2026-04-19) Turbo V3.0 sometimes produces images with fine patterns, textures, or a slightly blurry/soft look. After testing and tweaking, this version has clearer image quality and better sense of depth and space compared to the previous one. NSFW performance remains at the same level.
In 3.0, the generated images have fine mesh-like stripes and a warmer color tone. However, these can be preserved after downloading. For complex XXOO actions/poses, the probability of the image breaking down or deforming is theoretically lower.
(2026-04-19)Turbo V3.0 生成图片，有些图片会有图案细纹和模糊感。经过我的测试和调试，这个版本会比之前的版本画质更清晰，更有空间层次感。NSFW方面保持相同水平。
3.0生成图片有细小的网状条纹，再就是色调偏暖。但是下载了可以保留下来，XXOO复杂动作，出图崩坏概率理论上会更低。
(2026-04-19) v3 Turbo is finally here after more than a month of training!
When I originally released the V3 Base, I actually wanted to release the Turbo version at the same time. However, the LoRA dataset was simply too large and might not have been compatible with the Turbo model.
The Base version still has the best image quality, but it’s quite slow to generate. Some people suggested using acceleration LoRAs, but I haven’t tried that myself.
If you’ve already downloaded the Easy Prompt pack from before, you can keep using your old prompts — they work exactly the same, no need to download it again.
I’ll update and add more AI example images whenever I have time.
Easy prompt for V3 download
Eason Zimageturbo workflow download
My setting:
（2026-04-19）v3 turbo经过一个多月的训练终于来了，原本发布V3 BASE的时候我是希望能同时发布Turbo，由于Lora的素材量太大可能无法被Turbo 大模型所兼容。Base版本的模型画质是最好的，但是出图速度慢，有人提倡可以使用加速LORA，但是我没有尝试过。为生成速度快和为了让Turbo版本能生成复杂的人物动作，之前下载了傻瓜提示词（Easy prompt）的朋友可以继续使用原有的提示词，是一样的，不必重复下载。我会有空更新一些AI例子图片。
傻瓜提示词 V3 download
Eason Zimageturbo workflow download
我的设置：
Easy promt for V3 download
For BASE only， Not compatible with the Z image turbo model
------------------------------------------------------------------------------------------------
Eason ZimageBASE workflow download
(3-7-2026) After nearly 5 weeks of hardcore, non-stop training, V3 is finally here! This beast clocked in at a mind-blowing 2.36 million steps — yeah, you read that right.
Compared to V2, which had all sorts of sexy outfits, lingerie, and heavy NSFW gear, V3 strips all that away. No more clothes cluttering things up. I also ditched every single AI-generated image from the training data to push for maximum realism. The focus is purely on authentic NSFW performance: genitals look way more lifelike and detailed, double-person sex scenes/positions feel stronger and more dynamic, and overall anatomy + motion just hits different.
https://civarchive.com/models/2304577
Of course, with super complex poses or crazy action, it can still break sometimes , but honestly? After tons of testing, I'm super happy with V3. It feels like it has huge potential — the NSFW output is insanely real, rich, and intense. Just check out the example images... you'll feel the raw power right away.
Download it, fire up some generations, and enjoy the ride! 🔥 Let me know what crazy stuff you create~
I also release V3 Lora Version, Important note: This LoRA is not compatible with Z-Image Turbo — throwing it on Turbo will usually result in total snowflake/noise artifacts or just broken garbage. Don't waste your time trying unless you're into abstract chaos 😂.
It's built specifically for Z-Image Base, where you can freely crank the LoRA strength anywhere from 0.7 to 1 (or even higher) for super flexible control over style, details, and that raw NSFW intensity.
(3-7-2026) 经过将近5个星期的超负荷训练，V3版本终于发布了，训练步数达惊人的236万步。它与V2不同的是，V2有各种情趣衣服和NSFW的内容，而V3则没有那些衣服，同时为了更真实移除了所有AI图的训练素材，更专注于NSFW方面的训练，比如阴部显得更加真实，双人做爱运动也有所加强，当然也有时候会因为动作复杂而崩图。但是经过我的测试，我对V3非常满意，我感觉它是非常有潜力的模型，因为她的NSFW的表现非常真实和丰富，当你看见例图你也能感受的到NSFW的强度。欢迎你下载和生成图片，enjoy it！
https://civarchive.com/models/2304577
同时，我也释放出炼V3的BASE Lora版本，注意，这个LORA与Turbo 模型并不兼容，在BASE上可以自由的调节LORA的强度, 但在turbo上生成是雪花
纸飞机群交流 https://t.me/+gTE-veRY4qxkMDQx
Eason ZimageBASE workflow download
傻瓜提示词 for V3 download
I saw a user's comment today saying that after using my checkpoint, some features become weak or stop working properly (like "pantyhose, small breasts").
This usually happens because when you merge a LoRA into a checkpoint, the merging process forces a fixed weight.
If the weight is set too low → the LoRA effect (e.g. pussy, nipples, detailed underwear etc.) becomes very subtle or almost invisible.
If the weight is set too high → it starts to override and damage the base model's original style/quality.
To solve this problem properly, I released this adjustable-weight version of the LoRA. You can now freely control the LoRA strength yourself — just lower it a bit if it's too strong, or increase it if you want the effect more obvious. Adjust it to whatever best matches your taste and prompt!
Hope this helps~ Feel free to experiment with different strengths (0.7–1. is usually a good range to play with) 😊
Why is this LoRA file so freaking huge? Well, because it was originally designed to be merged into a checkpoint — like a big base for blending. It was trained on 9,300 pictures and went through a crazy 1,040,000 steps.
That said, you can still totally use it standalone if that’s what you need for some specific reason. But if your VRAM is tight and you don’t wanna deal with loading this monster every time, just grab the merged checkpoint I already prepared — it’ll save you a ton of hassle.”
😄
https://civarchive.com/models/2304577/eason-zimageturbo-realistic-nsfw-v2
今天看见一个网友留言说使用了我的Checkpoint，有些功能会失效，比如panty stockings。这是因为merge checkpoint选择融合LORA的时候会强制选择权重，权重低的时候，LORA本身数据会不明显，比如pussy，nipples。当权重强时候，会影响到原始基座模型的数据，所以为了解决这个问题，推出了这个可以调节权重的LORA，你可以搭配你的需要来适度增减LORA的强度，达到你的要求。
为什么这个LORA的容量如此大，因为这个模型最初设定是给checkpoint作为融合之用，含有9300张图片，训练步数达到1040000步。但是它也能独立使用，取决于一些人的特殊需求，如果你的显存有限，可以直接下载我已经做好的merge checkpoint
https://civarchive.com/models/2304577/eason-zimageturbo-realistic-nsfw-v2