Test SDXL Framing with Different Canvas Shapes
Academy is temporarily unavailable while it is being rebuilt.
A common SDXL problem is asking for a tall portrait or wide landscape on a canvas that does not fit the subject. The result may be a cropped head, a duplicated figure, empty space in the wrong place, or repeated texture bands. One useful first test is to compare canvas shapes used by SDXL before adding more corrective words. Canvas shape can influence framing, but it is not the only cause of cropping or repeated subjects.
Start by deciding the frame, not the pixel count. For a clothed standing portrait, test a vertical canvas such as 832x1216 or 768x1344. For a close portrait or centered object, 1024x1024 provides a square comparison. For a landscape scene, try 1216x832 or 1344x768. These are examples near the original SDXL model’s one-megapixel training area, with different aspect ratios. Check the requirements of your particular checkpoint and workflow before using them.
In a ComfyUI workflow, set the width and height on the Empty Latent Image node if you are building the image from noise. Keep the prompt unchanged for the first comparison. For example, use the same prompt for a clothed adult hiker on a mountain path, then generate it at 1024x1024, 832x1216, and 1216x832. Changing only the canvas makes it easier to see whether the issue was framing rather than wording.
Check success by looking for three visible signs: the main subject fits naturally inside the frame, the background supports the composition instead of fighting it, and repeated subjects or texture bands are absent. Those artifacts can have several causes, so record what changed without assuming the dimensions explain every defect. If a vertical version fits the requested figure better than the square version, keep that canvas shape for the next test.
Only after this test should you adjust prompt detail, sampler settings, or a refiner pass. Those controls address different parts of an SDXL workflow and may also affect composition. Comparing them separately keeps the cause of a change easier to investigate. A controlled resolution comparison gives you a clear next action instead of guessing.