Generate images and video with vLLM-Omni on SageMaker AI – Part 2
AWS published Part 2 of a series on running generative media models with vLLM-Omni on Amazon SageMaker AI. The walkthrough deploys two endpoints from the same pinned AWS vLLM-Omni Deep Learning Container (DLC): a real-time endpoint for FLUX.2-klein-4B image generation and an asynchronous endpoint for Wan2.1-VACE-1.3B video generation. The deployment script changes SM_VLLM_MODEL to load each model. The script pins omni-sagemaker-cuda-v1.6 and writes endpoint resource names to a state file.
In the described workflow, a text prompt goes to the real-time endpoint's /v1/images/generations route, which returns an OpenAI-compatible JSON response containing a base64-encoded PNG. The sample resizes that image to the video dimensions (480x320), encodes it as JPEG, and builds a multipart request for the Wan VACE endpoint's /v1/videos/sync route. Amazon S3 stores the multipart video request and the generated MP4; the asynchronous endpoint returns output and failure locations that the script polls. The sample uses InvokeEndpointAsync with InputLocation rather than inline Body because the multipart request exceeds the documented 128,000-byte inline limit. Default video settings are 17 frames and 30 diffusion steps; the post notes four steps is suitable only for a smoke test because it did not preserve source composition in the authors' review.
A sample command-line workflow and an optional Streamlit application are included. A single validation run in US East (N. Virginia) reported the ml.g6.xlarge image endpoint reaching InService in 9 minutes 30 seconds, the ml.g6e.xlarge video endpoint in 8 minutes 40 seconds, a warm image invocation returning in 4.7 seconds, 8.9 seconds of Wan VACE model latency, and the CLI retrieving the MP4 in 11.8 seconds. The post states these single-run values are reproduction checkpoints, not benchmarks. Cleanup is via a cleanup script; the post notes running GPU endpoints continue to incur charges.
Based on reporting from the original publisher. Visit the source for full context and later updates.
Publisher excerpt
Deploy two generative media models from one AWS vLLM-Omni Deep Learning Container on Amazon SageMaker AI. Generate an image with FLUX.2-klein through real-time inference, then animate it into video with Wan2.1-VACE through asynchronous inference, and retrieve the MP4 from Amazon S3.