Image-to-Video Workflow Recipes for Specific Cinematic Shots

The most common mistake in image-to-video generation is treating every source image the same way. A portrait shot, a landscape, a product still, and an abstract texture all fail in different ways under the same workflow. The model does not understand your scene. It predicts pixel motion from statistical patterns, and those patterns break differently depending on what your source image contains. The question is not which model is best. The question is which workflow matches the shot you need to produce. This page is a routing hub. It directs you to the correct shot recipe based on what you are trying to achieve. Each section below specifies what the source image must contain, gives the prompt and settings that work, and diagnoses the two or three ways that particular shot usually fails. The individual guides own the detail. This page owns the map. These image-to-video shot recipes exist because a shot that works for a slow camera pan across a landscape will collapse when you try to animate a human figure in motion. The recipe is the difference between a usable clip and a discard.

Every generative video shot workflow follows the same pipeline structure. The source image goes through preprocessing where it is resized, cropped, and adjusted to match the model's training aspect ratios. Then the model runs inference, encoding the image into a latent space, injecting noise, and denoising across frames with temporal attention layers. Finally, postprocessing applies frame interpolation, upscaling, and format encoding into an MP4 container with H.264 or H.265. That sequence is stable across every current tool from Runway Gen-4 to LTX-Video running on a local ComfyUI node. What changes between workflows is what happens inside each stage. A workflow for a close-up portrait needs a pre-processing step that centers the face and crops tightly to avoid background sliding. A workflow for a wide landscape needs a high motion bucket ID in Stable Video Diffusion or a strong motion strength slider in Kling 2.0 to produce visible movement instead of motion collapse. A workflow for a product shot requires aspectratio constraints that match the model's training data, typically 16:9 or 9:16, to prevent aspectratio distortion that stretches the product into an unnatural shape. The workflow also determines which failure modes you must watch for. Subject drift destroys a portrait workflow. Seam artefact destroys a workflow built on concatenating multiple short clips from a model that does not support startandendframe conditioning, like Luma Dream Machine which gives you 5 seconds per clip with no true endpoint specification. The workflow tells you where to look for the break. The pipe is the same. The recipe is what changes.

Image-to-video cinematic techniques depend entirely on whether the model gives you real camera-motion control or only a motion strength slider. Runway Gen-4 and its predecessor Gen-2 are distinguished by this single attribute. Gen-2 had only a motion strength slider. Gen-4 exposed a director-mode camera control panel, though as of September 2026 the documentation does not list separate pan, tilt, zoom, roll, dolly, or truck axes. Compare that to Kling 2.0, which gives you six camera motion presets for horizontal and vertical pan, zoom in and out, tilt, roll, and rotation, all controlled through numeric parameters. Luma Dream Machine offers six camera motion presets as well: pan left and right, tilt up and down, zoom in and out, and orbit. Pika 2.0 gives you pan, tilt, zoom, and rotate with numeric parameters alongside its sceneingredient and regionmasking controls that let you animate only a selected area of the frame while keeping the rest static. The difference between a preset and a full sixaxis control matters when you need a dolly move, a trucking shot, or a combined zoom and pan that no single preset captures. For those, you either need a model like Kling 2.0 that offers the individual axes or you need to postprocess the clip in FFmpeg by extracting frames and applying affine transforms yourself. The cinematic technique you can execute is bounded by the camera controls the model exposes. If the model does not expose them, you are limited to whatever camera motion emerges from the text prompt alone, and prompt adherence on the motion axis is weak across most models. Luma Dream Machine has strong wholeframe camera moves. Pika 2.0 has stronger selective animation. Choose the technique by the control you need, not by the model name.

