Is this still in active development or abundant
didn't saw any updates in a while
No more updates, Larry doesn't have enough compute to train further.
really?
Nah, it's still under active but pretty low priority development. The current version already works well at 8 steps and I don't have much ambition to push it further. 4 steps is the major challenge and what I'm working on right now — I have many internal versions but no good outcome yet. The major problem is a hard-to-resolve tension between macro composition/physics and micro texture/details: with a budget of only 4 denoise steps, no matter how I schedule the grid, these two sides just can't stop fighting each other. Compute is not the problem here — I have enough compute to do whatever's needed, but the recipe just isn't good enough to drive me to any satisfying results. It's an old problem with DMD2: you can't get infinitely better with simply more compute and more training steps; artifacts start showing up when you train too much. I'm struggling to find a scalable approach. I'd welcome anyone who's experienced with these problems and would like to work together on this.
Nah, it's still under active but pretty low priority development. The current version already works well at 8 steps and I don't have much ambition to push it further. 4 steps is the major challenge and what I'm working on right now — I have many internal versions but no good outcome yet. The major problem is a hard-to-resolve tension between macro composition/physics and micro texture/details: with a budget of only 4 denoise steps, no matter how I schedule the grid, these two sides just can't stop fighting each other. Compute is not the problem here — I have enough compute to do whatever's needed, but the recipe just isn't good enough to drive me to any satisfying results. It's an old problem with DMD2: you can't get infinitely better with simply more compute and more training steps; artifacts start showing up when you train too much. I'm struggling to find a scalable approach. I'd welcome anyone who's experienced with these problems and would like to work together on this.
Then may you make ref2video turbo Lora like v4 600 steps ema - it will be the most useful turbo Lora for ref2video model? Because I see the main problem of all 4 steps turbo Lora in wan 2.2 and minimax H3 - this is really too little steps for a video models to somehow generate a decent video
From my experience of using all the distilled models, SCFM seems to give the best result that are most close to the original model, and it is very cheap to train, please have a look at their github and paper, https://github.com/caitree/scfm
There is no point to stick to 4 steps. 8 or 10 steps is very fine.