ComfyUI

# Follow Up H3 Video Gen Test:    Same RTX 5090. Same resolution. Much faster generation time.

Yesterday, my best eight-second H3 text-to-video shot took 9 minutes and 36 seconds on an RTX 5090.

Today, the fastest version took 1 minute and 19 seconds.

I didn’t lower the resolution or remove the audio. I changed the generation workflow.

The goal was to explore how far we could push generation speed and build a practical framework for what is possible on an RTX 5090. Visual and audio quality vary between these workflows. These are speed experiments, not a claim that every result looks or sounds equally good.

## Why I ran the follow-up

My original comparison covered the RTX 4090, RTX 5090, RTX 6000 Ada and NVIDIA DGX Spark. I wanted a practical answer: how long would I actually wait for a shot?

This follow-up focuses on one question: how much faster could the 5090 get with a different H3 recipe?

I rented another RTX 5090 with 32GB of VRAM and an enforced 62GB system-RAM limit—close to the 64GB home PC I’m considering. Yesterday’s comparison machine had a 60GB RAM allocation. These are home-PC proxies, not a benchmark of the exact prebuilt I might buy.

## What stayed the same

Every output in today’s test used:

- 1376×768 resolution

- Eight seconds: 192 frames at 24fps

- Native stereo audio

- The original bakery prompt, with the same reference image for image-to-video

- The same base diffusion model, text encoder, video VAE and audio VAE

The main comparisons used the same seed. One plain-Turbo repeat used another seed.

I measured the server’s complete generation time, including model loading, encoding, sampling, decoding and saving the video. Initial downloads and the time to transfer the clip back to my computer are excluded. Graph caching was disabled.

The machines also used the same ComfyUI revision. Host CPU and driver differences remain, so these are practical workflow comparisons—not a perfectly controlled measurement of each optimization in isolation.

## The numbers

Yesterday’s 20-step baseline: 9m36s text-to-video / 10m15s image-to-video.

4-step Turbo + SageAttention: 1m41s / 1m48s.

Above + Sol-Attn and fused modulation: 1m44s / 1m52s.

Above + Spectrum: 1m19s / 1m27s.

8-step Turbo + SageAttention: 2m32s text-to-video. Image-to-video was not tested.

The plain four-step Turbo text-to-video repeat finished in 1m33s.

Yesterday’s numbers are medians of three runs per mode. Today’s rows are individual runs, apart from that separately reported repeat. Treat them as an exploratory comparison, not a universal performance guarantee.

The fastest text-to-video run was about 7.3× faster than yesterday’s median. Image-to-video was about 7.1× faster.

## More optimizations didn’t always mean more speed

The biggest initial change was moving from the 20-step workflow to a matching four-step Turbo LoRA, together with SageAttention. That combination already brought the wait down to roughly a minute and a half.

Adding Sol-Attn and fused modulation didn’t beat plain Turbo in these short four-step tests. That’s a result for this particular schedule and configuration, not a verdict on every Sol setup. I tested a ComfyUI Sol-Attn implementation, not NVIDIA’s complete Sol Engine stack.

Adding Spectrum produced the fastest results. Spectrum uses approximations to skip some expensive computation. In this test, that bought more speed, but it also means there’s another reason to inspect the output carefully.

I also looked into VDN. I did not run a VDN benchmark in this round, so there’s no VDN performance claim in these results.

## Speed is the framework. Quality is the decision.

After watching the initial four-step Turbo clip, my reaction was:

> “Audio quality is worse but not that bad.”

Both visual and audio quality can vary as the recipe changes. The purpose of this round was to map the speed range we could reach, then use the clips to decide which tradeoffs make sense for a particular shot.

The output was still the same resolution, duration and frame rate, with native audio. But those matching specifications do not establish matching perceptual quality.

I generated an eight-step Turbo sample in 2m32s as another option to compare. I’m not claiming that it fixes the audio; that needs listening and viewing, not an assumption based on the step count. All eight saved clips passed full decoding and format checks. That establishes that the files work—not that their dialogue, motion or detail are equally good.

## Watch and listen to the test clips

These are the actual eight-second outputs. Compare the dialogue, motion and detail yourself. Times below are generation time, not playback length.

Example 1 — 4-step Turbo + SageAttention: 1m41s (text-to-video).

Example 2 — Turbo + Sol-Attn + Spectrum: 1m19s (text-to-video, fastest measured).

Example 3 — 8-step Turbo + SageAttention: 2m32s (text-to-video, quality comparison sample).

## What this changes 

A 5090’s usefulness for local video looks different when the wait for this shot is around one to two minutes instead of around ten.

For exploring shots and testing ideas, that shorter feedback loop could matter a lot. For a final dialogue shot, I may prefer a slower recipe if it delivers better sound or motion.

It doesn’t change the 5090’s 32GB VRAM limit, and it doesn’t replace the earlier tests on the other machines. It does change how I interpret the 5090 result: a hardware benchmark needs the workflow attached to it.

Before buying a local video workstation, I’d test the actual model, resolution, duration and audio requirements I plan to use—and then watch the output, not just the stopwatch.

For the same eight-second H3 shot, would you choose roughly 1m20s with more approximation, 1m30–40s with plain four-step Turbo, or a slower recipe for the quality you need?

No claim that these are the fastest possible settings or equal-quality results.

Tools used: H3 Turbo, ComfyUI Sol-Attn, Spectrum. (https://github.com/ModelTC/Minimax-H3-Turbo) (https://github.com/Saganaki22/ComfyUI-sol-attn) (https://github.com/xmarre/ComfyUI-Spectrum-MiniMax-H3)