A 4000-pixel photo sounds better than a 900-pixel photo. Sometimes it is. But if both images come from the same blurry original, the larger file may only contain more pixels representing the blur.
Look at the face at actual size
Zoom to a normal viewing size and inspect the eyes, mouth, nostrils and jawline. If those features are sharp and natural, the image gives the model useful information. If they are smeared by motion blur or aggressive noise reduction, a higher export dimension is not going to recreate their true shape.
Compression can destroy small details
Messaging apps and screenshots may reduce quality, particularly around hair, eyelashes and thin glasses. Repeatedly saving JPEG files can add halos around high-contrast edges. Prefer the original image file where possible.
A clean JPEG is usually fine. PNG or WebP may also work if your chosen provider accepts them. The format matters less than whether the visible detail survived.
Cropping is a trade-off
A tight crop makes the face larger, which can help when the original picture is a wide group photo. But crop too aggressively and you remove the ears, top of the head or jaw. Those features help establish the identity during motion.
Leave space around the head. If the output scene includes shoulders and neck, having some of that context in the reference can help the model interpret proportions.
Don't upscale just to meet a number
An upscale may be useful for compatibility with a model's minimum input dimensions, but it is not guaranteed to add trustworthy identity information. Overprocessed upscales sometimes invent eye texture or alter face shape.
If you have multiple source photos, choose the one with the clearest natural features, even if its pixel count is lower. That is the better starting point for a reliable video.
