使用LLaVA和CogVLM进行打标，训练集使用了之前发布的各种材质以及特效LORA的训练集以及额外内容共约三万张，训练完成后使用cosineA轻微混合了0.2的leosamsHelloworldXL作为调剂，已合并VAE
LLaVA and CogVLM were used for marking, and the training set used a variety of previously released materials, as well as the special effects LORA training set and additional content of about 30,000 photos. After the training was completed, cosineA was used to slightly mix 0.2 leosamsHelloworldXL as a supplement, Baked VAE.
例图均未加载其他LORA或扩展插件，纯大模型提示词直出，如果效果不佳可以适当搭配之前发布的那些LORA，权重要适当降低
The example image no loaded with other LORA or extension plug-ins, Belongs to the large model text to image direct generation, If the effect is not good, you can appropriately use those previously released LORAs, and the weight of the appropriate reduction.
绘图参数
范围限制: 12 >= CFG >= 4
推荐参数: 9 >= CFG >= 7
CFG步数 >= 15
40 >= 推荐CFG步数>= 20
采样器: Euler a /DPM++ 2M Karras
Drawing parameter
Range constraint: 12 >= CFG >= 4
Recommend Range: 9 >= CFG >= 7
CFG STEP >= 15
40 >= Recommend CFG STEP >= 20
Sampler: Euler a /DPM++ 2M Karras
关于2D和3D的风格该模型也是支持的，但未来还会专门发布相关的模型
2D and 3D styles are also supported, and professional models will be released in the future
描述:
训练词语:
名称: quantumDemonXLRealV4_v40.safetensors
大小 (KB): 6938819
类型: Model
Pickle 扫描结果: Success
Pickle 扫描信息: No Pickle imports
病毒扫描结果: Success
