AI & Compute

Video Resolution Calculator

Find video and image generation sizes divisible by your model's required multiple (8, 16, 32 or 64) for a target aspect ratio — or check whether a resolution you already have divides cleanly.

Most image and video diffusion models require both output dimensions to be divisible by a fixed number — LTX-Video requires 32, Wan 2.1/2.2 enforce 16. This calculator answers the two questions that come with it: what sizes does my model actually accept, and will the size I have work? It ships with the rules of the popular open video models built in (see model presets) — pick one, or use a custom divisor for anything else.

Find a valid size

Pick your model/app from the presets — it fixes the divisibility rule, with a note explaining where that rule comes from — or choose Custom divisor… for manual control. Then pick an aspect ratio (buttons). The result has two lists:

  • Exact — sizes whose width and height are both multiples of the divisor and exactly match the ratio. These are solved, not scanned: for 16:9 at ÷32 they come out as 512×288, 1024×576, 1536×864 … because only every 32nd multiple of the base ratio lands on the grid.
  • Close (within 1 %) — sizes that are divisor-clean but land slightly off the target ratio, listed by how far off. When the exact list skips the size you want (it usually will at ÷32), these are what you render and crop or letterbox later.

Clicking any row loads it into the check tab so you can see its full verdict.

Check a resolution

Enter any width × height and the calculator reports:

  • whether it is a known standard size (1080p, QHD, 4K UHD …) or the closest one by pixel count
  • the simplified aspect ratio and megapixel total
  • which powers of two it divides cleanly by, dimension by dimension
  • an explicit pass/fail against the common “divisible by 32” requirement, with the nearest divisor-clean frame that keeps your size

Model presets and their rules

PresetDimension ruleFrame ruleWhere the rule comes from
LTX-Video÷32 (both dims)8n+1 framesOfficial model card; off-grid input is padded then cropped
Wan 2.1 (T2V / I2V)÷16 (both dims, hard error)4n+1 framesdiffusers WanPipeline.check_inputs raises unless both dims are divisible by 16; official examples (832×480, 1280×720) are all ÷16-clean
Wan 2.2 (TI2V-5B / A14B)÷16 (both dims, hard error)4n+1 framesSame diffusers check; the Wan2.2 VAE compresses 16×16×4 and the official “720p” size is 1280×704 — 704 = 44×16, chosen to sit on the grid
MiniMax H3 (FL2VA / Ref2VA)÷32 (both dims, required)Current Diffusers integration requires height/width multiples of 32 (effective spatial downsampling is 32× = VAE 16× + 1×2×2 patchification); shorter side defaults to 768. See the MiniMax H3 guide

The check tab applies the same rule: its pass/fail line and nearest-valid-frame suggestion follow whichever preset you have selected.

Why it matters

Diffusion models downsample in fixed-size patches. The VAE alone typically requires dimensions divisible by 8; latent patchification usually adds a factor of 2, making 16 the common minimum for image models. Several video models (and their reference-image / audio pipelines) require 32 or even 64.

The non-obvious part: standard display sizes frequently fail these rules. 1080p fails ÷32 because 1080 = 33.75×32; 4K UHD fails the same way (2160 = 67.5×32). QHD 1440p, by contrast, passes cleanly (1440 = 45×32) — which is why “1440p-ish” sizes are common defaults in video-model tooling. Passing an invalid size does not always fail cleanly: you may get a crash deep in the VAE, silently cropped output, or distorted aspect ratio — all of which are wasted render time. Checking takes five seconds.

Examples

  • 1280 × 832, divisor 32 → valid (1280 = 40×32, 832 = 26×32). Divides cleanly by up to ÷64; the safe ~16:9 default for video models.
  • 1920 × 1080, divisor 32 → invalid on height (1080 = 33.75×32). The check tab’s nearest clean frame is 1920 × 1088 — or step down to the exact 16:9 size 1536 × 864 from the find list.
  • 1000 × 1000, divisor 64 → neither dimension qualifies; nearest clean frame is 1024 × 1024.

FAQ

“Divisible by 16” — width and height separately?

Yes. Both dimensions must satisfy the constraint independently. A size like 1920×1080 passes for width (120×16) but fails on height (67.5×16).

Should I round up or to nearest?

To nearest keeps aspect ratio closest to your target; rounding up guarantees you never go below a minimum model size. For strict minimums, prefer rounding up and re-checking the aspect you can live with.

Does the divisor affect quality?

No — it is a hard constraint of the architecture’s patch grid, not a quality setting. Quality scales with total pixels and your training data’s distribution, so stay near sizes the model was trained on.

Sources and upstream documentation

  • LTX-Video model card — states that resolutions must be divisible by 32 (frames divisible by 8+1).
  • Wan2.1-T2V-14B and Wan2.2-TI2V-5B — official model cards (Wan2.2 VAE compression 16×16×4; TI2V 720p is 1280×704).
  • diffusers pipeline_wan.pycheck_inputs requires height and width divisible by 16 (shared by the 2.1 and 2.2 pipelines); (num_frames − 1) must be divisible by the VAE’s temporal factor (4).
  • MiniMax-H3 model card — visual latents patchified 1×2×2 for an effective 32× spatial downsampling (temporal 4×); shorter side defaults to 768.