niklaslenzFour video models with sound on one RTX 5090: MiniMax-H3, H3 with Pika-H3-Flex, Kandinsky 6 Pro and LTX-2.5
The same five prompts, seed 42, made into 5-second clips with sound by four video setups on one RTX 5090, timed from request to finished MP4. The page shows one original clip per setup, unedited, with its prompt and time.
Not ranked. Measured on the devices shown here only. No rating and no rank: this result has no rating by me yet.
RTX 5090 32 GB · MiniMax-H3 base · 20 steps
one RTX 5090 32 GB
RTX 5090 32 GB · MiniMax-H3 + Pika-H3-Flex · 6 steps
one RTX 5090 32 GB
RTX 5090 32 GB · Kandinsky 6.0 Pro-distill · 10 steps
one RTX 5090 32 GB
RTX 5090 32 GB · LTX-2.5 22B distilled · 8 + 5 steps
one RTX 5090 32 GB
The clips
Prompt
integrated_multimodal_description: A low tracking shot on a narrow gravel mountain road at golden hour. From the very first frame an orange rally car with a white roof is already sliding sideways through a tight hairpin at high speed, front wheels turned into the slide. The camera runs alongside just above the ground, so gravel and a big cloud of dust are thrown toward the lens. Pine trees and a steep drop with a sunlit valley are in the background, low sun backlights the dust. The car snaps straight out of the corner and accelerates away up the road, the dust cloud hanging behind it. Realistic car, believable suspension and wheel motion, motion blur on the gravel, one continuous five-second shot, no cuts, no text. Nobody speaks. overall_soundscape: A loud turbocharged engine at full revs from the first second with a crackling anti-lag pop on the corner exit, gravel hammering on the car body and underbody, tyres scrubbing loose stones, the engine note rising and fading as the car pulls away up the valley. No voices, no speech. non_diegetic_music: None.
Prompt
integrated_multimodal_description: A low tracking shot on a narrow gravel mountain road at golden hour. From the very first frame an orange rally car with a white roof is already sliding sideways through a tight hairpin at high speed, front wheels turned into the slide. The camera runs alongside just above the ground, so gravel and a big cloud of dust are thrown toward the lens. Pine trees and a steep drop with a sunlit valley are in the background, low sun backlights the dust. The car snaps straight out of the corner and accelerates away up the road, the dust cloud hanging behind it. Realistic car, believable suspension and wheel motion, motion blur on the gravel, one continuous five-second shot, no cuts, no text. Nobody speaks. overall_soundscape: A loud turbocharged engine at full revs from the first second with a crackling anti-lag pop on the corner exit, gravel hammering on the car body and underbody, tyres scrubbing loose stones, the engine note rising and fading as the car pulls away up the valley. No voices, no speech. non_diegetic_music: None.
Prompt
A low tracking shot on a narrow gravel mountain road at golden hour. From the very first frame an orange rally car with a white roof is already sliding sideways through a tight hairpin at high speed, front wheels turned into the slide. The camera runs alongside just above the ground, so gravel and a big cloud of dust are thrown toward the lens. Pine trees and a steep drop with a sunlit valley are in the background, low sun backlights the dust. The car snaps straight out of the corner and accelerates away up the road, the dust cloud hanging behind it. Realistic car, believable suspension and wheel motion, motion blur on the gravel, one continuous five-second shot, no cuts, no text. Nobody speaks. <AUDCAP>A loud turbocharged engine at full revs from the first second with a crackling anti-lag pop on the corner exit, gravel hammering on the car body and underbody, tyres scrubbing loose stones, the engine note rising and fading as the car pulls away up the valley. No voices, no speech.<ENDAUDCAP>
Prompt
A low tracking shot on a narrow gravel mountain road at golden hour. From the very first frame an orange rally car with a white roof is already sliding sideways through a tight hairpin at high speed, front wheels turned into the slide. The camera runs alongside just above the ground, so gravel and a big cloud of dust are thrown toward the lens. Pine trees and a steep drop with a sunlit valley are in the background, low sun backlights the dust. The car snaps straight out of the corner and accelerates away up the road, the dust cloud hanging behind it. Realistic car, believable suspension and wheel motion, motion blur on the gravel, one continuous five-second shot, no cuts, no text. Nobody speaks. Sound: A loud turbocharged engine at full revs from the first second with a crackling anti-lag pop on the corner exit, gravel hammering on the car body and underbody, tyres scrubbing loose stones, the engine note rising and fading as the car pulls away up the valley. No voices, no speech.
