[Kolors] OpenKolors v2.4 Multiple style general kolors model

V2.4 update

OpenKolors是一个多风格通用大模型，使用快手的可图模型作为基底进行微调，训练集约10万张图片，多轮多分辨率微调，过程中使用了blip rewards对模型进行图文匹配训练，能更准确的遵从提示词。

追加更多的图片训练，运用blip反馈训练优化图文配对，更好的遵从提示词。

多分辨率训练，尺度更广的分辨率生成，下至512到512，上到1344到1344

OpenKolors is a multiple style general model. It is using Kwai/kolors as a base model.

The total training set is around 100K images.

Trained with multiple resolution and around. In the training process, it used blip rewards to increase the prompt following.

In v2.4, it added more diverse training set and high quality images, down to 512 up to 1344 resolution.

V2.2 update

Trained on Chinese short text dataset to improve text ability

Merged multiple time to adjust the appearance and aesthetic

The default output might simplify the image, you could try to put vector art in negative prompt or put the details in first than main object.

Trained with style, natural language, tags to improve their ability.

Improved anime aesthetic

Improved prompt following.

Total training image on version 2 is around 100K.

It contains multiple round training to improve different ability.

Chinese Short Text Generation Prompt Example:

图片左侧有两行白色中文字，第一行写着“万圣”，第二行写着“快乐”。图片右侧有一个带着巫师帽子的南瓜头，眼睛发着黄色的光芒

V2.1 update

Multiple anime character including genshin, re:zero, naruto, etc

Improved realistic aesthetics

Vector art style

Improved anatomy(?)

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

该模型和OpenKolors-unet完全一致，只是加载形式变成了传统的sd格式

更新事项：

图真模型的后续版本将会更新在OpenKolors上，

相对于专注真实的图真模型，OpenKolors会更综合一些。

提供更广域的泛化性，涵盖动漫和真实方向。

对于真实系的生图，建议把“真实照片, raw photo”等相关tag前置，以及以自然语言描述画面，更好的触发相关权重

对于动漫的生图，建议把"anime artwork, niji style"等相关tag前置，以及可以使用自然语言+tag的形式来触发相关权重

使用方法:

下载模型，把模型放到comfyui/models/checkpoints里面

使用MZ的comfyui插件加载 https://github.com/MinusZoneAI/ComfyUI-Kolors-MZ

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 model is as same as OpenKolors-unet. It is just a more traditional way to load the model.

The follow up version of RealisticKolors which the alpha version is focus on realistic

but the following update of this model is more a general model which provide wide range of anime and realistic update.

For realistic generation, please put real photo or simular tags or description in the first place. The realistic generation is more pefer natual language description.

For anime generation, please put the anime artwork or general use tags style for generation.

Usage:

Download the model and place the model in checkpoints folder

Use https://github.com/MinusZoneAI/ComfyUI-Kolors-MZ this comfyui plugin to use it.