Xelawk

I tried to change the pool_res to [10000, 5000, 1200, 1000], then this error occured ``` #training meshes = 1004 ---------- Network initialized ------------- [Network] Total number of parameters :...

And now, I trying to add a small float like 1e-3 to avoid zero problem, then the new problem happens, as below: ``` arch: mconvnet batch_size: 16 beta1: 0.9 checkpoints_dir:...

Okay, I see. But many thanks to your reply :)

> 还有，train.py是estimator的格式，finetune是常规的，无奈。 方便加QQ么，刚好我在踩坑，求助

> Following [the tutorial](https://github.com/lllyasviel/ControlNet/blob/main/docs/train.md#step-3---what-sd-model-do-you-want-to-control) I can successfully download SD, add ControlNet, and train it. > > Now, I want to continue training the [OpenPose model](https://huggingface.co/lllyasviel/sd-controlnet-openpose) using a different pose estimation...

> Do you guys use it through API? I encountered this error but don't remember clearly what was the reason But if you use API - there may be arguments...

> > > It has already been implemented. For example, you can first use functions ii2s.invert_images_in_W() and ii2s.invert_images_in_FS() to invert all of your images . When you do the inference,...

> @lllyasviel yes, this is an optimization of decreasing loading time when switching stable diffusion model. In my test, this would decrease several seconds. How did you implement the loading...

Looking forward to the training code+2

> Hi @mikedh > I'm getting a step closer! > > I succeeded to add the image to the shape, but because my surface coordinates are random, the texture it...