AI video shot planning starts with a single decision: does the shot need the source image to remain recognizable at the end of the clip? If yes, you need a model with high sourceimage fidelity at clip end, which means a model that supports startandendframe conditioning. Kling 2.0 and Luma Dream Machine both let you upload an optional end frame image alongside the start frame. Pika 2.0 also supports startandendframe control. For these models, the workflow asks for a second image that shows what the scene should look like at the final frame. The model then interpolates between the two. That single feature eliminates subject drift and identity shift over the clip duration because the end state is anchored to a real image rather than to a statistical guess. For models without startandendframe support like Runway Gen-4, Sora, or Stable Video Diffusion, the source image fidelity degrades measurably after 4 to 6 seconds, roughly 40 to 60 percent of the claimed clip length. The usable clip length is shorter than the advertised maximum. Plan your shot around that limit. If you need 8 seconds of a face remaining recognizable, do not use a model that cannot anchor the end frame. Concatenate two shorter clips with a seam artefact risk that you will need to mask in postprocessing, or switch to Kling 2.0 which gives 10 seconds with endframe support. The planning question is not which model has the highest resolution. Kling 2.0 and Runway Gen-4 both output up to 1080p, but 1080p from a model that has lost the source image by frame 40 is worse than 768x512 from LTX-Video that holds the subject stable for its full 5 seconds. The effective resolution is a function of fidelity, not pixel count.

Every image animation shot guide must start with what the source image must contain, not what the model can do. The source image is the only anchor the model has. If the source image has poor contrast between the subject and the background, the model cannot distinguish them and will produce background sliding where the background moves independently of the foreground. If the source image contains a human figure with limbs that cross the body boundary, the model will produce limb mutation because it lacks a skeletal model and interpolates the arm across the torso. If the source image has a face in threequarter profile, the model will produce identity shift as it tries to interpolate toward the average face in its training distribution. The rule is simple: the source image must look the way you want the output to look at every frame, because the model has no persistent 3D world model and no physics engine. A cluttered source image produces a cluttered clip. A source image with a plain background and a centered subject produces a clip with less background sliding. A source image that matches the model's training aspect ratios, typically 16:9 for landscape shots and 9:16 for portrait vertical shots, avoids aspectratio distortion that stretches or compresses the subject. If the source image is square, crop it to 16:9 before uploading rather than letting the model's preprocessing pipeline do a naive resize. The source image requirement is the first filter because it is the only variable you control entirely. The model's output is a prediction from that single reference frame. Give it a clean reference and you reduce every failure mode: subject drift, limb mutation, background sliding, identity shift, flicker, and motion collapse.

Source Image Requirements

The source image must have a clear depth gradient from foreground to background. An image with a flat horizon line and no foreground element produces a static clip because the model has no cue for parallax motion. Include a tree, a rock, or a building in the foreground quarter of the frame. The foreground object provides the parallax anchor that the model uses to infer the camera is moving past the scene. Without it, the model defaults to minimal motion and you get motion collapse.

Prompt and Settings

Use a text prompt that describes the camera motion explicitly: "slow camera pan to the right across a landscape, foreground trees pass in front of distant mountains, steady movement." Avoid abstract terms like "peaceful" or "beautiful." They contribute nothing to the motion prediction. Set the motion strength to a middle value. On Kling 2.0, set the motion strength slider to 40 to 60. On Pika 2.0, set the motion strength slider to 5 or 6 on the 1 to 10 scale. On Stable Video Diffusion, set the motion bucket ID to a value that maps to 15 to 20 frames per second, which produces smoother panning than a higher frame rate. Use a clip length of 5 seconds maximum for models without startandendframe support. The model's temporal attention loses the camera direction after roughly 120 frames at 24 fps.

Failure Mode Diagnosis

Failure one is background sliding. The background moves but the foreground stays still, creating a layering effect that breaks the illusion of camera motion. Cause: the foreground element in the source image was too small or too similar in color to the background. Fix: recompose the source image with a larger foreground object. Failure two is motion collapse. The clip is nearly static. Cause: the motion strength was set too low or the prompt did not use explicit camera direction words. Fix: increase motion strength and rewrite the prompt to start with "slow camera pan." Failure three is flicker. Rapid brightness changes across frames. Cause: the source image had a high contrast area like a bright sky meeting a dark tree line. Fix: reduce contrast in the source image before uploading or use a model with stronger temporal consistency like Luma Dream Machine.

Source Image Requirements

The source image must be a frontalfacing portrait with the face centered and occupying at least 30 percent of the frame. The background should be plain and lowcontrast. A face in profile or threequarter view will produce identity shift as the model interpolates toward its training distribution. A cluttered background will cause background sliding behind the face. If the portrait includes hands, crop them out. Hands are the highest failure rate for limb mutation across every model. Crop to a headandshoulders composition.

