Recommended settings推荐设置: re_sde + simple, 8-16 steps, 1 cfg V3 resolves the overly glossy, wet-texture rendering issue, refines NSFW generation with improved convergence and stability, and boasts far better prompt comprehension. It follows your prompts much more faithfully and tones down the hyper-photorealistic look of earlier versions. This release ships as an INT8 quantized checkpoint, delivering a 30% speed increase for local inference. With so many LoRA models already available in the ecosystem, we set out to build a base checkpoint that stays closer to the original model’s native behavior, offers more reliable prompt adherence, and produces cleaner, more consistent outputs. Enjoy creating! v3解决很湿的问题，收敛nsfw性能，更加懂你的提示词，不再那么写实。 而且v3是int8版本，在我本地运行速度上升30%。现在已经有很多lora了，我们只需要更接近原版和更听话更干净的检查点。 祝你们玩的开心。