4:5 AI video for Instagram: generate 3:4, then crop with video-filter

None of the Video Router models lists a 4:5 aspect ratio, so you cannot ask for it directly. The route that works is to generate at 3:4, which Wan 3.0, MiniMax H3 and H3 Max all list, and crop the result to 4:5 with Sume's video filter. A 3:4 frame keeps all of its width and loses 6.25% of its height, so the crop is height: 0.9375 centred, and the encode costs $0.02.

Kling 3 lists only 16:9, 9:16 and 1:1, and Gemini Omni Flash 1.1 lists only 16:9 and 9:16, so neither can reach 4:5 this way without a larger crop.

Which models can start from 3:4?

The catalog's aspect_ratios field is the source. Read it from GET /v1/video-router/models before you submit, because the sets differ per model.

How big is the crop?

The filter's crop op takes a rectangle as fractions of the source frame: x and y between 0 and 1, width and height between 0.05 and 1, with x + width and y + height each at most 1. The compiler rounds to even pixel sizes for yuv420p. Starting from 3:4, 4:5 is wider by a factor of (4/5) / (3/4) = 16/15, so you keep full width and cut the height to 15/16, which is 0.9375. Centred, y is 0.03125. The snippet below computes it, and prints the op you paste into the request.

def crop_3_4_to_4_5():
    # A 3:4 frame is 3 wide by 4 tall. Keep the full width,
    # then the height that makes the frame 4:5.
    width_units, height_units = 3, 4
    keep_h = width_units * 5 / 4          # 3.75 units tall
    height = keep_h / height_units         # fraction of source height
    y = (1 - height) / 2                   # centred
    return {"op": "crop", "x": 0, "y": round(y, 5),
            "width": 1, "height": round(height, 5)}

op = crop_3_4_to_4_5()
assert op["y"] + op["height"] <= 1
print(op)

What does the request look like?

Video filter takes one clip hosted on your workspace's media.sume.com, so the generated clip must already be an artifact there. It requires an Idempotency-Key, defaults to async, and accepts mode: "sync" to wait up to 30 seconds. The same program can be checked first with POST /v1/video-filter/check, which is free, creates no job and reserves nothing.

The body is {"video_url": "https://media.sume.com/artifacts/artf_demo/clip.mp4", "ops": [{"op": "crop", "x": 0, "y": 0.03125, "width": 1, "height": 0.9375}]}. The result is a new MP4 with a new artifact id; the source stays untouched.

What if I only have a 9:16 clip?

The same arithmetic gives a bigger loss. A 9:16 frame is 9 units wide and 16 tall, and a 4:5 frame at that width is 11.25 units tall, so you keep full width and 11.25 / 16 = 0.703125 of the height, which loses about 30% of the picture. That is a hard crop for a clip framed for vertical video, because heads and captions near the top or bottom will be cut.

The benefit of generating at 3:4 is that the loss is only 6.25% of the height. A tall, centred subject, such as a product on a table or a person at a desk, survives it easily. Anything with text near the top or bottom edge does not, so add captions after the crop rather than before it.

If you also need the 9:16 version, generate once at 3:4 and crop once for 4:5, or generate a 9:16 clip separately on a model that lists that ratio. The filter's check endpoint validates a program without creating a job, so you can try both crops for free before paying for either.

What does the whole route cost?

A 5-second Wan 3.0 clip at 720p is $0.63 on Sume, and the crop adds $0.02, so $0.65 in total. The same clip on H3 Max at 768p is $0.50 plus $0.02. Generating at the higher resolution first leaves more pixels after the crop, but pays for them: the same Wan clip at 1080p is $1.25.

Check one thing before you ship: a crop removes picture. If the model placed a caption or a face near the top or bottom edge, 6.25% of the height is enough to clip it, so keep important content centred in the prompt.