MODEL FIELD NOTES · H3
Can H3 Run on an RTX 5090? A 32GB VRAM Benchmark
Yes, within a specific configuration. One RTX 5090 completed a roughly 15-second H3 clip at 352×640 in our test, but memory use was close to the card limit. This record explains what ran and what the result does not establish.
在黄果制片阅读中文版The tested configuration
This experiment ran on August 11, 2026. It used a pruned INT8 ConvRot H3 Ref2VA package, not an unchanged deployment of the official BF16 weights. Quantization packages and task variants are not interchangeable.
The reference was a 3.5-second clip downsampled to 352×640 at 24 fps. The final output contained 362 frames at the same resolution: 15.083 seconds. Sampling used 12 steps with res_multistep and the beta scheduler.
The recorded environment was an RTX 5090 with 32GB memory, PyTorch 2.7.1+cu128, CUDA 12.8 and ComfyUI 0.31.0. These identify the historical test; they are not a recommendation to downgrade a working installation.
Smoke tests and the final run
Client wall time and server execution time are reported separately. The two low-step checks tested whether the pipeline and output length worked; they are not quality benchmarks.
| Run | Output / steps | Server / client time | Peak VRAM |
|---|---|---|---|
| Short smoke | 2.333 s / 2 | 115.14 / 116.37 s | 31,098 MiB |
| Length check | 15.083 s / 1 | 113.07 / 114.47 s | 29,994 MiB |
| Final run | 15.083 s / 12 | 258.45 / 260.57 s | 31,658 MiB |
Why memory handling mattered
The pipeline needs a text encoder and audio/video VAEs in addition to the video model. A checkpoint size alone does not describe the memory needed across loading, sampling and decoding.
This run required ComfyUI to discard models between stages. The recorded flags were --reserve-vram 1.0 and --disable-smart-memory. They describe the tested environment, rather than a universal fix for every newer node or runtime.
Peak use in the final run left little room. More reference frames, larger outputs or extra reference inputs may cross the limit. This result does not establish that arbitrary 768p jobs fit in 32GB.
A practical order for checking your setup
Validate the weight format and node compatibility first. Then separate loading, sampling and decoding measurements so a slow first request does not get mistaken for a slow sampler.
- Record GPU capacity, runtime versions, weight variant, input dimensions, reference duration and sampling steps.
- Start with a neutral reference you have permission to use, a small output and a few steps. Verify that the media decodes completely.
- Increase output length before raising step count. Keep the resolution unchanged during that check.
- If the process runs out of memory, check encoder and VAE residency, then reduce the input. Repeating the same oversized request is not a stability test.
- Measure peak memory and full output time, and keep first-run timing separate from warm requests.
What finishing the job did not prove
The contemporary visual review described reasonably stable clothing, hair and background. It also found repeated small hand gestures and camera crops rather than a convincing extended action arc. The face stayed obscured, so facial identity retention could not be evaluated.
The final client time was approximately 17.3 times the output duration. This is one configuration, not a broad model comparison or an optimized current-runtime result. Original reference media is not included here; visual observations are attributed to the test record and are not independently inspectable from this article.
Official references
Use these references for model identity and the official setup. The measurements above come from our dated experiments, not a vendor performance guarantee.