Everlyn-1

Why we included this project

Everlyn-1 is an open research release for teams who want video generation built on an autoregressive foundation rather than the diffusion pipeline most models follow. The repo splits that goal into three parts, each with its own linked project you can inspect on its own: a Wasserstein-distance vector quantizer that stabilizes tokenization and improves codebook utilization, an efficient autoregressive backbone for joint image and video generation, and a decoding strategy called TAME that reduces hallucinated objects in multimodal LLMs. Because the pieces live separately, you can borrow just the one that fits your work, say the quantizer if you are training a video tokenizer or TAME if you are tuning an MLLM. It reads more like an architecture blueprint and a set of methods than a turnkey model you download and run. Researchers and engineers prototyping autoregressive video will get the most out of it, particularly where training stability and codebook utilization have been the stumbling blocks.