, not a confirmed diagnosis of your environment.