Prompt and Settings

Use a prompt that specifies what moves and what does not: "the person blinks and smiles gently, the hair moves slightly in a breeze, the face remains still, the background remains static." Avoid prompts that describe emotion. The model cannot infer happiness from pixel motion. Use a motion strength setting well below half. On Kling 2.0, set motion strength to 20 to 30. On Pika 2.0, set it to 3. On Wan 2.1, set the motion strength slider to 15 to 25. Strong motion on a portrait guarantees limb mutation and identity shift. Use the shortest clip length available, 4 seconds on Sora or Pika 2.0, 5 seconds on Kling 2.0 or Luma Dream Machine. Longer clips on a portrait will always drift away from the source face.

Failure Mode Diagnosis

Failure one is identity shift. The face changes ethnicity, age, or gender by the fourth second. Cause: the model's temporal attention lost the source image conditioning. Fix: use a model with startandendframe support like Kling 2.0 and upload the same portrait as the end frame. The model will interpolate between identical images, producing minimal change. Failure two is limb mutation. If hands are in the frame, fingers multiply or detach. Cause: the model has no skeletal model. Fix: crop hands out before uploading. Failure three is flicker around the eyes and mouth. Rapid frame to frame changes in the eye shape or lip position. Cause: the model cannot maintain fine facial detail across frames. Fix: reduce motion strength further or switch to a model with higher native resolution like Kling 2.0 which outputs 1080p and preserves small details better than 1024x576 from Stable Video Diffusion.

Source Image Requirements

The source image must show the product against a plain, highcontrast background. A white product on a white background produces seam artefact and background sliding because the model cannot separate the subject from the background. Use a background color that is the complement of the product color. The product should occupy the center of the frame with at least 20 percent margin on all sides. The source image must be at the correct aspect ratio for the model. For a vertical product shot intended for social media, use 9:16. For a horizontal product shot, use 16:9. Square source images will produce aspectratio distortion that compresses the product horizontally.

Prompt and Settings

Use a prompt that describes the rotation axis: "the product rotates slowly clockwise on its vertical axis, the background remains static, the lighting remains consistent." Do not use the word "spin" because it triggers a fast rotation that most models cannot produce without motion collapse. Set the camera motion control to a slow zoom in or out rather than a rotation. Models with separate camera controls like Kling 2.0 can rotate, but the rotation will introduce background sliding unless the product is masked. For Pika 2.0, use the regionmask feature to brush select the product and set it to animate while the background is set to static. For models without region masks like Luma Dream Machine, keep the rotation minimal and accept some background movement that you can crop in postprocessing using FFmpeg to extract the center of the frame. Use a clip length of 4 seconds. Longer clips on a rotating product will cause the product to morph into a different object as the model loses the source image.

Failure Mode Diagnosis

Failure one is subject drift. The product changes shape or color over the clip. Cause: the model has no persistent object identity. Fix: use startandendframe support with the same product image as the end frame. Failure two is background sliding. The background moves independently of the product during rotation. Cause: the model cannot infer that the background is a fixed plane behind a rotating object. Fix: use Pika 2.0 region masks to freeze the background entirely. Failure three is motion collapse. The product does not rotate at all. Cause: the prompt used the word "spin" which the model could not execute, so it defaulted to static. Fix: rewrite the prompt with "slow rotation" and increase motion strength on Kling 2.0 to 50 or on Wan 2.1 to 40.

Source Image Requirements

The source image can be abstract, but it must have clear edges and high contrast between color regions. A soft gradient image produces flicker because the model has no hard boundary to anchor the morphing. Use an image with defined shapes, lines, or patterns. The model will morph these into different shapes, and the clear edges give it a starting point for the transformation. Avoid images with text. The model will hallucinate new text that reads as gibberish.

