Z-Anime in ComfyUI: model comparison and upscaling tutorial

This tutorial explores Z-Anime, a fine-tune of Z-Image for anime-style image generation in ComfyUI. The presenter describes it as Apache 2.0 licensed. The presenter says it can run on an 8 GB graphics card and produce images in four steps. These are the speaker's claims, rather than a hardware benchmark with a specified test system.

The workflow compares four-step and eight-step turbo versions with the base model. The four-step example uses CFG 1 at 1216 × 832; a base-model comparison uses 42 steps at CFG 4.5. The presenter prefers the eight-step version for its speed and prompt following in these examples. The base model also supports negative prompting, and the tutorial includes a 2048 × 2048 generation without upscaling.

Tests expose limits in counting, panel placement and requested details. ERNIE Image follows the illustrated two-panel prompt more closely in the comparison. Its prompt enhancer can improve spatial layout, but it also changes the results: a later gritty detective prompt becomes neater than intended.

The local AI workflow covers image-to-image generation, inpainting and several upscaling methods. Separate model, prompt and VAE switches let Z-Anime generate an image that ERNIE Image then upscales, or reverse those roles. The examples show that upscaling can alter style and anatomy as well as sharpen detail, including changing a mount's paws into hooves.