Welcome to my 💫🌄 Friendly Z-Image-Turbo
✨ Less mess, more magic
Z-Image-Turbo - everyone already knows it's overkill!)
Z-Image is a 6B parameter efficient image generation foundation model with new Scalable Single-Stream DiT (S3-DiT) architecture where text, visual semantic tokens, and image VAE tokens are concatenated at the sequence level to serve as a unified input stream, maximizing parameter efficiency. It has a great understanding of prompts, styles and has a high attention to detail.
🚀 Z-Image-Turbo has a speed and quality balance advantage over other local models such as Flux 1, Flux 2, Qwen and others!
For more optimization use v1.1.2 (MultiGPU) version - the same as 1.1.3 but with MultiGPU optimizer. v1.1.3 is more stable but a little bit heavier.
Generate 2048 x 2048 images even with 6GB VRAM!
I offer my workflow, with which you can generate on a full BF16 model without sacrificing quality and at the same speed. This requires at least 6GB of video memory and at least 32GB of RAM. No Tritons or SageAttention! Simply configure the required amount of video memory dumped to RAM in the "VRAM Optimizer+" node.
The latest version of ComfyUI is required.
💻 System requirements:
Minimum system requirements:
RTX 3000-s, 6GB VRAM, 32GB+ RAM, 8-core processor, SSD, latest ComfyUI
📌 Detailed tips and links to models in the workflow
✨ Workflow features:
Extremely user-friendly interface
Maximum performance even on low-end systems with at least 6GB of VRAM
Enchance details parameter
Support for up to 3 Lora
Detailed tooltips, including recommended samplers for best quality
Manual random seed for complete control over generations
🤗🙏🏼 Thanks to Tongyi-MAI
Original repo — GitHub