z_image_turbo_nvfp4

Originally Posted: https://huggingface.co/Comfy-Org/z_image_turbo

🚀 Z-Image Turbo [NVFP4] – The Speed of Light Meets High-Fidelity Realism

Welcome to the next generation of efficiency. This is a specialized quantization of Z-Image Turbo using the cutting-edge NVFP4 (4-bit Floating Point) format. Optimized specifically for NVIDIA Blackwell (RTX 50-series) architecture, this model delivers unmatched inference speeds without sacrificing the photorealistic soul of the original weights.

Why NVFP4?

Unlike standard 4-bit quantizations, NVFP4 utilizes a dual-level scaling strategy (FP8 micro-blocks + FP32 tensor scaling). This minimizes quantization errors, preserving the fine textures, skin details, and lighting accuracy that define high-end diffusion models, all while slashing VRAM usage.

✨ Key Features

• Insane Speed: Generate 1.5MP to 2MP images in seconds.

• VRAM Efficient: Extremely lightweight footprint (approx. 4.2GB), making it accessible for GPUs with lower memory while flying on high-end cards.

• Zero-CFG Distilled: Designed for 8-10 step generations with no CFG overhead.

• Bilingual Mastery: Exceptional at rendering both English and Chinese text within the image, also Spanish

• Raw Realism: Maintains the "organic" look—perfect for those who hate the "AI plastic" aesthetic.

🛠 Recommended Settings

• Steps: 8 to 12 (10 is the sweet spot).

• CFG Scale: 1.0 (The model is distilled; higher values are not needed).

• Sampler: Euler a & bong tangent.

• Resolution: Optimized for 1024x1024 and 1536x1536.

• Architecture: Works best in environments supporting CUDA 13 and Forge UI.

Pro Tip: For the best results, use descriptive, natural language prompts. This model excels at "unposed" photography, cinematic lighting, and complex material textures like weathered skin or textile weaves.