MiniMax H3 settings for short product clips

Three modes and when each fits

Image-to-video from a first frame is the mode we tested, and for products it is the safest: the bottle, jar or box in the clip is the one in your photo. Reference-to-video takes up to 12 files and suits a campaign where the same product has to appear in several different clips. Text-to-video invents everything, which is fine for mood shots and wrong for a product a buyer will receive.

The image-to-video row follows the shape of your start frame rather than offering a ratio setting, while the text and reference rows let you pick 9:16, 1:1, 16:9 and other ratios. If a placement needs an exact ratio, crop the start frame before you upload.

768P or 2K

H3 is priced per second, and 2K costs more per second than 768P. The From figure in the table is 5 seconds at 768P, the default. Our four tested clips were all made at 768P and six seconds, which was enough to judge the motion, the light and whether the product held.

Approve the motion at 768P, then rerun the keeper at 2K only if the placement needs the larger file, such as a product-page video viewed on desktop. For a vertical ad watched on a phone, 768P is often all you need. Do not count on 2K to rescue a clip whose motion was wrong at 768P.

Length works the same way, since every extra second is billed. Six seconds was enough in our tests for two drops of serum or a full fall of petals. Pick the shortest length that lets the motion finish, and add seconds only when the action visibly needs them.

Output quirks to plan for

H3 did not return exactly what we asked for. Six-second requests came back at 6.58 seconds. A 9:16 start frame came back at 768×1344, which is 4:7, slightly narrower than 9:16. Every clip carried an audio track, although we wanted silent clips for muted autoplay. None of it is hard to fix, but plan a trim to length, a crop or pad to the exact ratio and an audio pass before you upload anything.

The clips took about 3 to 4 minutes each to come back, quicker than our Seedance 2.5 runs in the same session.

Three recipes from our beauty tests

Mist: describe the mechanism, not the look. “Presses the nozzle down twice, a fine horizontal jet from the side of the nozzle” produced two distinct sprays, while the word “mist” on its own sent a cloud straight up out of the cap like smoke. A dark background with a backlight is what makes fine mist visible at all.

Drops: describe how the liquid behaves and give it a rhythm. A drop that “swells and falls”, followed by a second drop that forms and falls, stretched before letting go the way serum does and gave six seconds with a natural loop point. The bottle and pipette stayed still because the prompt told them to keep exactly the same shape and position.

Loops: a still product with moving light loops most cleanly, because the first and last frames are almost identical. Palm shadows crossing a cream jar, explained as leaves moving in a breeze outside the frame, moved only the shadows and added no leaves to the shot. It is the pattern we would reuse for any product-page loop.