WorldCrafter-Base by TencentARC: a 14.3B-parameter image-to-video model

On September 21, 2026, the model TencentARC/WorldCrafter-Base was published on HuggingFace with the image-to-video task. This is the base part of the WorldCrafter project: transformer weights, a camera adapter, and LoRA for the inference code. The repository contains 55.4 GB of weights in pth and safetensors formats, with the largest file at 9.3 GB. It runs via diffusers. The author released the code as open source but did not specify the model's license on the page.

What it means

The model has 14.3 billion parameters — noticeably heavier than what is already in our video generators section. For comparison: LTX-Video from Lightricks solves the same image-to-video task and, as described by the author, requires only 1 GB of VRAM, while its entire repository weighs 236.4 GB. For WorldCrafter-Base, the base part alone takes up 55.4 GB, and that's without counting the shared components from WorldCrafter-Fast — repencoder, text_encoder, tokenizer, vae, and scheduler, which live in a neighboring repository. TencentARC does not publish VRAM requirements. For comparison, Maestro from Blizaine lists a range of 12—24 GB, but that's a different class of tool — a full-fledged studio based on WanGP with image-text-to-video. CogVideoX from Zhipu AI handles text-to-video rather than image-to-video, and its weights take up 20 GB. So far the model has zero downloads, and the author provides neither hardware requirements nor license terms. Running it locally will require not only this repository but also WorldCrafter-Fast with the shared components — it is not a standalone build.