I have provided a FREE "Follow Me Tier" for those who want to be notified of my new posts. This is at no charge, and provides no other benefits other than the ability to follow me.

I had been using the MimicPC site to train my Hunyuan Video LORAs. However, after a while I added up what it was costing me, and it was costing quite a bit. So much in fact, that I decided it would be cheaper, long term, to just put the money into buying a 5090 with 32 GB of VRAM. So that is what I did. 32 GB still isn't enough to do videos conventionally, so I spent some time modifying the musubi tuner code to handle an image series in sequential order, which is the next best thing. In reality, all of these video trainers are already extracting the images from the video anyway. The only difference in what they are doing, is that they handle 65 frames in a single batch. I modified the code to be able to do 65 frames (or whatever number of frames) in order, one after the other in sub-batches. With the right resizing of the video frames, I can fit a sub-batch size of 3 (for a 128 dim/rank model) and simulate the 65 frame batch size. This seems to work pretty well. The Katie Holmes model was trained on the new board in that manner. So any contributions here will be used to offset the cost of the board, which was over $2600 with tax. As always, all contributions are appreciated.

I have been training Hunyuan Video LORAs using my own money for the past couple of weeks. I was not originally going to share them. However, a couple of people have contributed, so what I have decided to do is share them with contributing members only. They WILL NOT be made available to the free tier, nor to the public.