Prompt and Settings

Use a prompt that describes the transformation: "the shape morphs into organic flowing forms, colors shift slowly from blue to purple, the motion is continuous and liquid." Set the motion strength to a high value. On Kling 2.0, set it to 70 to 90. On Pika 2.0, set it to 8 to 10. On Wan 2.1, set it to 60 to 80. Unlike the portrait and product workflows where strong motion destroys the subject, abstract shots benefit from strong motion because there is no identity to preserve. Use the longest clip length available. 10 seconds on Kling 2.0, 8 seconds on Wan 2.1, 5 seconds on LTX-Video. The morphing is more interesting over longer durations. Use a model that handles fast action well, such as Hailuo Minimax, which is documented as strong on complex motion. The higher motion strength on abstract shots will still produce limb mutation if the image contains any humanlike form. Crop or blur human features before uploading.

Failure Mode Diagnosis

Failure one is motion collapse. The abstract shape stays still. Cause: the motion strength was set too low for an abstract image that lacks realworld motion cues. Fix: increase motion strength to the maximum on the model you are using. Failure two is flicker. Rapid color changes across frames that look like a strobe effect. Cause: the model's temporal consistency fails on abstract gradients. Fix: reduce the contrast in the source image or use a model with higher temporal consistency like Veo from Google DeepMind, which generates synchronised audio alongside the video and has strong motion fidelity. Failure three is contentmoderation false positive. An abstract image with certain shapes or patterns triggers the model's safety filter. Cause: the classifier flagged a shape that resembles prohibited content. Fix: preprocess the image to remove ambiguous shapes or switch to an openweights model like LTX-Video under Apache 2.0 licence which has no safety filter built in, though you assume responsibility for output compliance.

Source Image Requirements

The source image must show the subject in the middle of the intended action, not at the start. A sprinter in midstride with one leg forward and one back produces better motion than a sprinter standing upright. The model cannot invent the start of an action from a static pose. It can only continue the motion that the pose implies. The background must be plain. Fast action with a complex background causes background sliding at high speed, which creates a distracting jitter effect. Use a high native resolution model like Kling 2.0 at 1080p to preserve detail during the fast motion. Lower resolution models like Stable Video Diffusion at 1024x576 lose fine detail when the subject moves quickly across frames.

Prompt and Settings

Use a prompt that specifies the speed and direction: "the runner sprints forward at high speed, arms pump, legs stride, the background blurs with motion." Include motion blur in the prompt because the model will not add it without instruction. Set the motion strength to maximum. On Kling 2.0, set it to 90 to 100. On Wan 2.1, set it to 80 to 100. On Hailuo Minimax, use the platform's default which is optimized for fast action. Use a short clip length. 4 seconds on Sora or Pika 2.0, 5 seconds on Kling 2.0 or Luma Dream Machine. Fast action degrades quickly because the model's attention mechanism cannot track fastmoving subjects across many frames. The usable clip length for fast action is roughly half the advertised maximum.

Failure Mode Diagnosis

Failure one is limb mutation. The sprinter's legs warp or multiply during the stride. Cause: the model has no skeletal model and the fast motion increases the statistical uncertainty of limb positions. Fix: use startandendframe support if available, uploading an image of the sprinter at the end pose. Failure two is motion collapse. The sprinter stays in place with a running motion but does not move across the frame. Cause: the model interpreted the motion as cyclic limb movement rather than forward translation. Fix: add the phrase "moves across the frame from left to right" to the prompt. Failure three is flicker. Rapid frame to frame changes in the background during the fast motion. Cause: the motion strength was so high that the model prioritized the subject's movement over the background's temporal consistency. Fix: reduce motion strength to 70 on Kling 2.0 or 8 on Pika 2.0 and accept a slightly slower action in exchange for a watchable clip.

Source Image Requirements

The source image must have multiple depth layers. A tree in the foreground, a building in the midground, and a mountain in the background. The parallax effect depends on the model inferring that closer objects move faster than distant ones during a zoom. If the image has only one depth plane, the zoom produces a flat scaling effect that looks like a digital zoom, not a camera zoom. The source image must be at the model's native aspect ratio. A zoom on a 16:9 image produces consistent results. A zoom on a 4:3 image will produce aspectratio distortion as the model scales the image to fit its latent space.

