Video generation demos are easy to admire and hard to compare. When a tool takes a source clip and replacement images, a single attractive frame tells us little about whether the motion survived or the new subject stayed consistent. A better test uses a fixed input set and a short review rubric.
This is a browser based evaluation workflow, not an API tutorial. It is useful for developers building content pipelines, product teams assessing creative tools, or anyone trying to make feedback on generated clips less subjective.
Define the task before choosing the tool
Reference based recasting starts with an existing video. The system uses its timing, movement, and composition as the basis for a new clip while changing selected visual elements. That differs from asking a text to video model to invent the entire scene.
Choose a source clip that you own or have permission to adapt. For an initial test, use one subject and one clear action: a person walking across a plain room, or a product turning on a table. Keep the clip short. Fast cuts and frequent occlusion may look impressive, but they make it difficult to identify why an output failed.
Write down the invariant before generating: “The walking pace, camera position, and background should remain recognizable; the subject’s outfit should change.” This sentence becomes the basis of your review.
Build a fixed input set
Put the source video, reference images, prompt, aspect ratio, and output in the same folder. The images should show the angles required by the motion. A front only product photo is a weak reference for a rotating object; inconsistent colors across images make identity harder to preserve.
An example folder can be simple:
experiment-01/
source.mp4
references/
product-front.jpg
product-side.jpg
prompt.txt
result-01.mp4
notes.md
The prompt should identify one replacement and name what must remain. For example: “Replace the plain ceramic mug with the mug in the reference images; keep the camera movement, table, and lighting.” Avoid stacking new settings, camera moves, and style changes in the first run.
You can run such a test in AI genjutsu by providing a reference video or template, images, and a prompt. The site currently offers browser based video recasting; the workflow here makes no assumption about an API or programmatic access.
Review the whole clip, then the failure frames
I use four questions rather than one overall score:
| Dimension | Question |
|---|---|
| Temporal fidelity | Does the action happen at roughly the expected moments? |
| Identity | Does the replacement remain the same person or object throughout? |
| Physical contact | Do hands, surfaces, shadows, and occlusions make sense? |
| Publishing fitness | Is the result accurate and readable in the target format? |
Watch once at normal speed and record timestamps where something feels wrong. Then inspect only those moments frame by frame. A clip may pass the first three checks and still fail the fourth because a product label changed or a vertical crop cuts off the subject.
Iterate on one variable
If an object deforms during rotation, add a better side view or choose a source with less rotation. If a character changes clothing midway, remove conflicting reference images. If the framing fails on mobile, set the target aspect ratio before generation. Change one input per run and record the result. Otherwise, an improvement is hard to explain or repeat.
This process will not turn every draft into a usable video. It does turn a vague reaction into actionable feedback. AI genjutsu can shorten the route from a motion reference to a candidate clip, but the final evaluation still belongs to the team, especially when real products, people, or claims are involved.
Make the notes useful to the next reviewer
A review note such as “the output looks weird” forces the next person to repeat the entire inspection. Write observable findings instead: “At 00:04, the mug handle changes sides during the turn,” or “The subject leaves the 9:16 frame before the final beat.” Include the timestamp, the object, and the type of failure. A short note is enough if it points to a specific frame.
It also helps to separate creative preference from correctness. A team may prefer a slower reveal, but that is a design choice. A logo that changes shape is an accuracy issue. Treating these as separate categories keeps feedback productive: the first may call for another reference clip, while the second may call for clearer images or rejection of the output.
If several people review a candidate, ask them to watch independently before discussing it. One person can check the pacing, another the identity of the replacement, and someone familiar with the actual product can check physical details. The goal is not a formal benchmark score. It is a shared record of why a clip can or cannot move to the next stage.
Use a simple acceptance rule
An acceptance rule can be written before generation: the subject must remain recognizable throughout; no product details may be invented; the action must read in the target aspect ratio; and there must be no visible failure in the opening and closing seconds. You can add project specific requirements, such as legible packaging or an unobstructed face.
This rule prevents a strong single frame from outweighing a broken sequence. It also gives a clear stop condition. If a clip fails a requirement after a couple of targeted iterations, replacing the source footage may be more efficient than continuing to refine the prompt. Generative outputs have variability, and not every reference is equally suitable for recasting.
Finally, keep permissions in the acceptance rule. Owning the output file does not automatically grant rights to the source video, music, or a recognizable person's likeness. The team should know where the source came from and what use is allowed before the clip is published.
Takeaway
A useful AI video test has a permitted source, a coherent image set, one clear replacement, and a small rubric. Compare motion, identity, contact, and publishing fitness separately. Keep the inputs and notes so the next attempt benefits from what the last one revealed.
Top comments (0)