- Establish the shot first, then scene, action, characters, camera, audio. Use LTX's published camera vocabulary verbatim: pushes in, pulls back, pans across, circles around, tracks, static frame. A synonym is a worse prompt than the documented term.
- The tip almost nobody follows: describe how the subject appears after the move completes. That gives the model a destination rather than a direction.
- Frame count bounds move length. At 49 frames you have under two seconds at 25fps, and a travelling move truncates rather than slowing to fit.
Camera language belongs in the prompt because LTX-2.5 reads the prompt as a shot description, not a caption.
The LTX-2.5 prompt guide documents six elements a strong prompt should cover, and one of them is reserved for camera movement. Shot size, camera behaviour and subject blocking are a slot, not decoration.
This post covers what goes in that slot, where it sits, which words the model was actually trained to recognise, and how long a move can hold given the frame-count rule.
Which surface this applies to
Prompt structure applies wherever you write an LTX-2.5 prompt: the open-source pipelines, ComfyUI, and the hosted LTX API.
The frame-count arithmetic further down uses --num-frames, an open-source flag. On the API you choose a duration from a fixed set and a frame rate of 24, 25, 48 or 50 instead.
Running locally needs Linux, Python 3.12 or newer, CUDA 12.7 or newer, and an NVIDIA GPU with 80GB or more of VRAM, with distilled variants supported on 32GB using quantization.
Where camera terms go
The documented order is: establish the shot, set the scene and lighting, describe the action, define the characters, identify the camera movement, describe the audio.
Camera arrives fifth, after the subject exists, so the model already knows what it is framing. Audio comes last and is not optional on a model that generates sound jointly with picture.
Note what comes first. Establish the shot, not the action. Open with shot scale and type, in the construction a shot list would use: "Interior medium close-up." Then the scene.
Write the whole thing as a single flowing paragraph in present tense. Match length to complexity rather than a word count, which for a single shot usually means four to eight descriptive sentences.
How do you prompt shot size?
Shot size is a literal clause, and the vocabulary you would write on a shot list is the vocabulary to write here.
Use the standard names: extreme wide shot, wide shot, medium wide shot, medium shot, medium close-up, close-up, extreme close-up.
Angle is a separate term from size, so state both when both matter. A low-angle medium shot is a different instruction from a medium shot.
One size per generation. LTX-2.5 introduced native multishot for cuts, so stacking two shot sizes into one description is the wrong tool for the job.
Use the published camera vocabulary
This is the part most guides miss. LTX publishes the camera language the model recognises, and using those exact terms beats a synonym every time.
The documented vocabulary is: follows, tracks, pans across, circles around, tilts upward, pushes in and pulls back, overhead view, handheld movement, over-the-shoulder, wide establishing shot, and static frame.
"Pushes in" is documented. "Creeps forward" is not. The first is a term the model was trained on; the second is a guess.
Static is a real instruction and worth writing explicitly. "Static frame" pins the camera and lets the performance carry the shot. Leaving camera language out entirely is not the same as asking for static.
Keep it to one move per shot. LTX also documents that highly chaotic motion introduces artifacts while simpler, plausible motion is more reliable, which is the same discipline applied to the camera.
The tip most people miss
Specify how and when the camera moves, describe the movement relative to the subject rather than in absolute terms, and then do one more thing.
Describe how the subject appears after the movement completes.
This is the highest-value instruction in LTX's guide and it is the one almost nobody follows. Saying "the camera pushes in" gives the model a direction. Saying what the frame contains once the push-in lands gives it a destination.
Compare "the camera pushes in on the courier" against "the camera pushes in on the courier; as the move completes, her face fills the left third of frame and the shutter fills the background." The second one tells the model where to end up.
What is blocking in a prompt?
Blocking is where the subject sits and moves within the frame relative to the camera, and it belongs in the action clause rather than the camera clause.
State position in frame terms, not world terms: centre of frame, left third, foreground right.
Then state the change chronologically: "he steps from the left third of frame into the centre, turning to face camera as he arrives."
Express emotion through physical cues rather than labels. "Shoulders drop, jaw tightens" works where "he looks sad" does not.
How long can a move be?
Move length is bounded by frame count.
The Video VAE encodes pixels to latents with the constraint that frame count must satisfy (F-1) % 8 == 0, where the latent frame count is F' = 1 + (F-1)/8. Width and height must also be divisible by 32. A number that fails either rule is invalid input, not a soft preference.
