Back office ops · Production

UCLA OARC delivers near real-time generative AI image and 3D pipeline for immersive theater production Xanadu using AWS

The problem

UCLA's REMAP center needed AI microservices capable of handling at least 80 concurrent mobile phone users per performance with a mean round-trip time under 2 minutes, with no tolerance for graceful degradation during live theatrical performances.

Workflow diagram · grounded in source
1
Audience submits mobile sketch
trigger
“Every module begins with an audience prompt, in which participants are asked to draw a sketch for a specific task, such as creating a background, rendering a 2D representation of a 3D object, or placing muses in custom poses and garments”
2
Sketches ingested via Firebase and SQS
integration
“User sketches were passed to the microservice using a low-latency Firebase orchestration layer”
3
Lambda routes by pipeline type
routing
“they were sorted into sub queues by an AWS Lambda helper function. Each queue was responsible for starting a pipeline based on the type of inference processing that the user sketch required (for example, 2D-image, 3D-mesh)”
4
Lambda validates and orchestrates
validation
“This function did the validation, error/success messaging, concurrency handling, and orchestration of the pre-processing inference and post-processing steps”
5
Vision model generates text description
ai_action
“we used either a DeepSeek VLM loaded onto an Amazon SageMaker AI endpoint or Anthropic's Claude 3.5 Sonnet model through Amazon Bedrock”
6
Diffusion model generates image
ai_action
“these descriptions, user sketches, and supplemental assets were provided as inputs to a local diffusion model paired with a ControlNet or similar framework to generate the desired image”
7
Upscale to high resolution
ai_action
“These lower-quality images were passed into either Nova Canvas in Amazon Bedrock or Stable Diffusion 3.5 to rapidly generate higher-quality images, depending on the module”
8
Human-in-the-loop review
human_review
“Since there was a human-in-the-loop, we did not perform automated post-processing on the images. We could safely trust that issues would be caught before they were sent to the shrines”
9
Assets delivered to performance screens
output
“the resulting media re-projected back to the shrines as AI generated 2D images and 3D meshes in the show's digital scenery”
Reported outcome

OARC successfully delivered 7 live performances with about 500 total audience members co-creating media, achieving processing times of 40–60 seconds on g6.4xlarge instances and 20–30 seconds on g6.12xlarge instances.

Reported metrics
Minimum concurrent users requirement80 mobile phone users
Mean round-trip time targetunder 2 minutes
Total audience members across performancesabout 500
Concurrent co-creators per performanceup to 65
Show all 8 reported metrics
minimum concurrent users requirement80 mobile phone users
mean round-trip time targetunder 2 minutes
total audience members across performancesabout 500
concurrent co-creators per performanceup to 65
processing time on g6.4xlarge instances40-60 seconds
processing time on g6.12xlarge instances20-30 seconds
SageMaker AI share of total cloud spendapproximately 40%
performances delivered7
Reported stack
Amazon SageMaker AIAmazon BedrockAWS LambdaAmazon SQSAmazon SNSAmazon DynamoDBAmazon EFSAmazon S3Amazon EC2 G6Amazon EventBridgeAmazon CloudWatchAWS CodeBuildHuggingFaceClaude 3.5 SonnetDeepSeek VLMNova CanvasStable Diffusion 3.5SDXLSPAR3DControlNetInstantIDFirebasePyTorchUnreal EngineGitHub
Source
https://aws.amazon.com/blogs/machine-learning/university-of-california-los-angeles-delivers-an-immersive-theater-experience-with-aws-generative-ai-services?tag=soumet-20
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

OARC successfully delivered 7 live performances with about 500 total audience members co-creating media, achieving processing times of 40–60 seconds on g6.4xlarge instances and 20–30 seconds on g6.12xlarge instances.

What tools did this team use?

Amazon SageMaker AI, Amazon Bedrock, AWS Lambda, Amazon SQS, Amazon SNS, Amazon DynamoDB, Amazon EFS, Amazon S3, Amazon EC2 G6, Amazon EventBridge.

What results were reported?

Minimum concurrent users requirement: 80 mobile phone users; Mean round-trip time target: under 2 minutes; Total audience members across performances: about 500; Concurrent co-creators per performance: up to 65 (source-reported, not independently verified).

How is this back office ops AI workflow structured?

Audience submits mobile sketch → Sketches ingested via Firebase and SQS → Lambda routes by pipeline type → Lambda validates and orchestrates → Vision model generates text description → Diffusion model generates image → Upscale to high resolution → Human-in-the-loop review → Assets delivered to performance screens.

Related back office ops cases