LTX-2.5 Slow-Motion Control: speed 0.025 to 1.0, 40x slower

The LTX-2.5 Slow-Motion Control LoRA adds a speed value from 0.025, which is 40 times slower, up to 1.0 for real time, while the clip still plays at 24 fps. Sume's video API has no speed or slow-motion field, so on Sume you steer slow motion only with the prompt, and the result is not a calibrated speed.

Specs are from the Slow-Motion Control card, read on 2026-10-02. Sume's side is from Video generation and Video filter.

How does the speed knob work?

The card says the LoRA decouples motion frame rate from playback frame rate, so physical motion speed becomes a controllable value while the output stays at its original fps. That is different from stretching a finished clip, which would show duplicated or blended frames. Here the model generates the motion as if shot on a high-speed camera, which is why splashes, cloth and similar actions are the examples.

• Input: one image for image-to-video, a normal caption, and a speed value from 0.025 (40x slow) to 1.0 (real time).
• Technical settings: 1920x1088, 121 frames at 24 fps, LoRA strength 1.0.
• Frame count must satisfy (frames - 1) % 8 == 0.
• Use identical prompts across different speed values so you can compare.
• The output includes audio generated from the prompt description.

How long is the result in real time?

A 121-frame clip at 24 fps lasts about five seconds on screen. At speed 0.025 those five seconds depict about an eighth of a real second of action, by simple arithmetic from the card's numbers. So the knob changes how much real motion fits into a fixed clip length, not the clip length.

What can Sume do for slow motion?

A Sume video request carries duration, resolution, aspect_ratio, frame_images, input_references, generate_audio and a few others. There is no speed parameter, and the docs say no v1 model accepts seed, so you cannot pin a result and re-run it at a new speed. You can write 'slow motion, high-speed camera' in the prompt and compare models, but the amount of slowing is up to the model.

Sume's Video filter route applies dim, crop or an allowlisted filtergraph to one clip and returns a new MP4. I found no retime operation described in its docs, so check the check endpoint before planning on one. Even then, stretching frames is not the same as generating high-speed motion.

Which route should you choose?

If a product shot needs a precise, repeatable slow-motion ratio, the LoRA is the tool, and you pay in setup and GPU time. If a clip only needs to feel slow, a hosted model with a clear prompt is quicker. See the earlier cinemagraph LoRA post for the sibling adapter. Licence is the LTX-2 community licence per the card.