H3 Max 1080p regenerates from 768p: the price of the extra step

H3 Max does not render 1080p natively. Sume's docs describe 1080p as a latent refinement from a native 768p pass, and it costs twice as much per second: $0.20 versus $0.10 after the 1.25 multiple.

MiniMax's own H3 page describes its higher-resolution path the same way in spirit: the base model regenerates its own low-resolution output in context instead of applying a conventional upscaler (MiniMax H3 announcement, read 2026-10-05). That is why the 1080p row is a priced step and not a free export setting.

What the docs say

The Video Router row for minimax-h3-max lists 480p, 768p and 1080p for 5 to 15 seconds. Native resolutions are 480p and 768p; 1080p is a latent refinement from 768p; there is no 2K or 4K. The fal list is $0.05, $0.08 and $0.16 per second, and Sume bills list x 1.25.

When the doubling is worth it

Refinement costs $0.10 more per second than the native 768p take. For a single 10-second hero clip that is one dollar more. For a batch of fifty candidates it is fifty dollars more. The practical pattern is to generate candidates at 768p and refine only the take you will ship.

Because refinement is a regeneration pass, the 1080p result can differ from the 768p one in fine detail. Check first frames and faces on the 1080p output before you trim or caption it. A refined clip is a new job with its own price, not a view of the earlier job.

• Spend the 768p price on picking the take, not on picking the pixels.
• Spend the 1080p price only on the clip you will deliver.
• Do not assume the 768p seed and the 1080p result match frame for frame.

If you need larger than 1080p

H3 Max has no 2K/4K row on Sume. The sibling minimax-h3 lists priced 2K and 4K upscales when a request includes them, at fal list $0.13 and $0.16 per second, or $0.1625 and $0.20 after the multiple. A 10-second 2K H3 clip rounds up to $1.63, which is under a 10-second 1080p H3 Max clip at $2.00. Confirm in the live catalog that the resolution is still offered on the row you call.

A decision table for one delivered clip

The point of the table below is the total spend to deliver one 10-second 1080p clip after n exploratory takes. Exploration at 768p on Max costs $1.00 per take; the final refinement is $2.00 once.

• Going straight to 1080p for every take at $2.00 would make ten takes $20.00.
• Whether a refined final matches the approved 768p take is a quality question; check it, do not assume it.

What to check on a refined clip

A refinement pass regenerates detail, so the checks are visual and specific. Pull the first frame and a mid-clip frame from both the 768p take and the 1080p result and compare faces, hands, small text on packaging and logos. If any of those drift, you have two choices: re-run the 1080p job, which bills another $0.20 per second, or deliver the 768p take and upscale it outside the generator.

Sound is part of the output too. The docs say H3 Max produces native stereo audio with no toggle, so the refined clip carries audio on its own. Listen to the refined clip before you deliver it, and do not assume it matches the 768p take.

• Compare frame 1 and a mid-clip frame at 100% zoom.
• Check any on-pack text and logos first; they drift most.
• Log usage.billable_amount_usd_micros from each submit so the refinement shows up as its own line in your spend sheet.