Prompt and Settings

Use a prompt that describes the zoom and the parallax: "camera zooms in slowly toward the distant mountain, the foreground tree moves out of frame, the midground building passes to the right, the background mountain grows larger." Use a model with explicit zoom control like Kling 2.0 which has a zoom preset, or Luma Dream Machine which has a zoom in and zoom out preset. For models without zoom control like Stable Video Diffusion, you can simulate a zoom by setting a slow camera pan toward the subject and using a prompt that describes the zoom. The simulated zoom will be less clean than a native zoom control. Set the motion strength to a middle value. On Kling 2.0, set it to 40 to 50. Higher motion strength on a zoom shot produces background sliding that destroys the parallax illusion.

Failure Mode Diagnosis

Failure one is background sliding. All depth layers move at the same speed, producing no parallax. Cause: the source image lacked distinct depth layers with clear separation. Fix: recompose the source image with a foreground object that occupies at least 15 percent of the frame and overlaps the midground. Failure two is motion collapse. The zoom does not happen. Cause: the model has no zoom control and the prompt did not convey the zoom direction strongly. Fix: switch to a model with zoom presets. Failure three is seam artefact. If you concatenate two zoom clips to extend the duration, the jump between them is visible because the model does not support true startandendframe conditioning for zoom. Fix: use a single longer clip from a model that supports 8 to 10 seconds natively, like Kling 2.0 or Wan 2.1, and accept the slight quality degradation over the longer duration.

Source Image Requirements

The source image must have clear boundaries between the area you want to animate and the area you want to keep static. A person standing against a plain wall works well because the boundary between the person and the wall is sharp. A person standing in a forest does not work because the leaves create a complex boundary that the mask will miss. The source image should be high contrast at the mask boundary. Pika 2.0 is the only model documented as of September 2026 with regionmask support, accessed through its sceneingredient controls that let you brush select regions to animate or hold static. The source image must be uploaded to Pika 2.0 at its native resolution of up to 1080p.

Prompt and Settings

Use the regionmask tool to brush select the subject. Set the selected region to animate and the unselected region to static. Use a prompt that describes the subject's motion only: "the person waves their hand, the face remains still, the background remains completely static." Do not describe any background movement because the mask will block it. Set the motion strength to a middle value. On Pika 2.0, set it to 5 or 6. The regionmask reduces motion collapse risk because the model has fewer pixels to predict motion for. Use the 4second clip length that Pika 2.0 provides. The regionmask feature is computationally expensive, and longer clips increase the risk of the mask bleeding into the static area.

Failure Mode Diagnosis

Failure one is mask bleed. The animated region's motion spills into the static area, causing the background to move. Cause: the brush selection was not precise or the source image had a soft boundary between subject and background. Fix: use a source image with a hard boundary and repaint the mask with a smaller brush size. Failure two is subject drift. The person changes appearance over the 4 second clip even though only the hand was meant to move. Cause: the model's temporal attention loses the source image conditioning for the subject even within the masked region. Fix: use startandendframe support if available alongside the regionmask, uploading the same image as the end frame. Failure three is motion collapse. The hand does not move. Cause: the motion strength was too low for the masked region. Fix: increase the motion strength to 7 or 8 on Pika 2.0 and ensure the prompt specifies the exact hand movement.

The single most practical thing you can do after reading this page is to identify the specific cinematic shot you need and work backward to the source image requirements. Do not start by comparing model specifications. Start by asking what the source image must contain for that shot to succeed. A slow camera pan across a landscape needs a foreground element. An animated portrait needs a frontalfacing face with a plain background. A product rotation needs a highcontrast background and a centered product. A fast action shot needs the subject in midaction pose. An abstract morphing shot needs clear edges. A zoom with parallax needs multiple depth layers. A selective animation shot needs a hard boundary between the moving and static regions. Once you know the source image requirements, the model choice narrows to one or two options. For shots that need the subject to remain recognizable at the end, choose Kling 2.0 or Luma Dream Machine for their startandendframe support. For shots that need selective animation, choose Pika 2.0 for its regionmask feature. For shots that need fast action, choose Hailuo Minimax or Kling 2.0. For shots that need strong camera motion control, choose Kling 2.0 or Luma Dream Machine for their camera presets. For shots that need openweights local generation without a safety filter, choose LTX-Video under Apache 2.0. The model is the last decision, not the first. The source image is the first. Prepare the image for the shot, and the model will follow.