LTX 2.5 🔥🔥🔥 Opens source model on the way.... Community is saying .... Its not a Minimax H3 rival model but a better upgrade from earlier 2.3 variant

Coming Today 2:00 PM EST 👇

Image quality-wise seems way better than H3, to be honest, from what I saw in this demo.

Wow.... expect that. Will it be better than minimax H3?

I am already too used to ref2va, if LTX is not having this - then H3 is a clear winner

If H3 didn't have the lic restrictions for US BS, then I would download. I'm not giving my info to UnoHu... LTX reasonable arrangement...

It had already leaked they were gonna skip LTX 2.4 and go straight to 2.5.. but yeah, the Minimax H3 comparisons at this point are unavoidable.

definitely curious to see how this compares to my local minimax h3 pipeline

I think if LTX wants to win against the Minimax H3, it needs to work on lowering the requirements for hardware. Minimax H3 with INT8 + TUR LoRA can currently run stably on a 3060 8GB(Seen in other posts). I wonder if they can achieve the same result. My PC Rig: Ryzen 7 7800X3D | RTX 5070 | 32GB RAM | 2TB SSD 👇👇

you can see here what kind of shit comparison is this no omni refernce mention and face distorion is still a thing

Another wonderful news today

I'm tired of AI videos. No one I know likes them

Why isn’t this going viral?

man, still learning minimax h3😅😅😇

💯💯💯

That dirt bike looked miles better than my minimax tests. Ltx2.3 had big issue with vehicles and dynamic motion. This looks great.

Calling it an upgrade rather than a rival is a fair distinction.

Oh? Disappointed

I love my life 🙂

Right, the upgrade from 2.3 is substantial with the multi-shot consistency improvements especially. We actually went into the specifics here:

@Ludedwolf

Agreed: LTX 2.5 should be judged against 2.3 first. Cross-model comparisons only help when prompt, duration, resolution and retry budget are fixed. Source-linked H3 prompt/output pairs:

You can run ltx onna mac mini H3 requires a lot more compute. It's the best optimize model for local video generation.

With omni models getting open weights , this has no place.

They'll need to have *drastically* improved their prompt adherence, and at least give the option of a less compressed VAE, for anyone to care now that H3 has raised the bar so high.

Has anyone managed to generate video with 8 GB of VRAM?