Scalable Pixel-Level Visual Regression Detection via On-Device MD5 Hashing of GPU Frame Buffers

Abstract

Visual regression testing for graphics-rendered user interfaces traditionally relies on golden image comparison, in which a captured frame buffer from the device under test is compared against a stored reference image captured during a known-good run. This approach scales poorly across device variants and display resolutions, because the storage cost of raw frame buffers grows with both the number of test cases and the number of supported display configurations. Storing one reference per test case per resolution per device variant in raw form quickly reaches gigabyte and double-digit gigabyte totals for a modest test suite. This paper presents a pixel validation infrastructure that addresses the storage problem by performing on-device MD5 hashing of GPU-rendered frame buffers. Instead of storing each reference as a raw image, the system stores a fixed-length hash. The size reduction is roughly six orders of magnitude. The reduction in per-reference storage from approximately eighteen gigabytes to approximately nineteen kilobytes is what makes pixel-exact regression detection across the production test matrix tractable. The system was deployed across more than sixty automated test cases spanning multiple device types and display resolutions, replacing a manual comparison process that did not scale. The paper describes the end-to-end design, including golden image lifecycle management, multi-resolution baseline handling, automated baseline regeneration on intentional UI changes, and a database-backed golden image service with object storage and REST APIs. It discusses the tradeoffs of cryptographic hashing as a validation signal compared with perceptual diffing, the failure modes that arise around antialiasing and timing-sensitive rendering, and the operational discipline required to keep golden references trusted over time.