@pikenrover awesome. I'll give it a run once I do all the triton install nonsense again.

for future training - if you need a easy way to clip and crop your training data I made a little tool specifically for video dataset curation. https://github.com/Tr1dae/HunyClip hopefully it's useful in a way. I find it faster than using normal video editors.

install ComfyUI(not going to link this as it would be an external site and I am not sure what the policy is for linking to external sites, its on github and huggingface), get one of the models from https://civitai.com/models/1295569/wan-ai-wan21-video-model-safetensors-andkijai-workflow-included?modelVersionId=1463630 along with a clip and workflow(on the same page). comfyui is a bit complex, and I havent been able to find a newbie guide for it. I can run any of the WAN workflows as I am missing a node, that wont install from the commands to install missing nodes, and the only thing I can find is that when having that happen means I need to update, but it wont update as I get a java error when trying to run the command to install the requirements. it probably related to the fact I have two Java versions on my system, but I am worried that removing the older version will break the Automatic1111 webui, so I just kinda gave up for now. automatic1111's webui is much easier to work with for AI images, but it cant do the video

https://civitai.com/models/1317373/runpod-wan-21-img2video-template-comfyui

@pikenrover I did yes

My guess would be something weird with the vae or the filepaths. Sometimes when using someone elses workflow it will do those guesses for the right file name, but if you have something named the same, it won't pick the right one. I didn't download to try though

@kvah you say we can drag and drop videos into comfyui to get the workflow? Same as with images?
It doesn't work for me. ANy Tip?

Where do I get the wan nsfw base model

Same issue. I paired it with the general nsfw lora and got movement, but not getting results anywhere close to what's shown here.

Perhaps that's due to those lightx2v or lightning LORAs that let us generate at 4 steps (2 low and 2 high)? Then we need to increase lightx2v or lightning to, say, 1.5, then this lora to 1.2 or higher, in low and high noise.

This lora is just kind of bad for motion, it barely moves even in the provided images. Its not worth your time

I can't cite half of the resources I use. The filters are terribly implemented. Nothing ever shows up, and certain choices lock out others, even though they can clearly work together. Manually add it in the prompt section.... which is also doesn't work as it should, never reads the metadata I restore to my posts after editing.