0.2.1 beta — Z-Image Turbo Two Character Composer

Fix release for the first public beta of the standalone Z-Image Turbo Two Character Composer.

Fixed

• Saved workflows that only carry positional values, from frontends that do not write widgets_values_named, restored most fields into the wrong widgets. The interface script used the Krea2 widget order instead of this node's. It now follows the node definition, and a test compares the two so they cannot drift apart again.
• The two example workflows list their values in the node's own order. Copies of the 0.2.0 examples are still restored correctly.

The node's inputs, outputs, widget order and generation behavior are unchanged from 0.2.0.

Download comfyui-zimage-two-character-composer-0.2.1.zip, extract it under ComfyUI/custom_nodes, restart ComfyUI and refresh the browser. The -source.zip also includes tests and development tools. The README is in English and includes installation, wiring and usage examples.

Validation: 32 Python tests, 30 frontend tests and 401 native ComfyUI hook/sampler cases passed on CPU against current ComfyUI master, including two real rank-one LoRA tensor projections. Model execution, CLIP and file parsing in the integration harness use test doubles.

Beta limitation: full Z-Image GPU generation with real checkpoints, LoRAs and the running ComfyUI interface has not been validated here. Native regional hooks typically require two model predictions per step, with a possible third for uncovered background. This is a different implementation from the Krea2 attention engine.

0.2.0 beta — Z-Image Turbo Two Character Composer

First public beta of the standalone Z-Image Turbo Two Character Composer.

• Two character prompts, trigger fields and independent regional LoRA hook groups.
• Left/right and top/bottom percentages, draggable Free boxes and optional masks.
• Live aspect-ratio preview and Swap characters.
• Common Prompts and inline character LoRA tags.
• Optional CFG 1 negatives through the separately installed NegPiP Z-Image patch.
• Matching negative/report outputs and runtime metadata for Cyberdelia Image Metadata 2.3.1+.
• Two core-only examples, including one Boolean that switches model, positive, negative and latent together.

Download comfyui-zimage-two-character-composer-0.2.0.zip, extract it under ComfyUI/custom_nodes, restart ComfyUI and refresh the browser. The -source.zip also includes tests and development tools. The README is in English and includes installation, wiring and usage examples.

Validation: 31 Python tests, 29 frontend tests and 401 native ComfyUI hook/sampler cases passed on CPU, including two real rank-one LoRA tensor projections. Model execution, CLIP and file parsing in the integration harness use test doubles.

Beta limitation: full Z-Image GPU generation with real checkpoints, LoRAs and the running ComfyUI interface has not been validated here. Native regional hooks typically require two model predictions per step, with a possible third for uncovered background. This is a different implementation from the Krea2 attention engine.