Lip sync still aspect ratio 0.4 to 2.5: which shapes pass

The still for Sume's MiniMax H3 Max Lip Sync must have an aspect ratio, width divided by height, between 0.4 and 2.5. A 9:16 portrait (0.5625) passes. A 21:9 ultrawide (2.33) passes. A 1:3 tall banner (0.333) and a 32:9 super-ultrawide (3.56) do not (as of 2026-10-09).

The check, ratio by ratio

The range comes from Sume's notes on this route, which describe the image as a public HTTPS still with aspect ratio 0.4 to 2.5, the same body as VEED Fabric except for the audio window. Compute width divided by height and compare.

What to do with a failing image

Crop before you upload. For a 32:9 image 3840 wide and 1080 high, a crop to 2.5 gives 2700 x 1080, which keeps 70 percent of the width. Center the crop on the face. For a 1:3 image 600 wide and 1800 high, crop to 0.4: height 1500, keeping 83 percent of the height, face centered.

The upload must still be a fetchable public HTTPS URL. Sume's media-input rules reject localhost and private-network URLs, so host the cropped file at a reachable public URL.

• Frame the face in the middle third; lip sync animates the mouth, so keep it clear of the crop edge.
• If the still is a Sume avatar, pass avatar_handle instead of image_url and skip the check.

Phone-first stills

Most talking-head content is made for phones, where 9:16 (0.5625) and 4:5 (0.8) both pass comfortably. 4:5 is a common feed crop and sits well inside the range. A 1080 x 1920 still has a ratio of 0.5625; a 1080 x 1350 still has 0.8; a 1080 x 2700 still has 0.4 and is exactly on the lower limit.

If you are unsure of the limit's inclusivity at the exact edge, stay inside it, for example 0.45. A 1080 x 2400 still has a ratio of 0.45 and avoids the question.

Which model takes which still

Treat 0.4 to 2.5 as the H3 Max limit. The Fabric docs on Sume do not publish a ratio range, so test the shape you plan to use with one short clip at 480p ($0.10 per second) before a batch.

When the still is a Sume avatar, pass avatar_handle and skip these checks. The range is for stills you supply.

A quick rule for batches: compute the ratio in code before you upload, and log any file outside 0.4 to 2.5. Rejecting a bad image locally is faster than waiting for a job to fail, and a failed job still costs you a round trip.