Player not loading? Watch on YouTube
This comparison tests LTX 2.5, MiniMax H3 and Wan 2.2 through ComfyUI using shared prompts and reference images. The speaker compares text-to-video, image-to-video and first/last-frame generation in widescreen, vertical and square formats. These are local AI tests on an RTX 3090, using mostly default or recommended workflow settings.
In the examples discussed, MiniMax H3 often follows the requested action more closely. The speaker generally favors LTX 2.5 for fine detail, camera movement and generation speed, and points to its wind animation in background plants and trees. All three produce consistency errors, including disappearing cups, distorted hands and objects that merge or appear unexpectedly. He recommends image references over text alone for better visual quality.
The timing results need context: output resolutions and frame rates differ between models. Wan 2.2 uses the most VRAM in these tests. For one first/last-frame example, the speaker reports roughly five minutes for LTX and nearly 32 for MiniMax. Wan takes about 90 minutes when memory spills into system RAM, falling to about 24 minutes after he closes other applications and reoptimizes.
The speaker's comparison app supports synchronized playback, zooming and frame inspection. The supplied resources include prompts, reference images and workflows for repeating the tests.
