Type | |
Stats | 632 59 |
Reviews | (62) |
Published | Jan 9, 2024 |
Base Model | |
Training | Steps: 20 Epochs: 14 |
Usage Tips | Clip Skip: 2 |
Trigger Words | nagisa |
Hash | AutoV2 8F2261AD64 |
v1
This model was trained on screenshots taken from within the game. It exhibits high levels of similarity and generalization. To enhance the expressiveness and liveliness of the generated characters, it is recommended to use emotional and action descriptors in the prompt. Furthermore, descriptions from different perspectives might produce surprising results.
这是通过在游戏内截图训练出的模型，模型相似度和泛化性都非常之高，建议在prompt中多使用情绪和动作描述以提高表情和动作的生动性，同时不同视角的描述或许也会有惊喜。
In order to maximize the authenticity of the characters, after testing more than a dozen underlying models, it was surprisingly found that the GhostMix mixed model yielded the best results. Consequently, this model was selected for training.
为了最大化角色的还原程度，并且在试过了十几个底层模型之后，很奇妙的，我发现ghostmix这个混合模型效果居然最好，因此我决定采用这个模型进行训练。
It is recommended to use a weight between 0.3 and 0.8 for image generation.
建议使用0.3-0.8的权重进行图像生成
The more complex the prompts are, the less likely the image is to collapse
提示词越复杂，图像越不容易崩坏
v2
New version v2 is released!
The training iterations have been increased and the base model has been replaced to enhance generalization and reconstruction accuracy.Better results can also be achieved with fewer prompts.
增加了训练次数，更换了底模，使其泛化性和还原性更好，即使提示词很少，也能出现较好的效果