been working on a bunch of upgrades - Seperable Causal Diffusion / Variational Flow Maps - 1 shot distillation / streaming

[johndpope]

https://wandb.ai/snoozie/vfm-v3b/runs/cgxdj1g6?nw=nwusersnoozie

https://docs.google.com/presentation/d/e/2PACX-1vRVuk1C9lVul_M7Wh_bAI2FRNbkQXmXlVxLMuO3GmqbJllUkWN5HRRiDn6eTOXSRNT8I-4i-SB-mJW6/pub?start=false&loop=false&delayms=15000

the work on SCD unlocks streaming - no upper limit on desktop gpu - can just keep spitting out video.

I did implement Ditto - but that paper is superceded with work from VFM
https://arxiv.org/abs/2603.07276

I was looking at DMD2 as opposed to matching 1 image / video to match the distribution - that seems ok with GAN but actually a paper dropped the other day SELF-E - from Adobe that transcends this with self evaluation - that is currently training.
https://wandb.ai/snoozie/vfm-v3b/runs/cgxdj1g6

all of this is running on 5090 - and the boss wants to cut my 20x claude account by end of this month..... so I'm just looking at power bill at the moment. need some sponsorshipt for compute

I have this training run - you can see video converging -
https://wandb.ai/snoozie/vfm-v3b/runs/kyubxb40

I need to run sanity checks - and if anyone wants to check my code - be my guest.
One of the losses VFM wasn't included in that run above - so im running it again now.

Code is here
https://github.com/johndpope/ltx2-castlehill