Valid frame counts include 9, 17, 25, 33, 41, 49, 57, 65, 73, 81, 89, 97 and 121. At 25fps that puts 49 frames at 1.96 seconds and 57 at 2.28, and the repository quick start generates with --num-frames 121.
Do the arithmetic before you write the move. Two seconds is not enough for a slow push to read as a push; it reads as drift. A move that needs to travel needs frames to travel across, so either raise the frame count or simplify the move.
The move does not slow down to fit the clip. It truncates.
If you would rather not do the arithmetic, the duration head can predict clip length from the prompt when you omit --num-frames.
Camera direction across cuts
Multishot changes what camera direction means, because one generation can carry two to four connected shots.
At every cut, name the transition in natural language, re-establish the new shot's scale and angle, reuse the same visual identifiers for recurring subjects, and say whether the audio continues or changes.
Give each shot a clear job: establish, then detail, then reaction.
Stay with a single continuous take when you want unbroken camera motion, an intimate performance, or dialogue that must stay lip-synced in one framing. For image-to-video from a first frame, prefer a single take unless you intend to cut away from that opening image.
When to stop prompting
When the move has to be exact and repeatable rather than approximately right, a prompt is the wrong tool.
IC-LoRA depth conditioning transfers camera and 3D geometry from a reference clip, which makes a move reproducible in a way a description is not. ICLoraPipeline handles that path and runs on the distilled transformer. The IC-LoRA guide walks through it, and the motion and camera control page covers the wider surface.
Seven pre-trained camera-move adapters also exist, covering dolly in, dolly out, dolly left, dolly right, jib up, jib down and static, with no reference clip required. They are published against the LTX-2 19B checkpoint. Most LTX-2.3 adapters run on LTX-2.5 unchanged per LTX's model card, with a small number of exceptions, so validate yours against your checkpoint before you build a shot list around it. The community library carries others worth checking first.
Pick one source of motion. If a reference clip supplies the camera move, describing that move again in the prompt gives the model two instructions for one behaviour and degrades the output.
An annotated example
This follows the documented order. The camera clause is bold.
Medium wide shot, exterior, narrow street at dusk. Overcast light, desaturated greys, wet asphalt holding a little reflection. A courier walks a bicycle in from the left third of frame, then turns to face camera as she reaches the centre, stopping to lean the bike against a closed shutter. She wears a yellow rain shell, wet at the shoulders, with a canvas bag across her back; her shoulders are loose and she is looking up rather than ahead. The camera tracks her from her left side at chest height; as the move settles, she sits centre frame with the shutter filling the background behind her. Tyre on wet ground, a distant car, the rattle of the shutter as the bike touches it.
Read it back against the order: shot, scene and light, action with blocking, character, camera with its destination, audio. One move, one shot size, and a frame count you would set before writing it.
Common mistakes
Stacking moves. "Pans left then pushes in" asks two moves to complete inside one short clip. Pick one, or use multishot and make it a cut.
Opening with the camera. Camera language before the subject exists inverts the documented order and hands the model a frame with nothing in it. Establish the shot, then the scene, then the action.
Contradicting a reference. Describing a camera move in the prompt while a reference clip supplies the motion sets two instructions against each other.
A move too fast for the frame count. A travelling move at 49 frames has under two seconds at 25fps, and it truncates rather than slowing.
Forgetting the enhancer rewrote it. LTX-2.5 ships a prompt enhancer that expands terse input using a system prompt. Read the enhanced version before generating, because it is what the model receives and it may have moved or softened your camera clause.
Summary
Camera language works in an LTX-2.5 prompt because the documented structure reserves a slot for it, fifth in an order that runs shot, scene and lighting, action, characters, camera, audio.
Establish the shot first, not the action. Write a single flowing paragraph in present tense and match length to complexity rather than a word count.
Use LTX's published camera vocabulary verbatim, keep it to one move per generation, and describe how the subject appears after the move completes.
Blocking belongs in the action clause, in frame terms, stated chronologically.
Frame count bounds how much motion can read. Valid counts satisfy (F-1) % 8 == 0, so 49 frames is 1.96 seconds at 25fps, and dimensions must be divisible by 32.
When a move must be exact, IC-LoRA depth conditioning or a pre-trained camera-move adapter carries it from a reference instead.
Write your next shot list as a prompt. One move per clip, and check the capabilities overview for what else the model will do in the same pass.
