Overview
Resolution and batch size are configured in the Empty Latent Image node.
For SDXL, the recommended starting size is 1024 by 1024. The image does not have to be square, but SDXL is treated as optimized around that scale. For Stable Diffusion 1.5 or earlier, the recommended size is 512 by 512. Lower dimensions can be used to accelerate an experiment; the ControlNet example uses 768 instead of 1024 to reduce generation time.
Batch size determines how many images are produced in one queue. A batch size of two creates two images. Later, a batch of four is used with an image chooser so that only the preferred candidate proceeds to upscaling.
The seed identifies the starting pattern of random noise. Randomize produces a different starting point for each generation. Fixed preserves the seed. When the seed and every other setting remain identical, the same image is generated again. This is essential when a workflow is paused after the first image and resumed later for upscaling; without a fixed seed, the second queue would generate and upscale a different image.
Steps determine how many rounds of noise removal occur. Too few steps leave the image noisy or incomplete. Increasing the count improves the result only up to a point. Beyond approximately 50 to 100 steps, the image may stop improving and can acquire noise or artifacts because little useful noise remains to remove.
A conventional setup can use 20 steps. A Lightning checkpoint needs far fewer: approximately five to eight steps, with some models capable of a usable image in as few as two. Using 20 steps merely because it is a familiar default ignores the checkpoint's design.
CFG controls how strongly the sampler follows the prompt. A very low value, such as 1, gives the generation more freedom and may not follow the prompt closely. A high value, such as 15, follows the wording very literally. Excessive literalness can create strange results. A value around 7 or 8 is a reliable general starting point.
Sampler name selects the algorithm that removes noise. Euler is a common and fast option. DPM++ 2M and DPM++ 2M SDE are suitable when quality is the priority. Scheduler adds another subtle variation to the sampling process. Karras and Exponential are suggested for higher-quality output, although the visible differences can be subtle and should be tested for the intended use case.
Denoise determines how much of the starting latent content is replaced. In text-to-image, the input is random noise, so denoise is set to 1.0 to remove 100 percent of that starting noise. In image-to-image, the same parameter becomes the main control over similarity to the uploaded image.