关于lada、jasna最新模型的测试

不兼容

训练模型需要更换的硬件和数据，加油

qinsese

真是个人才！

654550006

motbautroi 发表于 2026-5-16 02:12
I wrote a Secondary restore tool to upscale images using TGPlus's ESR-se. The results were excellent ...

The original model was trained at a resolution of 256, while I am currently working on a 512-resolution model. The main issue is that I don’t have access to the original author’s training parameters and workflow. Figuring everything out on my own will involve a lot of trial and error, and cost me considerable time on training and testing. By the way, have you released your secondary restoration tool yet?

motbautroi

654550006 发表于 2026-5-26 10:20
The original model was trained at a resolution of 256, while I am currently working on a 512-resol ...

I am using JAVPlayer to extract the mosaic region information file. My tool is written in Python, it crop images from the restored Lada video file, enhace the quality of those images using ESR-se_std_2x (PERFECT) and Topaz (increase PERFECT), and finally encodes them into a new video file.
Logically, I shouldn't spend time on the mosaic detection step using JAVPlayer, so I haven't released my tool yet.
I'm too lazy to write the mosaic detection code from Lada's source code.

djaafer

感谢楼主奉献

654550006

motbautroi 发表于 2026-5-26 13:05
I am using JAVPlayer to extract the mosaic region information file. My tool is written in Python, i ...

You focus on post-processing for video mosaic removal, while I'm working on optimizing the author's v1.2 model. So far, I haven't made any breakthroughs.

motbautroi

654550006 发表于 2026-5-27 11:44
You focus on post-processing for video mosaic removal, while I'm working on optimizing the author' ...

Post-processing still relies on the results from the basicVSRv1.2 model. Therefore, your solution will thoroughly process mosaic images, offering higher performance and saving time in video restoration.

17369515595

感谢分享，学习一下

654550006

从2月份来，我一直在训练新模型，到目前为止，新模型已接近1.2版本模型95%效果，各项指标都挺优秀，接下来就不断微调

654550006

lada作者stage1、stage2阶段(算上stage2.5、stage2.6微调)差不多跑了40-50W迭代，目前1.2应该是最好的模型。和lada作者相比，他的优势对代码很熟悉，他素材有5W场景；而我是一个代码小白，我的优势是有过类似模型5年训练的经验，素材虽然只有1.5W场景，但是我素材清晰度都是精选，硬件上应该比作者硬件好，最大的优势是作者训练的是256分辨率模型，我训练的是512分辨率模型。stage1阶段到30万迭代就很容易nan了，34W迭代后，我就转入了stage2阶段训练，测试stage2阶段15000迭代差不多到达1.2版模型95%水平，希望后面的训练结果能超过1.2模型。