Longcat Full Quality

• - 🏢 Meituan's open-source LongCat-Video, image-to-video full-quality variant.
• - 🆕 Single Diffusion Transformer handles text-to-video, image-to-video, video continuation.
• - 📏 Targets coherent, minutes-long sequences with consistent subjects.
• - 👁️ Outputs 720p at 30fps via coarse-to-fine generation.
• - 🧠 13.6B-parameter dense architecture using Block Sparse Attention.
• - ⚡ Distilled sibling variants are offered for faster inference.
• - 🔒 Weights released under the permissive MIT License.
• - 🎯 Pretrained on video-continuation for temporal coherence.

Meituan

Meituan is a Chinese technology company founded in 2010 by Wang Xing and headquartered in Beijing. Best known for its massive local services platform — spanning on-demand food delivery, consumer reviews, hotel bookings, and instant retail — Meituan listed on…

Longcat Full Quality is the image-to-video member of Meituan's LongCat-Video family, an open-source foundational video generator first detailed in the LongCat-Video Technical Report. It is built on a Diffusion Transformer (DiT) framework and, unusually, uses a single unified model to serve text-to-video, image-to-video, and video-continuation tasks. Per Meituan's report, pretraining on the video-continuation objective is what lets it sustain quality and temporal coherence across minutes-long clips, and it produces 720p, 30fps output using a coarse-to-fine strategy along both temporal and spatial axes.

Within the family, this "Full Quality" image-to-video model prioritizes output fidelity, while its companion Longcat Distilled applies step-distillation for faster generation. The same split exists for the text-conditioned variants: the Longcat Full Quality text-to-video model and the distilled Longcat Distilled text-to-video model. Choosing Full Quality means accepting longer inference time in exchange for higher visual quality versus the distilled path.

The underlying system is a 13.6B-parameter dense model that, per Meituan's technical report, employs Block Sparse Attention. The weights are distributed under the MIT License, which permits broad commercial and research use.

In the image-to-video configuration, the model takes a still input image and generates video output with consistent subjects, optionally guided by a text prompt.