How to Get the Most From LTX-2.5:
Prompts, Settings and Cost for AI Video

What LTX-2.5 is

LTX-2.5 is Lightricks' open-weights video model, released on August 11, 2026. It is a 22-billion-parameter audio-video diffusion transformer, so it generates the picture and a synchronized soundtrack together rather than adding sound afterwards. It works from text, from an image, or from an audio track, and the official model page lists what is new over LTX-2.3:

• A diffusion video decoder for sharper pixels and higher fidelity.
• Native multi-shot scenes: several connected shots in one generation, with characters, lighting, style and voice kept consistent across the cuts.
• Automatic duration: the model picks a length from what you describe.
• A better distilled model, output up to 4K, and five aspect ratios: 16:9, 9:16, 1:1, 3:2 and 2:3.

You can reach it four ways: the LTX API (ltx-2-5-fast and ltx-2-5-pro), hosted platforms such as fal, ComfyUI with Lightricks' node pack, or the open weights through Lightricks' own pipelines and diffusers. Limits differ by provider: on the LTX API both variants reach 4K, while fal lists its pro endpoints at up to 1080p. Check the provider you use.

Fast or pro

Pro spends more compute where the frame is busy, so it holds up better in crowded compositions and is the one to use for the final take. Fast is the drafting tool, and the only one that runs past ten seconds.

Prompting: write the shot, not the vibe

The most important thing to understand about LTX-2.5 is that whatever you leave out, it decides for you, including the soundtrack. Lightricks' prompting guide asks for six things:

Write it as a single flowing paragraph in the present tense, roughly four to eight sentences, and match the detail to the shot: a close-up needs more than a wide shot. An example that covers all six:

A medium close-up of a woman in her sixties in a flour-dusted apron, standing at a wooden counter in a small bakery before dawn. Warm tungsten light spills from an open oven behind her while blue light from the street window falls across her hands. She presses a thumb into a round of dough, tilts her head to listen, and the corner of her mouth lifts. The camera dollies in slowly until her hands fill the lower third of the frame. The oven fan hums, a radio murmurs low in the back room, and outside a delivery van rolls past on wet cobblestones. Without looking up she says, "Not yet. Ten more minutes."

Every sound in that paragraph belongs to something we can place: the oven, the back room, the street. Sounds that come from nowhere are where the audio drifts.

Multi-shot scenes

Multi-shot works best with two to four shots in one chronological paragraph. At every cut, name the transition (hard cut, match cut, dissolve), re-establish the framing and angle, re-identify recurring characters with the same words you first used, and say whether the sound continues:

A wide shot of a narrow harbor at dusk: fishing boats knock gently against the pier, gulls wheel overhead and a man in a yellow raincoat coils rope at the far end. Hard cut to a close-up of the same man in the yellow raincoat, eye level, his weathered hands tightening a knot as the harbor bell rings twice. The gulls and the water slapping the hull continue under the cut. He glances up toward the sea and says quietly, "Storm's early."

Image-to-video

When you start from an image, the picture is already decided. Spend the prompt on motion, action and sound, name the framing and light that should stay, and say what must not move. If you also supply an end frame, keep its dimensions and camera position identical and describe the change in the order it happens. Lightricks suggests a single continuous take, not multi-shot, for image-to-video.

Dialogue and pacing

• Stay at 24 or 25 fps for dialogue. 50 fps pulls the result toward a video look and weakens performances.
• Give speech a moment of stillness first. A line that starts on the first frame leaves lip sync nothing to lock onto.
• Write the beats. Automatic duration "won't stretch a moment or add a pause you didn't prompt for", so write "she pauses" or "a beat of silence", or set the duration yourself.

What to avoid

• On-screen text and signage: spelling and frame-to-frame stability are unreliable. Keep any text short and check it, or add it in the edit.
• Chaotic physics: splashes, crowds colliding and tumbling objects are the weakest results.
• Mood without motion: "tense" or "dreamy" with nothing visible moving gives the model nothing to animate.
• Prompts written for other models: Lightricks recommends rewriting keyword-style prompts from other tools into a flowing paragraph.

The settings that matter

• Frame rate: 24 or 25 for anything cinematic or spoken; 48 or 50 for fast motion, sports or product spins.
• Duration: automatic ("duration": null on the API) for exploration; fixed when a beat has to land on time. Automatic cannot be combined with an end frame.
• Camera motion: the camera_motion option takes dolly_in, dolly_out, dolly_left, dolly_right, jib_up, jib_down, static, focus_shift. It overrides moves you describe in prose, so use one or the other. The parameter is handy when the same move repeats across a batch.
• Audio: sound is on by default. Set generate_audio to false when you will score the scene yourself.

Spend on the final render, not the drafts

Official LTX API prices, per second of output video:

The cost difference between tiers is what makes a two-pass workflow pay. Ten six-second drafts on fast at 720p cost $5.40. One ten-second final on pro in 4K costs $3.90. Rendering all ten drafts on pro in 4K would cost $23.40. Explore cheaply, keep the seed and prompt of the take you like, then render it once at the quality you ship. With automatic duration you pay for the length produced, though prepaid accounts hold credits for the maximum length (20 seconds on fast, 10 on pro) until the job finishes.

Running the open weights yourself

Lightricks lists a minimum of an NVIDIA GPU with 32GB of VRAM, 32GB of RAM and 100GB of disk, and recommends an A100 80GB or H100. The full download is about 66GB. Five things make it practical:

• Start with the distilled pipeline. It is the fastest path and the one Lightricks recommends as a starting point: a fixed eight steps at half resolution, then a 2× spatial upscale refined in four more.
• Move to the production pipeline for finals. Lightricks' Diffusion Fidelity Rendering pipeline adds a detailing pass and can double or quadruple the frame rate, at the cost of time and memory.
• Quantize and offload when memory is short. FP8 casting, CPU or disk offload, NVFP4 on Blackwell GPUs, and in ComfyUI an INT8 transformer and text encoder with the lighter conv VAE.
• Respect the shape rules: frame counts of 8k+1 (97, 121, 193) and width and height in multiples of 32.
• Use the prompt enhancer locally only. It helps short, rough prompts, but Lightricks says not to use it for direct API requests.

From Lightricks' pipeline and ComfyUI documentation. On the distilled model a negative prompt has little effect, because guidance runs at CFG 1.

A workflow that works

Our read

This is the production shift we wrote about in What If a Movie No Longer Needed to Cost Millions?: the cost of a shot is collapsing, and what remains scarce is the direction. A good LTX-2.5 prompt is direction written down.

Sources: Lightricks, LTX-2.5 model page, prompting guide, API pricing and system requirements; Lightricks/LTX-2 on GitHub (pipelines, optimization and license); fal, How to use LTX-2.5, August 18, 2026. Prices and limits are as of October 26, 2026 and vary by provider. Example prompts are our own.