Median wall-clock time per 5-second clip with sound, 5 prompts, seed 42, one RTX 5090 32 GB. Different output sizes and different pipelines: this is a timing of each setup as run, not a quality or rank list.
| Setup | Pipeline | Output | Steps | Median time (5 clips) | Fastest to slowest |
|---|---|---|---|---|---|
| MiniMax-H3 base | ComfyUI 0.44.0 | 1344×768, 124 frames | 20 | 268.7 s | 268.3 to 271.3 s |
| MiniMax-H3 + Pika-H3-Flex | ComfyUI 0.44.0 | 1344×768, 124 frames | 6 | 86.2 s | 84.5 to 86.4 s |
| Kandinsky 6.0 Pro-distill 5s | own inference code, RTX 5090 preset | 864×480, 121 frames | 10 | 237.1 s | 236.6 to 238.4 s |
| LTX-2.5 22B distilled | ComfyUI 0.44.0, two stages | 1344×768, 121 frames | 8 + 5 | 27.8 s | 25.2 to 28.1 s |
How it was run
- One RTX 5090 32 GB, models warm (one warm-up clip per setup, seed 1, not counted; MiniMax-H3 base had none of its own, see notes), 5 prompts, seed 42, one run per prompt, median of 5.
- Time is the wall clock from request to finished MP4 with sound (ComfyUI: submit to saved file, Kandinsky: wall time of its own driver).
- MiniMax-H3 base: ComfyUI 0.44.0, pruned int8 convrot weights, 20 steps (res_multistep), 1344x768, 124 frames, 24 fps.
- MiniMax-H3 + Pika-H3-Flex: same graph plus the Pika-H3-Flex LoRA at 1.0, lcm sampler with the 6-step sigma list from the Pika template, 1344x768, 124 frames.
- Kandinsky 6.0 Pro-distill 5s: Kandinsky inference code with its RTX 5090 preset (not ComfyUI), 10 steps, 864x480, 121 frames, built-in prompt beautifier on.
- LTX-2.5 22B distilled: ComfyUI 0.44.0, our own earlier two-stage workflow (stage 1 at 672x384 with 8 steps, x2 latent upscale, stage 2 with 5 steps), 1344x768 output, 121 frames.
What to keep in mind
Measured on this device only (one RTX 5090 32 GB), not ranked. No other device was measured for this comparison.
What the clips show (by eye, one viewer, frames of the rally clip): the LTX-2.5 clip is the softest of the four. The car is heavily motion-blurred in its first second, cut off at the left frame edge in the first frame, and its number plate is not readable. Later frames show the car small on the road behind a dust trail. The H3 base, H3 + Pika-H3-Flex and Kandinsky clips show the car with clearer body lines and dust in the frames we looked at. LTX-2.5 renders stage 1 at 672×384 and upscales to 1344×768.
The four setups got the same scene text, but each in its own prompt format: three fields for MiniMax-H3 (also used with Pika-H3-Flex), an <AUDCAP> block for Kandinsky, a 'Sound:' sentence for LTX-2.5. Kandinsky's built-in beautifier rewrote its prompt before generation (the rewritten text is in the raw files).
The setups differ in size and pipeline: Kandinsky 6.0 delivered 864×480 with this preset, the others 1344×768. Kandinsky runs in its own inference code, the others in ComfyUI. What the Kandinsky driver's wall time includes is not documented in the raw file, and its separate pipeline load time was 3.4 s. The times are therefore not a per-pixel comparison.
MiniMax-H3 base had no warm-up run of its own (the Pika run before it loaded the same weights). Its clips P1 and P2 were timed in the first session, P3 to P5 in a second session after a pause. P3 took 271.3 s, the other four 268.3 to 268.8 s.
Pika-H3-Flex: its Hugging Face card says an end-to-end ComfyUI render was not verified for the upload and outputs can differ from the original inference runtime. We used the 1344×768 size from the Pika size list instead of the template default of about 0.98 MP.
Sound is in all four clips (AAC), but we did not listen to it, only loudness was measured. The rally prompt asks for no speech.
The test times one prompt set of five and does not measure prompt following, motion quality or sound quality; the clip shown is one of five.
License info (one line, not blocking): MiniMax-H3 Community License (also covers Pika-H3-Flex) lists excluded territories including the EU, UK, South Korea and USA (license text read 4 Nov 2026). Kandinsky 6.0 is MIT and LTX-2.5 uses the LTX-2.x Community License (model pages read 6 Nov 2026).
Exact command · All timings (summary of the 20 timed clips) · Per-clip job files (timings and ComfyUI graphs) · Resolution, frames and audio of the original clips · sha256 of the original clips · Kandinsky steps (driver log excerpt) · Prompts of the rally clip · Speech check of the fisherman clip (whisper-base) · All raw files