H3 video model breaks at low resolutions, runs 3x slower than LTX 2.5 in user tests
iJohnnyRobert · reddit · 2026-11-02
A Reddit user tested the new H3 video model at small resolutions (256×192) and found it badly underperforms LTX 2.5:
- Speed: for a 5-second 24fps clip, H3 (FastH3 Omni MLX INT4, sparse attention) took 12m32s and 10m36s with full attention, while LTX-2.5 Q8 finished in 4m16s — roughly 3x faster
- Quality: both H3 outputs showed severe visual artifacts (see comparison image); switching to full attention made things worse, and re-decoding with the larger 36-layer decoder didn't fix it
- The author notes different reference mechanisms: H3 uses the still as an OmniRef reference, LTX as a middle-frame guide
- They're asking other H3 users to reproduce the issue
Takeaway: for low-resolution, low-compute video generation, LTX 2.5 is currently the far safer choice.
More from Multimodal
- AI generates a fully interactive 3D neighborhood plan: 84 homes, 7 housing types — NathanSRobinson · 2026-11-02
- AI-generated "Bob the Hanger" creepy product mock goes viral — Delahuntagram · 2026-11-02
- Retro cel-shaded AI art series: a detective and her rainbow roses in 80s anime style — LudovicCreator · 2026-11-02
- Higgsfield Genjutsu turns Ice Age's Sid into an uncannily realistic human — KookyClothes407 · 2026-11-02
- Krea2 q6k quant runs image generation on pure CPU in 5-10 minutes — Robert__Sinclair · 2026-11-02
- Seedance 25 spec ad 'First Light' shows near-photoreal slo-mo water VFX — Sigita-YlariStudi00 · 2026-11-02