Case studies / Vision

• Google23 Oct 2026Vision

  Bringing foundational models to the front lines of TB screening

  Problem Remote clinics lack radiologists, causing many TB cases to be missed.

  Solution Created and fine‑tuned a foundation model that reads chest X‑rays, runs on low‑resource devices, and flags TB like an expert radiologist.

  • Multimodal
  • Fine-tuning
  • Transformers

  Read on research.googleDetailsSave
• Google21 Oct 2026Vision

  Empowering frontline maternal care with AI-assisted ultrasound

  Problem Many regions lack trained sonographers, leaving pregnant people without ultrasound access.

  Solution They built an AI model that reads blind‑sweep ultrasound scans from cheap portable probes and runs on mobile devices, letting frontline workers assess gestational age.

  • Multimodal
  • Real-time
  • Embeddings

  Read on research.googleDetailsSave
• Google29 Sep 2026Vision

  How Diffusion Controller unifies and simplifies AI image generation

  Problem Generated images often missed specific details or looked distorted when trying to follow exact prompts.

  Solution They added a small controller network that continuously guides the diffusion steps, improving prompt alignment while keeping image quality, even on closed models.

  90% win rate over baseline

  • Multimodal
  • Fine-tuning
  • Transformers

  Read on research.googleDetailsSave
• NVIDIA29 Sep 2026Vision

  Lower the Cost of Building and Running Visual AI Agents with NVIDIA VSS Blueprint 3.3

  Problem Building visual AI agents was expensive and hard to maintain, especially for video ingestion, processing, and search.

  Solution They released VSS Blueprint 3.3, adding a Build Vision Agent skill for rapid deployment and Adaptive Efficient Video Sampling that cuts VLM token use and GPU load.

  80% fewer tokens, 46% more concurrent streams

  • Cost
  • GPUs
  • RAG

  Read on developer.nvidia.comDetailsSave
• Pinterest10 Sep 2026Vision

  Building Pinterest’s VLM Serving Stack on NVIDIA Dynamo

  Problem Serving vision‑language models needed low‑latency, cost‑effective inference for Pinterest’s multimodal features.

  Solution Pinterest built a VLM serving stack on NVIDIA Blackwell GPUs and Dynamo, using distributed inference, KV cache offloading, and custom routing to handle image‑heavy requests efficiently.

  • GPUs
  • Latency
  • Cost

  Read on medium.comDetailsSave
• NVIDIA9 Sep 2026Vision

  When to Use Encode-Prefill-Decode Disaggregation to Accelerate Multimodal Model Serving

  Problem Vision encoding blocks multimodal requests, causing slow first-token latency and hurting throughput.

  Solution They separate the vision encoder from LLM prefill/ decode into independent workers (EPD disaggregation), letting each stage batch and scale alone, which speeds up token generation.

  Up to 5x faster TTFT and 7x faster end‑to‑end response

  • Latency
  • Real-time

  Research behind this, on AI Papers DailyFastAFD: When Does Attention-FFN Disaggregation Pay on Rack-Scale NVLink?

  Read on developer.nvidia.comDetailsSave
• Google1 Sep 2026Vision

  Mapping global methane emissions from space with deep learning

  Problem Users needed a fast way to spot methane leaks from space.

  Solution They built MAPL‑EMIT, a deep‑learning model that turns hyperspectral satellite images into detections, enhancements, and source estimates of methane plumes.

  High recall of 84%

  • Transformers

  Read on research.googleDetailsSave
• Yelp20 Aug 2026Vision

  Building Menu Vision: Real-Time Dish Recognition

  Problem Diners had to type each menu item to see photos, slowing decisions at the table.

  Solution Created an on‑device camera app that reads menu text, matches it to Yelp’s dish database, and instantly shows user photos using real‑time vision and fuzzy matching.

  • Multimodal
  • Real-time
  • Latency

  Read on engineeringblog.yelp.comDetailsSave
• Google16 Jun 2026Vision

  From pixels to planning: Earth AI for nature restoration

  Problem Users couldn’t see small hedgerows and woodlands on satellite maps, making it hard to plan nature restoration.

  Solution They trained a high‑resolution deep‑learning model on satellite imagery to detect and convert tiny ecological features into precise vector shapes that can be used for planning and carbon accounting.

  • Fine-tuning

  Read on research.googleDetailsSave
• Google22 Apr 2026Vision

  It's all about the angle: Your photos, re-composed

  Problem Users want to change a photo’s angle after taking it, but normal tools can’t add missing parts.

  Solution They built a system that first estimates a 3D scene and camera from a single image, then uses generative inpainting to create a new view from a different angle.

  • Multimodal

  Read on research.googleDetailsSave
• JJJioHotstar28 Mar 2026Vision

  AI Video Generation at Scale — Prompts Don’t Scale

  Problem Prompt‑based video creation drifts and becomes inconsistent when producing large volumes.

  Solution They created a production pipeline that replaces free‑form prompts with structured contracts, adds objective quality checks, and optimizes GPU use for consistent, scalable AI video output.

  • Multimodal
  • Cost
  • GPUs

  Read on blog.hotstar.comDetailsSave
• Shopify19 Mar 2026Vision

  Building a Magic Mirror: AI retail experiences with Remix

  Problem In‑store shoppers wanted a fun, personalized experience that felt more interactive than just browsing.

  Solution Created an AI‑powered mirror with a hidden camera that reads a shopper’s look, runs a vision model to match makeup shades, and instantly shows recommendations or discounts via a web‑based UI.

  • Multimodal
  • Prompting

  Read on shopify.engineeringDetailsSave
• Google17 Feb 2026Vision

  Teaching AI to read a map

  Problem MLLMs can’t reliably trace paths on maps, missing basic navigation rules.

  Solution They built a synthetic data pipeline that generates map images with start/end points and correct paths, training models to learn spatial reasoning.

  • Synthetic data
  • Prompting

  Read on research.googleDetailsSave