This checkpoint character is aesthetically pleasing. Utilizing a multimodal training method, it can influence all styles and is capable of handling SFW and NSFW tasks.

NSFW Unlock Status: Except for special actions and postures, and sexual minorities, all have been unlocked.

The ZIB version offers greater diversity without significantly slowing down generation speed. Last week, I spent the entire time training and testing LoRAs based on the ZIB version. However, there is currently a serious bug in ZIB LoRA training: under normal conditions, the model still doesn't converge even after 80,000 steps (the issue appears to stem from the 8-bit optimizer). Training LoHa is even more challenging, so I've opted to train a larger number of smaller LoRAs instead to iteratively improve this checkpoint.

Recommended Sampling Settings: Euler_ancestral + Beta, 30 steps, CFG 4

Online usage: TensorArt | TensorHub | SeaArt

The more detailed the prompt, the closer they will be to the effect you want. Please check the fluency of your prompt. ZIB supports negative prompt, so please make good use of that as well.

Thinking Z-Image Base isn't quite right? You might just need to try my fork! My BF16 build works like a charm and handles LoRAs beautifully. Every single example image and post of mine comes from this BF16 version.

I've only briefly tested FP8, so I can't say for sure how much quality is lost there. If your rig is up to it, definitely stick with BF16.

Have fun experimenting!

If you like my model, you can try my other Lora models, which you can find here: Tensor

Because some people download the files, change the names and display images, and re-upload them as their own models, new versions will no longer be released on Civitai. If you want V3 and later versions, please purchase them from TensorArt.

因为一些人下载后修改名称和展示图当成自己的模型从新上传，所以不会在Civitai发布新版本了，想要V3和后续版本的请到Tensorart付费购买吧。