← sources

Fastest MiniMax H3 768p workflow on RTX 5090 Laptop 24GB without sacrificing quality?

source: reddit (reddit:StableDiffusion)published 2026-08-29 18:06ingested 2026-08-30 02:38

retained text

Hi everyone,

I'm looking for the fastest MiniMax H3 workflow currently available for 768p generation, while keeping the best possible quality and prompt adherence.

My setup:

\- RTX 5090 Laptop – 24GB VRAM

\- ComfyUI

\- Target resolution: 768p

\- Mainly T2V / I2V, sometimes with reference video

\- Native 20-step H3 quality is my reference

I'm completely open to using a Turbo/Lightning/Distilled LoRA or fewer steps if it can actually preserve quality and prompt adherence close to the native 20-step model.

I've already tested some faster LoRA configurations, but so far I noticed a significant loss in quality compared to native H3. So I'm not specifically looking for a 20-step workflow — I'm looking for the fastest setup that doesn't noticeably sacrifice quality.

I'm interested in any current optimization, including:

\- Turbo / Lightning / distilled LoRAs

\- SageAttention

\- First Block Cache

\- SLA / Sparse Attention

\- Spectrum

\- TeaCache / EasyCache

\- Sol Attention

\- Quantized models (INT8, NVFP4, etc.)

\- CUDA / PyTorch optimizations

\- Any combination of these that works well

\- Any newer H3 optimization I might have missed

What is currently the fastest H3 setup you'd recommend for a 5090 Laptop 24GB while maintaining quality and prompt adherence as close as possible to native 20-step H3?

If a LoRA can achieve that in 8, 10, 12 steps, etc., I'm absolutely interested.

I'd especially love to see:

\- Workflow JSON

\- Exact settings / number of steps

\- Resolution and video duration

\- Generation time

\- Any quality trade-offs you've noticed

For comparison, 5 seconds at \~768p would be a useful benchmark.

Thanks!

document fetched 2026-08-30 00:30 · canonical https://reddit.com/r/StableDiffusion/comments/1w1ti0n/fastest_minimax_h3_768p_workflow_on_rtx_5090