AWS shows FLUX.2-klein-4B and Wan2.1-VACE-1.3B on SageMaker AI

This post is Part 2 of an AWS series on specialized Deep Learning Containers. Part 1 used vLLM-Omni and SageMaker AI bidirectional streaming for real-time speech. Part 2 moves to generative media: vLLM-Omni extends vLLM beyond text to models that process or generate text, audio, images, and video through OpenAI-compatible APIs. AWS packages tracked vLLM-Omni releases into a DLC and adds routing middleware for SageMaker AI, so the same container image can serve different model families.

The sample deploys one pinned DLC image to two endpoints and switches models by setting SM_VLLM_MODEL. The image endpoint runs FLUX.2-klein and returns a base64-encoded PNG inline, while the video endpoint runs Wan VACE and writes the MP4 to Amazon S3. AWS says it chose asynchronous inference for video because generation is longer-running and carries an image-conditioned multipart payload, a workload choice rather than a rule for every video model.

The reported timings from one US East (N. Virginia) Region run are reproduction checkpoints, not benchmarks, so they should not be read as expected performance. Whether teams adopt this pattern will likely hinge on validating GPU memory and performance per model and on instance availability for ml.g6.xlarge and ml.g6e.xlarge, since the video endpoint here stays on a fixed instance type.