Hedra says its AMD tuning cuts Krea 3 image edits to 3.6 seconds

The San Francisco startup says custom ROCm kernels and lower-precision math cut a 2K edit from about 22 seconds on the same MI355X hardware.

        By [Ryan Merket](https://runtimewire.com/author/ryan-merket)
        · Published 

Primary source: [X](https://x.com/hedra_labs/status/2110763001372708992?s=46)

Why it matters #

Hedra is pitching its infrastructure as a way to serve third-party models, with AMD inference optimizations as part of the offer. The sixfold speed claim remains company-reported.

Hedra added Krea 3, a reference-driven image generation and editing model, to its platform on October 15th. In a six-post thread on X, the startup said it can return a 2K edit in 3.6 seconds on AMD Instinct MI355X accelerators, down from about 22 seconds on the same hardware before it rebuilt its inference stack.