The Raw BF16 build is made for training, while the Turbo INT8 build is for direct image generation.

Raw BF16 版本专供训练使用，Turbo INT8 版本拿来直接出图。

Tuned for niji and illustration styles, it works great both as a LoRA training base and a standalone generation checkpoint. NSFW is fully unlocked, and I've fixed all the classic native Krea2 tag bugs — you know the drill: mixing up stockings and pantyhose, confusing lip-biting with tongue-out, and misrendering clothing protrusions.

主打 niji 风与插画质感，不管是训练 LoRA 还是直接生图都更好用。NSFW 已全面解锁，还修掉了 Krea2 原生一堆经典标签识别 bug—— 比如丝袜和连裤袜傻傻分不清、咬嘴唇和吐舌头搞混、衣物激凸识别错位之类的老毛病。

Funny enough, this whole model started as a happy accident. I was training a LoRA that just wouldn't converge, so I said screw it, scaled the dataset up to 3,000 images and ran it for 50k steps. It ended up completely reshaping Krea2's core characteristics.

Now it primarily outputs 2D, semi-realistic, illustration and 3D style artwork — which actually makes it a fantastic dedicated base for non-photorealistic work. Whether you're generating directly or training LoRAs, you'll get results much closer to the look and feel of Anima or Illustrious. It's a great foundation for non-realistic datasets, elevating your LoRA skills to a new level.

说起来这模型纯属意外产物 —— 当初练一个 LoRA 无法收敛，我干脆直接把训练集堆到 3000 张，跑了 5 万步，结果直接把 Krea2 的底子给改头换面了。

现在它基本只输出 2D、半写实、插画、3D 这类风格的内容，反倒成了个素质过硬的非写实专用基座。不管是直接生图还是拿来训练，出图质感都更贴近 Anima 或者 Illustrious 的表现。它是非写实类的一个好基座，能使你的lora上升到另一个高度。

Same routine as always: exclusive first launch on TensorArt. The v1 build will be available for download once it hits 10k uses, and all v2+ updates will remain TensorArt exclusives.

Thanks for all the support, everyone!

老规矩，TensorArt 平台首发。V1 版本使用量破 1 万后开放下载，V2 及后续版本会是 TensorArt 独家内容。谢谢大伙的支持！

Ray Krea2 Semi-real Ω - Raw v1 bf16 nsfw