Wan2.1 or HunyuanVideo Poop Loras
Really big fan of your your work for Pony and IL, I was wondering if you had any interest in making a version for a text-to-video/image-to-video model. I'm currently looking into Wan2.1 which has only a few Loras so far, but it's already integrated with the diffusion-pipeline repo for Lora creation. I'm not sure how much it takes to create a good concept lora such as this and how much more effort would be needed to shift to a video related pipeline, so I imagine you would know better about how difficult it would be.
At a glance it might not be too difficult to actually do but my issue is that I do all my training locally and as far as I can tell training video LoRAs would require me to either get a much more expensive GPU or rent time on a GPU. Honestly I'm not interested in spending real money on this hobby right now. Thanks for the interest though.
hello @BlackHat404 I'm happy to train the Wan2.1 I2V Lora (I have the budget & enough tech skills to train the model with diffusion pipe on a H100 - 80 G Ubuntu machine), all I need is the training data (realistic videos, caption and Wan model config ie .toml file, like this https://github.com/tdrussell/diffusion-pipe/tree/main/examples), let me know if you're happy to collaborate and I do the training, thanks @2122zap fyi
hello @BlackHat404 I'm happy to train the Wan2.1 I2V Lora (I have the budget & enough tech skills to train the model with diffusion pipe on a H100 - 80 G Ubuntu machine), all I need is the training data (realistic videos, caption and Wan model config ie .toml file, like this https://github.com/tdrussell/diffusion-pipe/tree/main/examples), let me know if you're happy to collaborate and I do the training, thanks @2122zap fyi
Thank you for the offer but I prefer to keep everything under my control. Seeing as I would be figuring out a lot of this as I go I would prefer to be able to tweak and try things at any time. If you already have all the hardware ready then perhaps you should work on this since I have barely any idea how to do it and lack the hardware.