WorldFoundry
English中文

Krea Realtime Video 14B

Krea Realtime Video 14B from Krea AI: realtime-video-generation and text-to-video; Runner parity pending.

Edit this pageGitHubMarkdown

Krea Realtime Video 14B is Krea AI's open realtime/streaming video generation model. Unlike the batch T2V models in this catalog, it targets low-latency streaming generation: text-to-video and video-to-video in a streaming regime. The in-tree runtime lives under synthesis/visual_generation/krea_realtime, and the route exposes num_blocks, resolution, fps, and seed controls. Alias krea-rt-14b resolves here.

No official teaser or method figure is checked into this docs tree.

Project page: krea.ai. Code: krea-ai/realtime-video.

Build your run

Pipeline
KreaRealtimeVideoPipelineworldfoundry.pipelines.video_official.pipeline_official_video
TaskInteractive Video interactive-video
Environmentworldfoundry-unified-cu128
DeviceCUDA 12.8
Runnerworldfoundry.pipeline
Loadingfrom_pretrained
Invocationnative_or_standard_invocation
worldfoundry-eval run \
  krea-realtime-video \
  --pipeline.task-profile interactive-video \
  --json

What it is for

Use krea-realtime-video for catalog text-to-video inference. The recipe exists so you can launch a recorded T2V profile without assembling an upstream demo.

Upstream describes it as a 14B autoregressive/streaming video model distilled for realtime block-wise generation (weights at krea/krea-realtime-video, apache-2.0). This repository records the block-based generation controls (num_blocks) and the prompt/video input contract; deeper internals are not recorded here.

Catalog tasks include realtime-video-generation, text-to-video, video-to-video, and streaming-video. Catalog aliases include krea-rt-14b. Documented modalities: text and video in, video out.

Recorded evaluation entry points include VBench (Baseline quality check for streaming T2V output).

Runtime contract

Optional prompt, negative_prompt, frames, fps, and height; artifacts krea-realtime-video.mp4.

CLI task profiles: interactive-video. WorldFoundry binds KreaRealtimeVideoPipeline.

Family status: Runner parity pending.

Recorded interactive-video defaults: fps=16, height=480, width=832, seed=42.

Realtime throughput expectations depend heavily on your GPU — the 14B scale wants datacenter-class hardware for true realtime rates.

Running in WorldFoundry

The catalog launch is worldfoundry-eval run krea-realtime-video --pipeline.task-profile interactive-video. The command builder above switches --pipeline.task-profile among recorded tasks and still emits worldfoundry-eval run.

It runs in the unified environment worldfoundry-unified-cu128 (Python 3.11, CUDA 12.8).

Recorded weight repositories include krea/krea-realtime-video. Prepare the environment and check local assets with bash scripts/inference/prepare_model_infer.sh krea-realtime-video. Add --download to fetch missing public checkpoints.

Recorded licenses include apache-2.0; status: Runner parity pending.