I tested LTX2.5 a few days after it came out because I was curious.
LTX2.5 video
MINIMAX-H3 video
LTX-2.5 vs H3 comparison under the same conditions (832x480, 20 steps, seed 42)
LTX-2.5: 67 seconds, 121 frames, 1.04MB
H3: 294 seconds, 124 frames, 456KB
LTX-2.5 is 4.4 times faster
I initially asked the agent to test with the same prompt, but because the steps and bitrate were different, I measured again and added audio.
LTX2.5
MINIMAX-H3
LTX-2.5 vs H3 comparison including audio (832x480, 20 steps, seed 42)
First: LTX-2.5 — 168 seconds, 121 frames, audio AAC 48kHz
Second: H3 — 294 seconds, 124 frames, audio AAC 32kHz
LTX-2.5 is 1.75 times faster and has a higher audio sample rate
So far, LTX-2.5 seems to be faster.
This is based on DGX SPARK.
H3 has a resolution limit, but LTX2.5 doesn't seem to have one.
I tried creating a 16-second video with LTX2.5.
832x480, 385 frames, 24fps, 16.04 seconds
Generation time: 567 seconds (9 minutes 27 seconds)
VRAM peak: 39.9GB (out of 128GB)
Audio: AAC 48kHz stereo, 128kbps
File size: 2.92MB
No OOM
I made this one short in FHD.
LTX-2.5 FHD 5-second result:
1920x1088, 121 frames, 5.04 seconds
Generation time: 1142 seconds (19 minutes 2 seconds)
• Resolution: 1920x1088 (32x multiple)
• Frames: 121, 5.0 seconds
• Audio: AAC 48kHz ✅
• VRAM: 87GB / 128GB
• File size: 3.6MB
Overall comparison summary:
| Condition | Generation time | VRAM |
|------|-----------|------|
| 832x480 5 seconds | 168 seconds | ~40GB |
| 832x480 16 seconds | 567 seconds | ~40GB |
| 1920x1088 5 seconds | ~900 seconds (estimated) | 87GB |
FHD takes about 5 times longer because it has 4.5 times more pixels. A 10-second FHD video would be at high risk of OOM.
Practical combinations: 832x480 10-second generation → upscaling or FHD 5-second continuous generation are optimal.
My agent came to this conclusion, but upscaling makes it look like clay....
Conclusion: LTX2.5 is more useful than I thought.
