ComfyUI 9:16 resolution: exact sizes for Qwen, Krea, MOSS

A 9:16 Reel is 1080 by 1920, and no image model in ComfyUI can generate that size directly: 1080 is not a multiple of 16, and the model makers publish their own vertical presets instead. Pick the wrong size and you get stretched faces, a stray second figure or a crop that cuts the head off in the profile grid.

This guide gives the exact vertical size for Qwen-Image 2.1, Krea 2 and MOSS-Image, the one size that scales to 1080 by 1920 with no stretching, and where to put the face so Instagram's own interface does not cover it. Every preset below is copied from the model maker's own page; the arithmetic is ours and labelled.

Why 1080 by 1920 is not a size you can generate

Comfy's MOSS-Image tutorial says the model accepts any aspect ratio in multiples of 16, so width and height must each divide by 16. 1920 does (120 times 16). 1080 does not: 1080 divided by 16 is 67.5. The nearest legal widths are 1072 and 1088, both slightly off 9:16.

Both official vertical presets in this guide have the same small quirk. Qwen's 1536 by 2752 and Krea's 768 by 1376 each work out to a ratio of 0.558, while true 9:16 is 0.5625. That is a 0.8% difference, invisible in a Reel, but it means the preset is not a clean 1080 by 1920 scale-up.

The pattern is 144 by 256 multiplied by a whole number. Any of those sizes is exactly 9:16 and passes the multiple-of-16 rule, and 864 by 1536 is the one that lands on 1080 by 1920 with a clean 1.25 factor.

The vertical size for each model

Qwen-Image 2.1 ships a dictionary of sizes on its model card, and the 9:16 entry is 1536 by 2752. The same dictionary lists 1:1 at 2048 by 2048, 3:4 at 1792 by 2400 and 2:3 at 1696 by 2528. Comfy's tutorial says the model generates natively at 2K and tells you to set the megapixel target to about 4.0 for a 2048 square. For a Reel, 1152 by 2048 is a lighter and cleaner choice, but it is our pick, not a Qwen preset, so check a few seeds before you commit a batch to it.

Krea 2 lists 768 by 1376 for 9:16 and labels its resolution 1K. Its other portrait presets are 928 by 1152 (4:5) and 832 by 1248 (2:3). There is no 3:4 row, which matters if you plan for the profile grid (below).

MOSS-Image has no preset list. The tutorial says it is trained for 1 to 2 megapixels and extrapolates from about 0.25 to 4 megapixels. That puts 864 by 1536 (1.33 MP) and 1008 by 1792 (1.81 MP) inside the trained range, and 1152 by 2048 (2.36 MP) just past it. Stay inside the range for faces you plan to reuse.

krea.aiOpen the page

docs.comfy.orgOpen the page

Set it up in ComfyUI, step by step

Where the face goes on a 1080 by 1920 frame

Hopper HQ's 2026 guide, which attributes the figures to Meta's published safe zones, says to keep key content out of the top 14%, the bottom 35% and 6% on each side, leaving roughly 950 by 980 pixels. It also says the profile grid crops every thumbnail to 3:4, which is 1080 by 1440. Other guides we found give different margins, so treat these as a working rule, not a spec.

In practice the face sits in the upper middle of the frame: about 14% to 65% of the height. Prompt for it by name. A request such as "full body, head in the upper third, space below the knees" does more than any resolution tweak.

None of the three reported a 9:16 vertical test at the sizes above, so the thresholds are a warning about aspect ratio, not a limit for portrait. Reddit threads were not reachable from our research environment, so these voices come from blogs and Hugging Face.

Our verdict

Use 864 by 1536 as the default vertical for Krea 2 and MOSS-Image: it is exact 9:16, it is divisible by 16, it sits in the trained 1 to 2 megapixel range, and it scales to a Reel with one 1.25 resize. Use 1152 by 2048 for Qwen-Image 2.1 when you want the sharper source, and keep the official 1536 by 2752 for when you need the 2K look and have the VRAM. Our picks are arithmetic, not vendor presets, so run five seeds of your character at the new size before you switch a whole feed.

When you run these models for a recurring character, the size is one of the settings that should never drift between images. Musetta runs on your own ComfyUI and keeps each character's reference photos, so a vertical request starts from the same approved face; try the size in your own graph first, since the guide does not need Musetta.

Musetta studioDemo data

For the video side of the same Reel, see the open video models for vertical reels, and for the models behind these sizes the MOSS-Image review and the Krea 2 vs Qwen-Image 2.1 comparison. If a node is missing when you open a workflow, the missing nodes guide covers it.