← sources

Losing my mind trying to resolve Minimax H3 artifacts at top of frame

source: reddit (reddit:StableDiffusion)published 2026-09-16 18:21ingested 2026-09-17 02:37

retained text

I've spent a couple of days trying get clean H3 reference generations that do not have a couple of specific artifacts.

Most prevalent is a smearing/warping artifact at the top of the frame (see video).

I've been tearing my hair out trying to fix them. I must be missing something obvious or this is a known issue that I haven't found mentioned anywhere.

Anyone else had these issues?

Here's what I've tried without success...

1. Different sampler/schedulers. They appear with the stock res\_multistep/simple combo but I've tried others and the issue persists.
2. Different step counts (20 - 40).
3. Different resolutions. (from 768p up to 2.5MP, ensuring all divisible by 32).
4. Different models (pruned BF16 ref and fl models, smhfacct's hybrid models).
5. Different VAEs (bf16 & int8).
6. Switched to tiled VAE decode.
7. Different durations (5s - 15s).
8. Changing the number and size of reference images.
9. Disabling comfy kitchen and H3 SLA (no other attention or LoRAs in use).
10. Updated to latest PyTorch version.

ComfyUI is up to date as of today (issue was happening with older versions too).

The issue happens on my local machine (5090, python 3.13) and on a runpod (RTX Pro 6000, python 3.12), both running Linux cuda130 and pytorch  2.13.

https://reddit.com/link/1wi5acr/video/s302neox5xph1/player

I also have short horizontal lines (not grid patterns, as far as I can see) that appear in the top left quarter of of the frame intermittently when using res\_multistep / simple, across different resolutions, models and the other variables I listed above, although switching to er\_sde / beta does fix this but does impact quality a little and it seems odd to move away from the recommended.

document fetched 2026-09-17 00:32 · canonical https://reddit.com/r/StableDiffusion/comments/1wi5acr/losing_my_mind_trying_to_resolve_minimax_h3