A useful LTX 2 prompt guide should leave you with a shot you can evaluate, not a longer string of cinematic adjectives. Begin with something observable: a bird turns its head, a key rotates, the camera stays still. Then decide which parts of the source must survive.
This exercise uses a fictional brass wind-up bird with cobalt wings and a coral winding key. You can reuse the same source for several prompt variants and compare whether the intended action actually happens. The still is original editorial practice artwork. The examples below have not been rendered in LTX.
Choose the release and workflow before the prompt
LTX 2 names a model family, while the current official LTX 2.5 model card describes a newer release with audio-video capabilities and multi-shot support. It distributes separate components rather than one universal checkpoint. Check the model card and its linked license before choosing a local deployment.
Keep your provider, model identifier, input mode and requested settings beside the prompt. A local native pipeline, a ComfyUI workflow and a hosted API can expose different controls. A setting copied from one environment is not proof that another accepts it. For the wider deployment choices, read our open-source AI video guide.
| Starting material | First exercise | Keep separate |
|---|---|---|
| Text only | Describe the whole scene and one action | Resolution and duration controls |
| An accepted first frame | Describe motion and retained details | Actual input image versus prose |
| A planned sequence | Describe cuts and recurring identifiers | Single take versus several shots |
Check current LTX Fast and Pro API limits
The current LTX 2.5 API matrix lists text-to-video, image-to-video and audio-to-video for Fast and Pro, but not retake or extend. Both reach 4K. Fast offers up to 20 seconds at 720p or 1080p and 24 or 25 fps; its higher-resolution or higher-frame-rate combinations top out at 10 seconds. Pro lists 6, 8 or 10 seconds. Audio-to-video length follows its separate input-audio limits.
Automatic duration uses an explicit duration: null; omitting the field is different. It cannot be combined with last_frame_uri on image-to-video. These hosted API options are separate from the native frame planner below.
Move hosted requests to the current API flow
LTX’s migration notice says V1 generation endpoints stop after October 26, 2026 at 11:59 PM UTC. V2 returns a job ID for polling rather than a finished MP4 in the creation response. The upload endpoint stays at /v1/upload. Test the actual integration; a model name change alone does not migrate its response handling.
Build a prompt around observable motion
The official prompting guide recommends clear scene, action, camera and audio descriptions. It favors a flowing paragraph for a simple continuous shot. For current multi-shot prompting, describe the transition, new framing, recurring identity and audio continuity at each cut. Its image-to-video guidance prefers a continuous take unless a cut is intentional.
Our practical test is to underline the verbs. “Beautiful, cinematic, detailed, premium” contains no action to check. “The bird slowly turns its head toward the window, pauses, then returns to its starting pose” gives you three visible beats. Use a concrete ending so the review includes more than an attractive opening frame.
- Name the subject and the details that identify it.
- Describe the starting composition and physical setting.
- Write the action in the order it should occur.
- Choose a fixed camera or one deliberate movement.
- Specify the sound and whether speech is present.
- Remove competing actions before the first test.
Text-to-video example for a single take
A locked medium close-up frames one brass wind-up songbird on a pale wooden workbench beside a rain-covered workshop window. Its cobalt enamel wings are folded and a small coral-red winding key sits on its visible side. The bird slowly turns its head toward the window, pauses, then faces its starting direction again. Both feet remain planted on the same spot. Warm lamplight catches the brass while cool window light outlines the wings. A quiet mechanical tick accompanies the head turn, with soft rain behind it and no speech or music.
The action is intentionally small. Review the head, both feet and the side key throughout the clip. If the bird walks away, the shot fails the planted-feet requirement even if the motion looks attractive. If the key vanishes, record the identity failure separately from whether the head turn happened.
Image-to-video example using the practice reference
Open the original bird practice image and save it if your chosen workflow accepts a first-frame image. Confirm that the upload preview has the expected orientation and crop. Use only a supported input format in the actual tool.
Animate this reference as one continuous locked shot. The brass bird gently turns its head toward the rainy window and then returns to its opening pose. Keep its cobalt wings folded, its coral side key visible and both feet touching their original positions on the workbench. Preserve the workshop composition and the warm lamp against the cool window light. Add a subtle mechanical tick during the turn and quiet rain ambience. There is no dialogue or music.
This version describes a change to an existing composition. It avoids asking for a new workshop, a different species or a second bird. Preservation language is a request, not a guarantee. Compare the exported clip to the reference at the beginning, during the turn and at the ending.
Multi-shot example with explicit continuity
A medium view shows a brass songbird with cobalt folded wings and a coral side key standing on the workshop bench. It turns its head toward the rainy window as quiet rain fills the room. A hard cut moves to a close-up of the same coral key on the same brass bird; the warm lamp and rain ambience continue across the cut. The key makes one slow quarter turn while the bird remains still. A hard cut returns to the original medium framing, where the same bird holds its opening pose and the rain continues without music or speech.
Use this as a current LTX 2.5 sequence exercise, not evidence that earlier releases support the same behavior. First check whether the bird and key survive the cut. Next check that the key turns once rather than spinning indefinitely. Finally check whether the return shot matches the requested composition. A failed cut can be repaired as a separately generated shot and assembled in an editor.
Plan native frame counts without confusing duration
The native PyTorch pipeline documentation specifies dimensions divisible by 32 and frame counts in the form 8n + 1. For example, 121 is valid in that sequence, while 120 is not. These are native pipeline constraints; verify the controls exposed by a hosted provider separately.
The free planner above proposes a nearby grid and frame count. At 24 fps, 121 encoded frames occupy approximately 5.0417 seconds, while the timestamp difference between the first and last frames is 5 seconds. Those are different measurements. Rounding 1080 to 1088 pixels also changes the aspect ratio slightly; choose an actual crop or pad operation rather than silently stretching the image.
Grid compliance does not prove a job fits your GPU. The native documentation distinguishes native pipeline weights from ComfyUI-specific variants and the separate Diffusers distribution. Follow the setup instructions for the exact implementation. Avoid mixing a template, checkpoint and command from different releases.
Review a result with a small failure log
For another first-frame exercise, use the Hunyuan Video motion guide. Keep each implementation’s settings separate while comparing the action and subject continuity.
| Check | Pass condition for this exercise | If it fails |
|---|---|---|
| Action | One head turn and a return | Simplify to one turn before adding a return |
| Contact | Both feet stay on the bench | Reduce body motion and keep framing fixed |
| Identity | Cobalt wings and coral key remain recognizable | Return to the accepted reference |
| Camera | No unwanted cut or drift | Remove additional camera instructions |
| Sound | A tick aligns with the turn and rain remains quiet | Review audio separately or replace it in post |
| Ending | The final pose is intentional | Describe a pause at the end |
Save the exact prompt, model identifier, settings, reference and exported clip together. Change one thing for the next attempt. If you change the camera, subject and sound simultaneously, a better result will not tell you which change helped. Review at normal playback speed before examining individual frames.
Keep important labels and logos out of the generated scene when their exact appearance matters. Add them in an editor and inspect the full export. For a sound-led exercise, compare our AI ASMR planning guide. For another model's reference and mode distinctions, use the Seedance workflow guide.
Create visual references and original content in Clout. This article's planner does not run LTX or charge credits. Generation availability and settings belong to the actual app and selected workflow.
