Pusa V1.0 Surpassing Wan-i2v-14b With $500 Training Cost By Vectorized Timestep Adaptation

a nice project to check: its a free open source vibe coding app and supports the new kimi k2 model, built entirely with radio huggingface.co/spaces/akhaliq…

Thanks for sharing, AK. Pusa V1.0 is truly an exceptional model, surprising us with unprecedented low cost. Our approach fine-tunes the Wan-T2V-14B model using only $500 and 4K training samples, yet outperforms Wan’s official Image-to-Video model on VBench-I2V with 10 inference

Excellent work taking Wan T2V to sota I2V on just 4k samples and $500 of compute The key is really the right data and care of the underlying model structure which we currently brute force There are some interesting things in the approach, check it out!

Training a language model with raw rewards collapses when every hard question hands back zero. Guided Hybrid Policy Optimization (GHPO) fixes that by dropping small hints only on those stubborn items. RL with verifiable rewards is a game of 1 for correct, 0 for wrong. If the

From early last year, our group has been thinking what’s the boundary between SFT and RL. Is SFT truly useless w.r.t developing reasoning ability? The intuition back then was no. And now, we give our new work that really reveals how SFT can benefit reasoning!! Please check our