VellmAI video looks wrong? What to change before you try again
By Vellm ·
Watch the clip once and note the first moment that goes wrong. A face that is wrong from the opening suggests a different problem from one that changes during a fast spin. That difference helps you decide whether to replace the photo, simplify the movement or change the reference.
Before you rerun, save the inputs, prompt, model and settings. A new generation can use more credits. Changing one thing at a time makes each attempt easier to learn from.
The face or pet markings change
Check the source photo at full size. A tiny face, heavy filter or hand over the mouth gives the model less usable detail. Choose a clearer image with even light and one intended subject.
If the subject starts correctly and changes during a turn, reduce the turn's speed or size in a custom prompt. Keep the camera simple while testing. In a fixed template, choose a clearer photo or a routine with less demanding movement; the template's choreography is not an open text setting.
Increasing the resolution of a finished clip cannot reliably restore the exact face or markings that changed during generation. Check identity in the moving output before spending effort on a larger export.
Extra paws, fingers or sliding feet appear
For a full-body action, use a photo that shows the limbs clearly. Avoid legs hidden by furniture, fur that merges into a dark floor, or a close crop that removes the feet.
In Create, simplify the action. A dog shifting its weight and lifting one paw is easier to inspect than a spin followed by several jumps. For remakes, look for a reference where the performer's limbs remain visible and do not repeatedly cross other performers.
Clearer inputs reduce visual confusion, though AI video models can still produce anatomical flaws. Always review the entire clip, right up to the final second, before using the file.
The right subject appears in the wrong role
Read the upload labels again. A template with two roles may assign the first photo to the left performer and the second to the right. Swapping the photos changes the intended casting.
For Vellm's Hotel Lobby cat duo, Photo 1 identifies the left-side rapper and Photo 2 identifies the right-side hype pet. For La Calabaza, one photo supplies all three performers; extra photos select the side pets. Those are different mapping rules.
With custom references, explain which subject or visual element each file supplies. If both are called "the cat" in the prompt, use consistent positional labels such as "the cat on the left."
The movement is nothing like the dance
A text prompt with the name of a routine leaves the model to interpret the choreography. If an available template matches the dance, use that configured routine. If you have another clip you can use, try Remake a video with the reference and your subject photo.
If a remake misses the key action, trim the source to that section. Check that the original performer is visible and that the camera is not cutting away. A shorter, readable reference gives you a clearer comparison than a montage.
The camera cuts off the subject
For Animate a photo, begin with an image that already leaves room around the subject. Some models inherit the image proportions, so look at the selected model's controls before assuming a format setting will rebuild the scene.
Try a specific camera instruction such as:
A locked wide shot keeps the whole dog visible, with clear floor space below its paws throughout the movement.
If the template has fixed framing, you can change the photo or choose a different template. A prompt cannot change controls the page does not offer.
There is no sound
Vellm's current templates export silent video. A reference preview may have music while the generated download does not.
Create also starts with audio off. Look for the Audio chip on a model that supports sound. Enabling it does not guarantee a reference song, and naming a song in a prompt does not provide permission to use it. Add suitable music in your editing or posting workflow.
The generation failed, or the page stopped responding
A job marked as failed is different from a completed clip with an unwanted face or movement. Vellm's published policy returns the credits used by a failed generation. It does not promise a free rerun for every completed result you dislike.
If the browser loses connection or seems stuck, check your existing results and job status before starting another generation. The page may have stopped updating while the original job still exists. Keep the job details if you need to contact support.
For the next creative attempt, write a one-line diagnosis: "The face is correct until the fast turn" or "The paws are missing in the source photo." Change the input or instruction that addresses that observation. Use the photo guide or prompt examples for the next edit, and read the credit cost before generating again.