Appearance
Compare MiniMax H3 and FastH3 video generation times on Olares One, Mac Studio and NVIDIA DGX Spark. Explore 480p and 720p outputs using the same prompts and inputs across machines.
MiniMax's open omni-modal video model: it reads text, images, video and audio as context and generates clips of up to 15 seconds with native stereo audio. Released as FL2VA (text, first and last frame) and Ref2VA (reference media).
The generator is large, so GPU memory decides which lengths stay fully on the card and stay fast.
Time to generate one video with audio. Each cell shows the time at 480p / 720p. Lower is better.
Scroll horizontally to compare all configurations.
| Machine | 5 s clip, first + last frame | 7.29 s clip, 1 reference image | ||||
|---|---|---|---|---|---|---|
| FastH3 v2 nvfp4480p / 720p | FastH3 v2 int8480p / 720p | H3 fl2va int8480p / 720p | FastH3 OmniRef nvfp4480p / 720p | FastH3 OmniRef int8480p / 720p | H3 ref2va int8480p / 720p | |
| Olares OneComfyUI, default memory mode | 43 s / 2.1 min | 51 s / 2.4 min | 52 s / 2.8 min | 71 s / 2.8 min | 88 s / 3.2 min | 58 s / 2.6 min |
| Mac Studio M5 MaxComfyUI, --gpu-only | 4.5 min / 15 min | 5.6 min / 16 min | Needs more than 64 GB | 7.7 min / 26 min | 8.5 min / 27 min | Needs more than 64 GB |
| DGX SparkComfyUI, default memory mode | 84 s / 4.2 min | 2.2 min / 5.9 min | 2.3 min / 6.6 min | 2.4 min / 5.4 min | 3.4 min / 7.4 min | 3.4 min / 5.7 min |
Each cell is how long you wait for one clip at 480p / 720p, from pressing run to getting the mp4 with sound. One run each, same seed (123456). The left three columns make a 5 s clip from a first and a last frame. The right three make a 7.29 s clip from one reference picture.
FastH3 and OmniRef are FastVideo's fast builds (8 steps, sparse attention). fl2va and ref2va are MiniMax's own release with a speed-up LoRA (8 and 4 steps). nvfp4 is a smaller format that RTX 50 GPUs run natively. The Mac has no nvfp4 hardware and emulates it. fl2va and ref2va are int8 only, because the only nvfp4 versions are third-party quants (the fl2va one breaks text-only prompts).
The Mac has 64 GB, and fl2va and ref2va use 69–78 GB there, so it swaps to disk and one 720p clip did not finish in 90 minutes. They are not run on the Mac. A dash means not run yet.
Estimate video generation time and VRAM requirements on Olares One for your model, resolution, video length and reference inputs.
Explore 12 test cases covering text prompts, keyframes and reference media. Select a case to compare video outputs and generation times across hardware configurations.
Each case compares MiniMax H3 with FastH3 using the same inputs and seed. FastH3 outputs use nvfp4 and int8, and the MiniMax release uses int8. The Mac (64 GB) skips the cells that need more memory than it has.
Text to video
Keyframes
Reference media
5 s480p
subject_definitions: <Subject 1> is an adult man in a white T-shirt with dark curly hair. summary: A 5-second realistic live-action single continuous shot. retention_analysis: Preserve the supplied identities and environments; keep all reference roles distinct. detailed_description: In a warmly lit kitchen, <Subject 1> stands beside a wooden counter. A steady medium shot. He looks at the camera, smiles, waves once and says <d>[English] Hello, it is good to see you.</d> He lowers his hand and holds a natural expression for the remaining time. No cuts or added dialogue. overall_soundscape: Natural footsteps, quiet ambience and clear synchronized speech. non_diegetic_music: None.
Scroll horizontally to compare all hardware.
| Configuration | Olares One | Mac Studio M5 Max | DGX Spark |
|---|---|---|---|
| FastH3 v2 nvfp4 | 39.1 s3.19 s/step, 23.2 GiB | 255.6 s21.41 s/step, 43.0 GiB | 73.5 s6.52 s/step, 30.2 GiB |
| FastH3 v2 int8 | 46.7 s4.08 s/step, 23.4 GiB | 325.5 s21.92 s/step, 50.0 GiB | 109.2 s10.97 s/step, 38.1 GiB |
| H3 fl2va int8 | 46.4 s4.05 s/step, 23.2 GiB | Needs more than 64 GB | 116.7 s11.32 s/step, 37.0 GiB |
s/step is the sampler time per step. One run per clip, seed 123456. The Mac runs dense attention and emulates nvfp4, so its frames differ from Olares One's even with the same seed. Memory peak is the whole GPU on Olares One (nvidia-smi), the ComfyUI process on the Mac, and system memory used on Spark (unified memory), so the figures are not the same measure.
Most prompts come from the Olares MiniMax H3 page. The kitchen, portrait and headphone prompts are our own.