0.1 Series Overview
Every post in the Synthetic Data Series generates a single static image. Real industrial inspection is often a video problem – radioscopy scan heads and conveyors moving parts past an X-ray detector, with real ambient mechanical sound alongside the picture. This series asks whether general-purpose joint video+audio generation models, applied to that real domain, actually stay in that domain – or quietly drift back to whatever they were trained on most.
Same ethos as the rest of this blog: real code, real captured output, honest limitations reported alongside anything that works.
0.1.1 Posts in this series
- Part 1 – X-ray Inspection Video + Sound Generation. A real GDXray weld radiograph, LTX-2 (a real, non-gated 19B joint video+audio diffusion transformer, int4-quantized to fit real memory constraints), and a real, honest failure mode: the model holds the X-ray domain faithfully for the one real conditioning frame, then drifts into ordinary color industrial photography for the rest of the generated motion.
- Part 2 – Video Conditioning vs. Single-Frame Conditioning. A reader question about a real LTX-2 feature – conditioning on many real frames instead of one – tested directly against Part 1. Real, measured result: color drift drops roughly 12-30x, but content drift (X-ray texture to industrial-machinery footage) persists, and audio gets noticeably quieter. A real partial fix, not a full one.
Reuse
Citation
BibTeX citation:
@online{goni2026,
author = {Goni, Hasan},
title = {Video \& {Sound} {Generation} {Series}},
date = {2026-09-02},
langid = {en}
}
For attribution, please cite this work as:
Goni, Hasan. 2026. “Video & Sound Generation Series.”
September 2.