video-test版本：尝试使用静态图片练的视频lora，似乎过拟合了，使用其他的描述词效果不佳，后面试试视频炼制的效果。使用的训练脚本来自青龙圣者大佬：<a target="_blank" rel="ugc" href="https://civarchive.com/articles/10640/ez-hunyuanvideo-training-for-windows100-optimizerlycoris">https://civarchive.com/articles/10640/ez-hunyuanvideo-training-for-windows100-optimizerlycoris</a>强度建议0.7左右，到达1后很难出现动态效果-- v1.2版本，测试了一下flux，推荐强度1，没有添加wlop，后续版本加上试试--v1.1版本推荐使用 DPM++ 2M Karras | Euler a在原有模型的基础上，新增了200+的图片，wlop的图片占多数，所以风格会偏向wlop推荐强度：0.7--推荐采样方法：DPM++ 2M Karras | DPM++ 2M SDE Exponential放大算法：DAT x2数据集基于 宽 x 高：1024 * 768 进行的训练，所以这个效果应该是最好的描述就参考我上传的图片即可<hr />VIDEO-TEST VERSION: Tried to use the video lora practiced with static images, seems to be overfitting, using other descriptors doesn't work well, try video refining later.The training script used is from bdsqlsz: <a target="_blank" rel="ugc" href="https://civarchive.com/articles/10640/ez-hunyuanvideo-training-for-windows100-optimizerlycoris">https://civarchive.com/articles/10640/ez-hunyuanvideo-training-for-windows100-optimizerlycoris</a>Intensity is recommended to be around 0.7, it's hard to get dynamic effects after reaching 1--v1.2 version, tested flux, recommended strength 1, did not add the wlop picture, the subsequent version plus try--v1.1 recommended DPM++ 2M Karras | Euler aOn the basis of the original model, 200+ new images have been added, and wlop images account for the majority, so the style will be biased towards wlopRecommended strength: 0.7--Recommended Sampling Method: DPM++ 2M Karras | DPM++ 2M SDE ExponentialEnlargement Algorithm: DAT x2The dataset is trained on width x height: 1024 * 768, so this should be the best.Just refer to the image I uploaded for the description

厚涂特写 | ThickCloseUp | FLUX | XL | Hunyuan video

尝试使用flux进行训练，flux似乎确实对二次元的内容支持不太好，训练了3200步，四十张图，但loss只是到了0.214，但效果还不错。

Files with SHA256: 1f8e213be410b309b13bc15c77f18fb9ebeda5e8bf05cd4b0f255c68375ab464