注意，请使用把权重放到comfyui/models/checkpoints里面，

使用最新版的https://github.com/MinusZoneAI/ComfyUI-Kolors-MZ

用MinusZone - KolorsCheckpointLoaderSimple进行加载

这是一款基于 OpenKolors v1.3 训练的动漫特化模型，模型训练的图片采用 OpenKolors v1.3 生成的特定风格提示词。主题提示词来源于 SA1B 数据集，随机选取了 500 个男性提示词、500 个女性提示词以及 500 个其他提示词组成。

在完成图片生成后，使用快手的 MPS 评价模型对图片进行分类。根据 MPS_SCORE 的评分标准，分类结果如下：

• worst：MPS_SCORE 低于 25%

• bad：MPS_SCORE 在 25%~50% 之间

• good：MPS_SCORE 在 50%~75% 之间

• high：MPS_SCORE 在 75% 以上

该模型选取评分在 50% 以上的生成图片作为训练集，使用自行编写的 t2itrainer 进行训练。代码仓库可在以下链接查看：

GitHub - lrzjason/T2ITrainer

训练参数

• 模型类型：prodigy

• 学习率：1

• 训练周期：10

• 重复次数：4

• 秩：16

• 批次大小：4

• 验证比例：0.1

• 去偏差：是

• 最小信噪比：5

本模型的目的是验证模型自我迭代的能力，通过使用模型自身的图片生成能力来优化模型对提示词的响应。

Important!

Please place the weight in to comfyui/models/checkpoints

Update MZ plugin to the latest version https://github.com/MinusZoneAI/ComfyUI-Kolors-MZ

Use MinusZone - KolorsCheckpointLoaderSimple to load the weight

This is an anime-specialized model trained based on OpenKolors v1.3. The training images were generated using specific style prompts from OpenKolors v1.3. The theme prompts were sourced from the SA1B dataset, with a random selection of 500 male prompts, 500 female prompts, and 500 other prompts.

After generating the images, the MPS evaluation model from Kuaishou was used to classify the images. The classification based on MPS_SCORE is as follows:

• worst: MPS_SCORE below 25%

• bad: MPS_SCORE between 25% and 50%

• good: MPS_SCORE between 50% and 75%

• high: MPS_SCORE above 75%

The model selected images with a score above 50% for the training set, and the training was conducted using a self-developed tool, t2itrainer. The code repository can be found at the following link:

GitHub - lrzjason/T2ITrainer

Training Parameters

• Model Type: prodigy

• Learning Rate: 1

• Epochs: 10

• Repeats: 4

• Rank: 16

• Batch Size: 4

• Validation Ratio: 0.1

• Debiased: Yes

• Minimum SNR: 5

The purpose of this model is to validate the model's self-iterative capability, adjusting its response to prompts through its own image generation abilities.