# How sparse is too sparse for H3? — retained record

> Retained reddit record from reddit:StableDiffusion.

Canonical: https://highsignal.app/data/records/64023e9690933719

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


source: [reddit](https://highsignal.app/data/reddit) (reddit:StableDiffusion)  published 2026-08-28 17:40 ingested 2026-08-29 02:37

[original source ↗](https://www.reddit.com/r/StableDiffusion/comments/1w0xb4k/how_sparse_is_too_sparse_for_h3/)

## retained text



```
So me and various other people have implemented their own Sparse Attention nodes and you can see many people argue about what % you should actually run these nodes on to maintain prompt adherence etc.

So to help come to the bottom of this, I extensively tested various settings.

[https://huggingface.co/datasets/Zironic/h3-attention-breakpoint-10s](https://huggingface.co/datasets/Zironic/h3-attention-breakpoint-10s)



My main conclusion is that videos only truly visually break below 10% retained attention but various noticable semantic changes can still happen all the way up to 50%. However just increasing the attention at the early steps get you most of that semantic consistency back. I got the best result when tapering the attention down in a ramp which is why I've now implemented that as Denser Early ramp in my own node. 

Sidenote: That's not even the SLA version of the lora that's trained on 15% attention. It's really surprisingly viable to use the LTX turbo 8 step lora and sparse attention at the same time.
```



document fetched 2026-08-29 00:32 · canonical https://reddit.com/r/StableDiffusion/comments/1w0xb4k/how_sparse_is_too_sparse_for_h3
