xAI's Ethan He applies Monte Carlo Tree Search at inference to prevent semantic drift in long video generation

@EthanHe_42 That’s an interesting idea

@IamEmily2050 It can actually be applied to all models. It’s a search algorithm wrapper on top.

@Rafa_Schwinger framepack is very close to what you describe https://lllyasviel.github.io/frame_pack_gitpage/

@EthanHe_42 It will be interesting to see this against Seedance V2 🤔🤔🤔

@EthanHe_42 increasing the resolution and stitching older frames (but smaller) from 2^n seconds ago (training a lora or similar on this) could be an interesting approach.

That way you force the state to be {x_i} with x_0 being now but locking correlation with the past history.

@EthanHe_42 Seedance V2 has so much potential left for improvement 🔥

@EthanHe_42 Sorry but I am Rafadomis

UNO REVERSE TIME MACHINE ACTIVATE

@EthanHe_42 MCTS on video generation is inherently a bad idea.

@EthanHe_42 This is a smart wrapper. Look ahead rollouts to pick the best continuation instead of just hoping the model doesnt drift. Could be huge for longer form video. Nice work.

@EthanHe_42 COOL

@EthanHe_42 GitHub / hf / modelscope?

mcts for video makes sense on paper but the search space is the part i want to understand better. video rollout evaluation is expensive and noisy. if the reward model is a learned judge, are you effectively training a second model to prefer the first model's outputs? what does the reward signal actually look like here

@EthanHe_42 Can it be extended to I2V wrapper?

@EthanHe_42 What is this? I thought you left?

@EthanHe_42 What model it will be? P.S. Since you not working at grok imagine, Imagine updates will be slower?

@EthanHe_42 Nice one, team. Stoked to see AlphaGo's approach helping keep long vids coherent. Any chance you can share a demo or some samples?

@EthanHe_42 That's what @ericjang11 was talking about on a recent podcast with @dwarkesh_sp that how much Alpha Go's algo is better and can be used into other creative works , and your video generation is just one example of it.