Why can't video models re-render the same take at higher resolution? An LTX 2.5 complaint
iJohnnyRobert · reddit · 2026-11-04
The author's core gripe with LTX 2.5 (and likely H3 and other video models): after finding a great take via cheap low-res "blocking" passes, there's no way to regenerate that exact video at a higher base resolution — you just re-roll and hope. They argue for Blender-style deterministic fingerprints that reproduce identical video and audio at any base resolution, enabling a full 2h film to be blocked out at tiny resolutions then converted per-clip to 480p/720p and upscaled. They dismiss video-guided regeneration as slower, costlier and weaker than native fingerprint reuse.
More from Multimodal
- Dev uses Gemini 4 + Antigravity to auto-document a week of building into a BTS video — alexanderchen · 2026-11-05
- Ex-Google AR founder launches Atmee v2: real-time avatar model claims 50x speed, 100x cheaper — eptwts · 2026-11-05
- Creator shares Frozen + reverse Seedance video prompt combo — techhalla · 2026-11-05
- IV Studio: open-source local film studio on ComfyUI + MiniMax H3, idea to short film — itsVariance · 2026-11-05
- Running open source image/video models without a GPU: CozyGPU bills raw GPU time, cheaper than marked-up platforms — MissMoneyBalls · 2026-11-05
- From stills to a playable game-like world: a Weave creative workflow (plus discount code) — gizakdag · 2026-11-05