它是在 <a target="_blank" rel="ugc" href="http://FLUX.DEV">FLUX.DEV</a>完整模型上使用FP16的精度训练的，同时将clip-L模型换成了CLIP-GmP-ViT-L-14 模型，然后使用了joy_caption_two进行自然语言训练集打标。<a target="_blank" rel="ugc" href="http://这意味着你如果加载flux.dev">这意味着你如果加载flux.dev</a>完整模型全精度和T5-clip以及CLIP-GmP-ViT-L-14全精度进行生图时，可能需要35G显存。我同样没有设置触发词，你只需要直接了当的输入你想要的画面即可。或者你可以将我的示例图拖入comfyui，以查看我的提示词写法。需要注意的是：我的训练集中并没有包含任何NSFW画面，如果你需要绘制例如我的图集中图一和图二的裸体画面，你可能需要加载别的微调模型或者lora。训练集中包含的概念装束有：比基尼，皮裤，厚底高跟鞋，短裙，抹胸，人种以欧美为主，包含多人。It is trained on the <a target="_blank" rel="ugc" href="http://FLUX.DEV">FLUX.DEV</a> full model with FP16 precision, and the clip-L model is replaced with the CLIP-GmP-ViT-L-14 model, and then joy_caption_two is used for natural language training set labeling. This means that if you load the <a target="_blank" rel="ugc" href="http://flux.dev">flux.dev</a> full model with full precision and T5-clip and CLIP-GmP-ViT-L-14 with full precision for raw images, you may need 35G video memory.I also did not set trigger words, you just need to directly enter the picture you want. Or you can drag my example picture into comfyui to see how I write my prompt words.It should be noted that my training set does not contain any NSFW pictures. If you need to draw nude pictures such as Figure 1 and Figure 2 in my collection, you may need to load other fine-tuning models or lora.The concept costumes included in the training set are: bikini, leather pants, thick-soled high heels, short skirts, tube tops, and the races are mainly European and American, including multiple people.

Files with SHA256: 0c9cbc921a7309495e0d7b24dfae392e2807454a6b7b2d0a21f81d32c0c193ea