重返未来1999风格/Reverse: 1999 Style（on_z_image_turbo)（anime)触发词：R style使用技巧：1.使用ai-toolkit在Tongyi-MAI\Z-Image-Turbo上用90图训练4750步。在z_image_turbo_bf16.safetensors上面运行风格稳定。2.如果用非官方的z_image_turbo大模型运行此Lora，可能效果弱甚至无效。3.这个Lora会导致人物手部结构劣化。在z_image_turbo上训练的人物远旅Lora手部正确率比无lora有所下降，但多次生成很容易找到手部正确的图。（在z_image_turbo上训练的Lora不一定手部结构劣化）这个lora的训练图大多都是手部良好的官图。在illustrious上训练的1999风格Lora也多次生成很容易找到手部正确的图。（同样的训练图在illustrious训练可以轻微手部结构劣化）所以我暂时难以确定问题出在哪里，4.如何解决结构劣化：提示词中说明手部动作，如比耶，放在脸上。可以提高正确率。但无法根治。最好不要生成手。5.人物还原不如我的illustrious1999风格Lora，大多只能还原个大概。 6.部分提示词是我在即梦主页随机复制的。English:Trigger Word: R StyleUsage Tips：1.Use ai-toolkit to train with 90 images for 4750 steps on Tongyi-MAI\Z-Image-Turbo. Running on z_image_turbo_bf16.safetensors ensures style stability.2.If you run this LoRA with a non-official z_image_turbo base model, it may result in weak or even ineffective performance.3.This LoRA may cause hand structure degradation of characters.The hand accuracy of the Yuanlü LoRA trained on z_image_turbo is lower than that without LoRA, but it is easy to generate images with correct hand structures after multiple attempts. (LoRAs trained on z_image_turbo do not necessarily lead to hand structure degradation)Most of the training images for this LoRA are official images with well-rendered hands.The 1999-style LoRA trained on illustrious also allows easy generation of images with correct hand structures after multiple attempts. (Training the same set of images on illustrious may cause slight hand structure degradation)Therefore, I cannot temporarily identify the root cause of the problem.<a target="_blank" rel="ugc" href="http://4.How">4.How</a> to resolve structure degradation: Specify hand movements in the prompt (e.g., "making a peace sign", "resting on the face")—this can improve accuracy but cannot completely eliminate the issue. It is best to avoid generating hands altogether.5.The character restoration effect is not as good as my illustrious 1999-style LoRA; it can mostly only capture a general resemblance of the characters.6.Some of the prompts were randomly copied from my Jimeng homepage.

Files with SHA256: b7b07761ac14fc2f1de2c5100823440301d6a0b333a9858fd8e24acd8cc1c707