Hunyuan Video有非常棒的汉服和丰富的古风元素，但是在艺术意境的表现上略有不足，这个Lora正是为此而训练，希望会有人喜欢。
Hunyuan Video features excellent Hanfu and rich traditional Chinese elements, but its artistic expression is somewhat lacking. This LoRA was trained specifically to address this aspect, and I hope someone will enjoy it.
Description
FAQ
Comments (34)
画面很美！视频很稳！绝了！👍👍👍
是用的默认参数吗
大概练到多少步
@gx_ground136 降低了学习率，6195 steps
@fok3827 感谢 ，默认BS有4，7000步要挂一个晚上了 ~~
@gx_ground136 是训练了1天
这画风绝了
Very nice quality! great job.
Could you let me know the number of dataset images and training resolution?
It's about 270 images. I first trained for approximately 4000 steps at 512 resolution, then continued with an additional 1000 steps at 768 and 1024 resolutions respectively.
@fok3827 Wow that's a nice method. Is it possible to do that in diffusion-pipe? or are you using some other trainer? I've tried changing lr in the middle of the training with diffusion-pipe with saved checkpoints but It didn't seem to change lr. maybe i'm missing something.
@strongman69 Yep! I'm using diffusion-pipe.
@fok3827 I think I found the way. Thanks a lot.
我刚想说评论区不尊重作者 都是NSFW 一看原来是作者自己喜欢
训练和推理用的是h100吗。
示例图的工作流打开一看非常夸张，无fp8分辨率还是1280*720，我在32G的卡上试，爆了
大佬請教一下，這模型只能載入自身訓練的lora而不能直接搭配sd1.5 sdxl和flux的lora嗎
是的，它们是完全不同架构的模型，没有办法混用
希望大佬能夠訓練age slider之類的LORa，機器不好無法自行訓練😭,對了不知道您覺得此模型和LTX模型哪個比較好，感覺這個比較好
@zczcg 单纯比较最终品质的话，毫无疑问混元video是目前的开源第1，但是如果只比较推理速度，那么LTX无敌。至于slider，目前没有能训练hunyuanvideo slider lora的公开脚本（不公开的也没有），所以这恐怕超出了我的能力范围。
@fok3827 LTX好像只能做出制式的NSFW的樣子，我以為只要有訓練lora的program就可以無限生成各式各樣的LORA。
@fok3827 無論如何，謝謝大佬帶給我們這麼好的LORA
How did you manage to make the videos so smooth? Can you share your workflow?
感谢大佬带给我们这么高质量的lora，希望大佬多来点带工作流示例图，谢谢
大佬牛逼，可不可以求个工作流
Many thanks for this, now I need to learn more about Chinese clothing and Characters, can't wait to do some video's, just wish I knew more about Hanfu, but ill get there.
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.