liveBook · Manning

Image size

As explained in Local Image Generation

Image size is the width and height configuration that determines the spatial dimensions of a generated image.

Image size is the Width and Height used to generate an image. It sets the pixel dimensions and, more importantly than the sliders might suggest, helps determine the frame’s aspect ratio and can change what the model produces. Think of it as choosing the canvas shape before asking for the painting. section 1.0 section 1.2.4

Figure 1.10 The options for setting our Width and Height in the UI.

Set image size with the Width and Height controls. A square 512×512 starting point is useful, but changing either value changes the composition’s available space. section 1.2.4

Use image size as a composition control, not merely a resolution dial:

  • Choose Width and Height to fit the intended orientation: 640×384 produces a 5:3 landscape frame, while swapping values makes a portrait frame. section 1.2.4
  • Use supported dimensions. Stable Diffusion generally requires each dimension to be a multiple of 8; multiples of 128 are a practical rule of thumb, and A1111 limits inputs to working values. section 1.2.4
  • Watch total pixels: Width × Height increases VRAM demand quickly. Large direct generations can fail with CUDA out-of-memory errors. section 1.2.4 section 7.1.2
  • Keep the seed fixed when comparing sizes, so the visible differences can be attributed to the changed dimensions rather than a different random result. section 1.2.4

A wider coffee image

Start with 512×512 images, then set Width to 640 and Height to 384 for a 5:3 landscape result. Generate several images with seed 1337, then return to 512×512 with that same seed. The comparison shows that size changes can alter generated content as well as make the image rectangular. section 1.2.4 section 1.2.4

Do not assume that a larger direct render is a better high-resolution image. Beyond VRAM limits, base Stable Diffusion was trained on 512×512 images, so a 2048×2048 request may yield repeated or tiled subjects. A practical high-resolution workflow begins with a smaller image at the target aspect ratio, then upscales it—for example, 856×360 as the starting point for a 3440×1440 result. section 7.1.2 section 7.1.3

Panel dimensions as layout tools

Give each graphic-novel panel its own prompt, height, and width: 800×1024 for a tall panel, 320×1800 for a wide panel, or 800×400 for a narrow one. Storing those values with each prompt lets a generation loop produce panels that support the intended page layout. section 3.5 section 3.5

How Local Image Generation covers it

How deep it goes
Explains UI settings, valid multiples, aspect ratio, VRAM costs, and why direct high-resolution generation can fail or tile.
Examples
Demonstrates 512×512 coffee images, 640×384 landscapes, 856×360 upscale sources, graphic-novel panels, and a 752×1000 ControlNet scene.
What it assumes
Teaches size controls from basic image generation while introducing aspect ratio and VRAM as needed.
What you can do after
Apply it via A1111 or Diffusers to select supported dimensions, control framing, and prepare images for upscale workflows.
How it connects
Connects image size with aspect ratio, seed, VRAM, upscaling, Stable Diffusion, and ControlNet.

Also explained in

Compare how 3 books explain image size →

Related concepts

This explainer was generated by Manning's knowledge graph from the text of Local Image Generation. The section links point to the parts of the book it draws on.