Artifacts in high-frequency regions when generating videos

[aryanmikaeili]

Summary

Generating with LTX-2-19B (image-to-video), I see persistent flickering / shimmering on high-frequency textures (foliage, jungle background, distant trees). Smooth regions (sky, tarmac, bike body) look fine.
The same artifact reproduces in a pure VAE encode → decode roundtrip of the source video — no text encoder, no transformer, no diffusion, no upsampler. That points at the VAE decoder itself rather than the diffusion stage(s).

Reproduction

1. I2V generation with TI2VidOneStagePipeline at 768×448×121, seed 42, conditioning on the first frame of an input clip. Look at the jungle hills in the background — the foliage shimmers frame-to-frame even though the camera is locked-off and the source shows it as a stable texture:

video_6_decoded.mp4

2. Pure VAE encode → decode roundtrip of the source mp4 (resize → encode → decode, no transformer, no prompt). The shimmer is still there in exactly the same regions:

video_6_i2v.mp4

Question

Is this a known limitation of the LTX-2 VAE on dense organic textures, or does it look like something specific to my setup? Any recommended preprocessing or decoder settings to mitigate it? Happy to share the source clip and run any diagnostics.