Everlyn-Labs/Everlyn-1

Everlyn‑1 Overview

Everlyn‑1 is an open‑source research suite from Everlyn Labs that tackles three problems that are central to modern generative AI for video and multimodal language models:

Component	What it does	Why it matters
Distribution Matching for Vector Quantization (Wasserstein‑VQ)	Introduces a new way to train the codebook used in vector‑quantized video models. It matches the distribution of latent vectors to the codebook using the Wasserstein distance, which improves codebook utilization and reduces quantisation error.	More stable training and higher‑quality video generation because the discrete latent space is better‑behaved.
EfficientARV	An efficient autoregressive architecture that can generate both images and videos. It supports conditional tasks such as animation, in‑painting, out‑painting, video prediction and interpolation, and is designed to be plug‑and‑play with multimodal large language models.	Provides a single model that can handle many video‑generation tasks without the heavy compute cost of typical autoregressive video models.
ANTRP (Intervening Anchor Token)	A decoding‑time strategy for multimodal LLMs that reduces hallucinations. It adjusts the variance of query‑key attention parameters (the TAME mechanism) so that “anchor” tokens do not dominate the attention spectrum.	Lowers the frequency of fabricated objects or facts in multimodal outputs without adding extra inference latency.

How the pieces fit together

The three repositories are linked under the Everlyn‑1 umbrella. The first improves the discrete latent representation that the autoregressive backbone (EfficientARV) consumes. The second provides the backbone that can be conditioned on a variety of visual tasks. The third sits on top of any multimodal LLM that uses the same token stream, making the whole pipeline more reliable when the model is asked to describe or generate visual content.

Who might use this?

• Research labs exploring video generation or multimodal LLMs that need a stable VQ layer.
• Developers building applications such as video‑to‑text, interactive video editing, or AI‑driven animation who want a lighter‑weight autoregressive model.
• LLM teams looking for a plug‑in decoding strategy to cut down hallucinations in vision‑language tasks.

Getting started

Each component lives in its own GitHub repo (links provided in the README). The typical workflow is:

1. Train or fine‑tune the Wasserstein‑VQ codebook on your video dataset.
2. Use the trained codebook with EfficientARV to learn the autoregressive generation of video tokens.
3. When deploying a multimodal LLM, enable the ANTRP/TAME decoding hook to improve output fidelity.

All three projects are open‑source, and the main Everlyn‑1 page (everlyn.tech) provides further documentation, model checkpoints, and links to pre‑trained weights.

---

This summary is based solely on the information provided in the repository’s README.