Z-Image
一个包含多个变体的AI图像生成模型家族，擅长写实人物生成，支持多语言文字生成，具备提示增强和推理能力
Timeline (last 90 days)
Image LoRA training enables fused AdamW by default on NVIDIA GPUs and introduces aspect-ratio buckets support; FP8 LoRA training has lower memory usage for FLUX and Z-Image on Ada, Hopper, and Blackwell GPUs.
Image models can maintain INT8 or FP8 precision when offloaded, and Z-Image generation time on a B200 with a 10 GiB VRAM limit dropped from 38.4 seconds to 2.3 seconds
Z-Image is an image generation model capable of producing cinematic, ultra-realistic ancient-style character images.
据B站UP主阿琪老师演示，Z-Image无需复杂部署环境，也不需要堆叠大量LoRA权重，新手可在三分钟内生成古风人物图像
使用Z-Image时K采样器推荐配置为：步数20步、CFG值1、降噪值1
演示中使用9:16竖版比例，适合手机端展示或短剧竖屏内容
All Facts (7)
新一代模型如FLUX、Z-Image由于架构改进，CFG值设为1甚至更低即可获得最佳效果，而传统Stable Diffusion模型通常使用7-12的CFG值
80%UnverifiedImage models can maintain INT8 or FP8 precision when offloaded, and Z-Image generation time on a B200 with a 10 GiB VRAM limit dropped from 38.4 seconds to 2.3 seconds
70%UnverifiedZ-Image is an image generation model capable of producing cinematic, ultra-realistic ancient-style character images.
60%UnverifiedImage LoRA training enables fused AdamW by default on NVIDIA GPUs and introduces aspect-ratio buckets support; FP8 LoRA training has lower memory usage for FLUX and Z-Image on Ada, Hopper, and Blackwell GPUs.
50%Unverified使用Z-Image时K采样器推荐配置为：步数20步、CFG值1、降噪值1
50%Unverified据B站UP主阿琪老师演示，Z-Image无需复杂部署环境，也不需要堆叠大量LoRA权重，新手可在三分钟内生成古风人物图像
50%Unverified演示中使用9:16竖版比例，适合手机端展示或短剧竖屏内容
50%