Build an Image-to-Video Prompt That Controls Subject Motion, Camera Movement, and Visual Style

How The Models Read Your Prompt

The correct image-to-video prompt structure is subject motion, then camera motion, then visual style, then negative prompt. That order matches how these models process their inputs: text conditioning attaches to the source image first, camera instructions come second, and style modifiers act as a final filter. Get the order wrong and the model will treat a camera instruction as a subject description, or ignore the style entirely. The prompt builder tool encodes exactly that structure. It is a local interactive guide. Nothing leaves your browser.

Subject Motion Leads

Every image-to-video model shares a pipeline: encode the source image into latent space, add noise, then denoise across frames using temporal attention layers. The text prompt enters after the initial image encoding, during the denoising steps. The first element the model reads has the strongest influence on what moves.

Start with the subject and its motion. Write what the person or object does. The model's text conditioning connects this to features it detected in the source image. A prompt that begins with camera motion, like 'pan left', will instead try to animate the subject as if it were an object performing a panning motion. The model has no prior expectation about what 'pan' means for a person. The result is a static subject with a background that slides incorrectly, or no motion at all.

Camera Movement Second

Second, the camera motion. This tells the model how the viewpoint moves through the scene. Kling 1.6 and Runway Gen-3 Alpha both publish explicit camera-motion vocabulary: 'zoom in', 'zoom out', 'pan left', 'pan right', 'tilt up', 'tilt down', 'roll', 'tracking shot'. Luma Dream Machine supports the same set. If you place camera motion before subject motion, the model often applies the camera move to the subject instead. You get a warped figure that appears to zoom or tilt rather than a camera that moves past it.

Style And The Negative Prompt

Third, the look. This is the visual style: lighting, colour palette, texture, time of day. The model applies style conditioning late in the denoising process, so it cannot override subject or camera motion. A style placed early will get interpreted as a motion instruction. Write 'cinematic lighting, shallow depth of field, golden hour' after you have told the model what moves and how the camera tracks it.

Fourth, the negative prompt. This is the text that tells the model what to avoid. Only Pika 1.5 and later, checked on 2026-09-26, supports a negative prompt field in its UI. Every other model in this comparison ignores it. Runway Gen-3 Alpha, Kling 1.6, Luma Dream Machine, Sora, Veo, Stable Video Diffusion, LTX-Video, Hailuo, Wan, and Seedance do not document negative prompt support. If you are using one of those, your negative prompt text is dead words. The prompt builder tool marks which models accept a negative prompt so you do not waste a generation.

Motion collapse is the most common failure mode in image-to-video generation. The output is nearly static despite a prompt that asks for significant movement. It happens because the model biases toward low-motion outputs to avoid visible artefacts. The prompt ordering is usually the cause.

If you wrote 'a woman walks forward, zoom in, cinematic lighting' and the clip shows a woman standing still while the background zooms, your subject motion was placed too late. The model read 'zoom in' as the primary instruction and treated 'a woman walks forward' as a style descriptor. Rewrite it as 'the woman walks forward, zoom in, cinematic lighting', moving the subject motion to the front. That reorders the text conditioning so the temporal attention layers anchor to the walking motion first.

Another sign of motion collapse is high motion strength paired with an empty subject description. Stable Video Diffusion uses a motion bucket ID from 1 to 255. If you set it high but write only 'a person, zoom in, soft light', the model has no subject motion to generate and defaults to a near-static clip. Give the subject a verb: 'a person walks, zoom in, soft light'. The motion bucket ID then amplifies that walking motion rather than creating motion from nothing.

Camera motion instructions work only if the model has been trained to recognise them. Kling 1.6, checked on 2026-09-26, documents this vocabulary: 'zoom in', 'zoom out', 'pan left', 'pan right', 'tilt up', 'tilt down', 'roll clockwise', 'roll counterclockwise', 'static'. Runway Gen-3 Alpha, same date, adds 'tracking shot', 'crane up', 'crane down', and 'roll' without direction. Luma Dream Machine supports the same set as Runway. Pika 1.5 adds 'roll' alongside the pan and tilt instructions.

Sora, Veo, Hailuo, Wan, Seedance, Stable Video Diffusion, and LTX-Video do not publish explicit camera-motion vocabulary. If you use those terms, the model may interpret them as scene descriptions rather than camera instructions. For those models, you rely on the subject motion to imply camera movement: 'a person walks toward the camera' achieves a dolly-in effect without a camera instruction.

Write camera motion as a single phrase. Do not combine multiple moves in one clause. 'Pan left and tilt up' produces a jittery diagonal movement because the model's temporal attention cannot split the instruction cleanly. Write two separate generations and stitch them in post-processing with FFmpeg.

Of the twelve models in this comparison, only Pika 1.5 and later, checked on 2026-09-26, provides a negative prompt field. The field is documented in the Pika web app UI. If you are using Pika, write your negative prompt after the positive prompt, describing what the model should avoid: 'blurry, distorted face, unnatural colors, jittery motion'. Pika's text conditioning applies this during the denoising steps, reducing the chance of those artefacts.

Every other model ignores the negative prompt. Runway Gen-3 Alpha, Kling 1.6, Luma Dream Machine, Sora, Veo, Hailuo, Wan, Seedance, Stable Video Diffusion, and LTX-Video do not have a documented negative prompt input. If you include one in the text field, the model treats it as a positive instruction. Writing 'no motion' in a negative prompt on Runway will produce a static clip, because the model reads 'no motion' as a request for stillness. The same text on Pika suppresses motion instead.

For models without negative prompt support, control artefacts through the motion strength parameter instead. Stable Video Diffusion's motion bucket ID from 1 to 255 lets you tune how much motion the model generates. A lower ID produces less motion and fewer artefacts, at the cost of a less dynamic clip. Set it to 127 for a balanced result, then adjust based on whether the output has too much drift or too little movement.

Take a source image of a woman standing in a forest, holding a red umbrella. The goal is a 5-second clip where she turns to look left, the camera pans right to follow her gaze, and the scene has soft morning light with no distortion.

The bad prompt. 'Cinematic lighting, soft morning light, a woman turns her head to the left, pan right, no blur, no warping.' The model reads 'cinematic lighting' first and treats it as a motion instruction. It attempts to animate the light itself, producing a flickering clip. 'Pan right' comes after the lighting instruction and is weakened. The negative prompt 'no blur, no warping' is ignored by every model except Pika. The woman remains static or moves unnaturally because the subject motion was placed third, after a lighting description that the model interpreted as motion.

The good prompt. 'The woman turns her head to the left, pan right, soft morning light.' The subject motion comes first. The model's text conditioning latches onto 'turns her head' and applies that to the woman in the source image. The camera motion comes second. 'Pan right' tells the temporal attention layers to shift the viewpoint. The look comes last, applying soft morning light as a style modifier that does not interfere with motion. The negative prompt is omitted because the model in use is not Pika. The clip shows a woman turning her head while the camera pans right, with consistent lighting and no subject drift for the first 4 seconds.

If you are using Pika. Add the negative prompt after the look: 'The woman turns her head to the left, pan right, soft morning light, blurry, distorted face, unnatural colors.' Pika's text conditioning will suppress those artefacts. On any other model, that negative prompt text is wasted.