Tencent Launches DisCa To Rocket AI Video Speeds

By vramkickedin | | 2 min read
An hourglass tilted has instead of sand tiny glowing rectangular film frames growing rapidly through the narrow center.

DisCa is a new acceleration framework designed to speed up AI video generation models. It reduces processing time by 11.8 times without sacrificing visual clarity.

Tencent researchers who also made HY-World-2.0 and HY3-Preview, introduced the system to address the heavy computing demands that typically stall open source video tools. By replacing older caching shortcuts with a small neural predictor, the tool maintains stable output even when sampling steps are cut down significantly.

Model Size: varied GB & VRAM GPU: requirements vary

Accelerating local video generation

  • Deployable lightweight predictor uses minimal system memory.
  • Compatible with both text-to-video and image-to-video tasks.
  • Stabilizes compressed outputs during rapid frame rendering.
  • Includes ready-to-run shell scripts for quick environment setup.

Users managing custom media projects can deploy the framework on standard workstation setups to render longer sequences without upgrading hardware. The straightforward command structure allows operators to adjust speed parameters while keeping generation quality consistent across different workloads.

Developer observations and next steps

The team acknowledges that pushing the current architecture too far occasionally introduces minor frame jitter during playback. Achieving higher compression rates will require refined training methods to keep visual elements stable across rapid transitions.

"It might be worth exploring more techniques (like DMD and post-training etc..) to further improve the performance,"

noted the researchers in their project documentation. Upcoming work will likely target these stability gaps to support longer video outputs. Download the model checkpoints on Hugging Face or review the full research paper for implementation details.