# Native 3K MiniMax H3 on a 4090 thanks to this node: 3072×1728, 8 s, 25 steps in 2h21m, no more "illegal memory access" — retained record

> Retained reddit record from reddit:StableDiffusion.

Canonical: https://highsignal.app/data/records/34fdad8e6f111a0e

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


source: [reddit](https://highsignal.app/data/reddit) (reddit:StableDiffusion)  published 2026-10-29 18:27 ingested 2026-10-30 02:38

[original source ↗](https://www.reddit.com/r/StableDiffusion/comments/1xis0z3/native_3k_minimax_h3_on_a_4090_thanks_to_this/)

## retained text



```
Edit: Youtube link for full quaility: [https://youtu.be/E0SZIlegaEE](https://youtu.be/E0SZIlegaEE)

For weeks, anything above about 1080p × 5 s with MiniMax H3 in ComfyUI just died on my 4090. Usually it didn't even OOM. It crashed at step 0 with this:

    CUDA error: an illegal memory access was encountered
    ...
    aimdo memory compile error

After that ComfyUI looks alive (the web UI still responds), but the prompt worker is dead and nothing runs until you restart. `/free`, a fresh restart, lower batch: nothing helped. I'd pretty much accepted "720p, upscale later" and was eyeing a 5090.

Then I found **H3-Optimizations by Zironic**: [https://github.com/Zironic/H3-Optimizations](https://github.com/Zironic/H3-Optimizations)

You drop **H3 Memory Optimization** between the model chain and the guider/sampler and leave it at its defaults. It chunks the attention and MLP work so the temporary memory spikes go away. It's meant to keep the model's math the same; it isn't a quality hack. (The same pack has an optional Sparse Attention node that trades quality for speed. I did NOT use it for any of this.)

Same graph, same seed: without the node, 1920×1088 × 10 s crashes at step 0 every time. With it, it renders. So I kept pushing.

**Setup:** RTX 4090 24 GB, ComfyUI 0.43.0, SageAttention on, H3 ref2v hybrid checkpoint, 25 steps (turbo LoRA off), res\_multistep/simple, 192 frames (8 s @ 24 fps) with audio, 3 reference images. Only the canvas size changes between runs.

|Resolution|Megapixels|s / step|Total time (25 steps)|
|:-|:-|:-|:-|
|1920×1088|2.1|\~65|28 min|
|2336×1344|3.1|125|54 min|
|2560×1440|3.7|166|1h 13m|
|**3072×1728**|**5.3**|**333**|**2h 21m**|
|3840×2176|8.4|805|\~5.6 h (estimate; a 2-step probe ran fine, I didn't do the full render)|

VRAM sat at about 23.6 GB on every size with zero OOMs; the node offloads instead of falling over. So on a 4090 the limit isn't memory any more, only how long you're willing to wait. Time per step grows faster than pixel count (4× the pixels of 1080p is about 11× the time per step).

At 10 s / 1920×1088 I also got:

* 8-step turbo: 17 min
* 25-step: 40 min
* a 10 s Motion Context extend: 18 min, with a seamless join

Things I noticed:

* The model holds together well above its native resolution. The attached 3K clip lands all five prompted fight beats, with no doubled limbs or tiling. Armour plating, sparks and spray detail are a clear step up from 1080p.
* Changing resolution re-stages the shot (different blocking and camera), so you can't take a 720p take you like and "re-render it bigger". You have to sweep seeds at the target size.
* Native 1080p was noticeably steadier frame to frame than 720p + SeedVR2 upscaled to 1080p. SeedVR2 gives crisper micro-texture but more shimmer.
* I tried a DLSS5-Enhancer pass on the 3K clip. At this resolution it mostly just regraded it (flatter and cooler), so I'm posting the raw output.

Big thanks to Zironic. This saved me about $10k on a 5090. [Workflow ](https://github.com/chenthedesignmaker/ComfyUI-Work)(drag into ComfyUI). It's the exact 3K graph, with the Memory Optimization node wired in and a note on the timings. Swap in your own reference images and prompt.
```



document fetched 2026-10-30 00:32 · canonical https://reddit.com/r/StableDiffusion/comments/1xis0z3/native_3k_minimax_h3_on_a_4090_thanks_to_this
