How to choose reference images for AI video

Step 1: gather more than you need

Collect 15 to 30 candidates in one folder: phone photos, earlier generations, product shots, frames from a video. Video frames are useful because they show angles people rarely photograph, but take them from a sharp, well-lit moment, not mid-motion.

Remove duplicates and near-duplicates first. Five almost identical front photos make the model more certain about the front and no more certain about anything else.

Step 2: run the five tests

Each image must pass all five. Be strict. A weak image does not just add nothing; it pulls the result toward its flaws.

1. Sharpness. Zoom to 100 percent on the eyes or the logo. If they are soft, reject the image. Motion blur, missed focus and heavy compression all fail.

2. Light. The subject should be lit evenly enough to see both sides of the face or product. Hard shadows across half the face, colored stage light and strong backlight fail. Light that is warm in one image and cold in another shows up as color shifts between clips.

3. Background. Plain or softly out of focus passes. A busy street, a patterned wall, other people or readable text fails, because the model may treat repeated background elements as part of the subject.

4. Same look. Hairstyle, makeup, glasses, beard length, outfit or product colorway must match the rest of the set. An older photo with different hair is a different version of the character, even if it is a great photo.

5. Clean. No filters, beauty retouching, watermarks, stickers, text overlays, or crops that cut through the head or the product. The model treats all of these as features of the subject.

Step 3: pick for coverage

From the images that pass, choose for angle, not beauty. For a person, the target set is a front view, a three-quarter view, a profile and one wider shot that shows the body and outfit. For a product, it is the hero angle, a side, the back and a logo close-up.

Sort the survivors into those angle slots, then keep the best image in each. If a slot is empty, note it. That is a gap to fill in step 4, not a reason to promote a failed image.

Size the set to the tool. As of October 2026, Kling's guide for Kling 3.0 Omni allows up to seven reference images, or four when a reference video is included, each at least 300 pixels on each side and at most 10 MB. Veo 3.1's reference mode takes three. When the limit is three, use front, three-quarter and one wider shot. Put the clearest front view first, since many tools treat the first image as the main one.

Step 4: fill the gaps on purpose

A missing profile is the most common gap. You can photograph it if the subject is available, generate it with an image model's edit mode using your best front view as input, or accept the gap and avoid profile shots in your scenes. Generated angles must pass the same five tests. Check that the ear, the jawline and the hairline match the real images.

Then run a short test: four clips in different situations with the final set. If one kind of shot keeps failing, the set is telling you which angle is missing.

Common mistakes and checklist

Using only the most flattering photos, which are usually all front views. Mixing photos from different years. Keeping a great photo with sunglasses when the character does not wear sunglasses. Cropping so tight that the hairline or the product's cap is cut off. Uploading the maximum number of images when only three are good.

Checklist: 15 or more candidates; duplicates removed; every kept image passes sharpness, light, background, same-look and clean; one image per angle slot; best front view first; set sized to the tool; gaps filled on purpose; four-clip test run.

RefVid takes 1 to 5 images per character or product, in JPG, PNG or WebP, at least 300 px and up to 10 MB each, and uses the first as the main view. You choose the set once, and every video of that asset is generated from it on Kling 3.0 Omni.