Z-Image
Z-Image by Alibaba's Tongyi lab: a 6-billion-parameter image model; its Turbo version draws good pictures in a few steps on ordinary cards.
What it is
Z-Image (late 2025) is a compact image model that reads sentences in English and Chinese and renders text well. Z-Image Turbo is distilled for about 8 steps and runs on cards with 12 to 16 GB (less with fp8 or GGUF files). Its small size and speed made it a popular choice for newcomers.
Files and nodes
Load Diffusion Model with the Z-Image file; Load CLIP with qwen_3_4b; Load VAE with the Flux ae.safetensors; the empty latent of the official template.
Settings that work
About 1 megapixel (1024×1024); Turbo: about 8 to 9 steps, cfg 1 (so no negative prompt); the template's sampler and scheduler. Prompt: detailed natural sentences.
What fits
Only Z-Image LoRAs and control patches. Flux LoRAs do not work, although the VAE is shared with Flux.1.
How the app helps
Shown as Z-Image on model cards. The lesson Z-Image Turbo: a picture in eight steps builds its graph.