# I'm getting unusable footage from H3 compared to LTX 2.5. What is going on ? — retained record

> Retained reddit record from reddit:StableDiffusion.

Canonical: https://highsignal.app/data/records/ded209369475e411

[← sources](https://highsignal.app/data)


source: [reddit](https://highsignal.app/data/reddit) (reddit:StableDiffusion)  published 2026-11-01 16:41 ingested 2026-11-02 02:40

[original source ↗](https://www.reddit.com/r/StableDiffusion/comments/1xl7pvz/im_getting_unusable_footage_from_h3_compared_to/)

## retained text



```
I was excited to finally try out this thing. I was expecting to see better results than LTX 2.5 at the same resolution. (I work at small resolutions). Well I finally managed to try it out and now I'm highly disappointed in it.

Not only is it about 3x slower than LTX 2.5 even at Q4 but it also seems to produce broken output at this resolution. I'm not sure if the model is so stupid that it cannot function properly at lower resolutions or if there's something else going on. Not to mention the quality of the sound is awful compared to LTX.

I also refuse to increase my resolution. I work on an M1 Air with only 8GBs of ram. If this model can only work with 480/720p+ then it's pretty much useless to me as it's way too slow at those resolutions and I won't wait around for hours for 5s of footage. So Im mainly interested in why it can't seem to do what LTX 2.5 seems to be doing with no issues at lower resolutions.

The video in my post compares three 5-second renders:

* **FastH3 Omni MLX INT4:** BF16 compute, 8 steps, 90% sparse attention, 8-layer NVFP4 decoder
* **The same H3 setup:** full attention instead of sparse attention
* **LTX-2.5 Q8:** 8 steps, Pruna decoder

All three use the same prompt and reference still at 256×192 and 24 fps.

In my first test H3 took **12m32s** with sparse attention and **10m36s** with full attention. LTX took **4m16s**. Both H3 outputs look broken and full attention didn’t fix the problem (it actually seems to have made it worse). I also decoded the same H3 result with its larger 36-layer decoder; the visual problems remained.

H3 uses the still as an OmniRef reference while LTX uses it as a middle-frame guide so their motion doesn't match frame for frame. The two H3 runs are the direct test of sparse versus full attention.

If anyone has any idea what the issue is with H3 do let me know. I'm curious for those of you which use H3, if you try rendering something at the same resolution on your side is the video output just as broken as mine ?
```



document fetched 2026-11-02 00:29 · canonical https://reddit.com/r/StableDiffusion/comments/1xl7pvz/im_getting_unusable_footage_from_h3_compared_to
