Why can only resolutions that are multiples of 8 be generated?

Looking at the Empty Latent Image node in ComfyUI, you can only set width/height in increments of 8. Even if you force input like 513px, the output is always rounded to a multiple of 8.

This is caused by structural constraints of latent diffusion model and VAE.

Latent diffusion model and VAE compression rate

Latent diffusion models such as Stable Diffusion compress images into a low-resolution latent space once before calculation. VAE is responsible for this "compression" and "restoration".

In commonly used VAEs (SD1.5 series and many derivative models), the structure is:

  • Compresses height/width to 1/8 (1/64 in area)

At this time, since the latent resolution must be an integer, fractions indivisible by 8 are truncated.

For example, even if you input an image of 513px, the fraction of 1px is ignored and processed as effectively 512px.

There are also VAEs with other compression rates

Not all VAEs have 8x compression, but vary by model, such as 4 / 8 / 16 / 32.

When using a new model, it is a good idea to check once which VAE (which compression rate) that model uses.

Video model VAE and frame count

In VAEs for video, there are those that compress the time direction (number of frames) together.

  • Wan 2.1 VAE: VAE that summarizes 4 frames into one latent

In this case, the number of frames that can be generated becomes a form like "1 + 4n", and a structure where detailed specification such as outputting only 2 frames is impossible in the first place may be adopted.

Just as the spatial direction (vertical/horizontal) is bound to multiples of 8, please remember that there may be constraints that "can only be handled in multiples of this number" in the time direction as well.

Points to note for image2image / image editing models

Even if you use a resolution other than a multiple of 8, generation is possible without error. However, please note that misalignment occurs in that case.

  • Input Image: Original resolution (e.g. 513px × 769px)
  • Output Image: Resolution rounded to a multiple of 8 (e.g. 512px × 768px)

Since input and output pixels do not correspond 1-to-1, if pixel-perfect editing is required, it is necessary to crop/pad to a multiple of 8 first.

If you feel "somehow the edge is cut" or "size is slightly different", please check once if the resolution is a multiple of the VAE's compression rate.