Try size: 768 x 1024 or 768 x 1280Weights: 0.7~1.0 (v3 and v2)Weights: 2.0~3.0 (v1)Pictures from Z-Image TurboNote: The visual effect is close to that of GPT Image 2, so it can sometimes look dirty and the noise has not converged. Because "Sigmoid (Timestep Type)" and "High Noise (Timestep Bias)" are used during training, the randomness of the composition may be enhanced. Of course, it also has a sharp focus effect on street scenes. Increasing the lora weights will increase the frequency of the black shoulder bag, just like the realistic model in SD1.5, so it will have a messy and unpredictable visual feel, similar to the MajicMix Realistic model. Why do this? Because turbo's soft focus is particularly prevalent, and secondly, there is very little variation between different seeds. When generating images on the platform, please use both alpha and beta ver. Shift is a parameter that can adjust "High Noise"; a value of 9 is recommended. / 視覺效果是接近GPT Image 2的，所以有時會看起來髒兮兮的，噪點收斂不乾淨。因為訓練時採用"Sigmoid(Timestep Type)"跟"High Noise(Timestep Bias)"，所以可能加強了構圖的隨機性，當然它也對街景有銳焦的效果。當提高lora權重時，會增加黑色側背包出現的次數，就跟SD1.5的寫實模型情況是一樣的，所以會有點麥橘那種雜亂、不按理出牌的既視感。為何做這個？因為turbo的柔焦特別氾濫，其次是不同種子間的變化很少。Shift是能調整"High Noise"的參數，建議採用9。<img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/01e1110f-7ded-46a5-af46-9d5e4974ea68/original=true/01e1110f-7ded-46a5-af46-9d5e4974ea68.jpeg" /><a target="_blank" rel="ugc" href="https://huggingface.co/Tongyi-MAI/Z-Image">https://huggingface.co/Tongyi-MAI/Z-Image</a>

random slider / natural pose slider

I recommend using beta2(dark). / 建議用beta2(dark)。

randomslider_zimagedeturbo-beta.safetensors

beta(dark)

v3.0

Files with SHA256: b67770c77538d725aa14cd12969b59385c4c63faf733eca4e4fc7a95b47003bf