ComfyUI SDXL Workflow
Workflow Name
sdxl
Workflow Description
This workflow is intended as a basic workflow for use with SDXL checkpoints.
Default Workflow Dependencies
Download sd_xl_base_1.0 SDXL checkpoint, save in ComfyUI\models\checkpoints:
https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0/blob/main/sd_xl_base_1.0.safetensors
Workflow Details
CheckpointLoaderSimple
Purpose: This node loads a Stable Diffusion checkpoint file, which contains the trained weights for the model, CLIP (text encoder), and VAE (variational autoencoder). It’s the starting point for defining the base model used for image generation.
Customizable Settings: You can select the checkpoint file to load. Changing this setting will fundamentally alter the style and subject matter the model is capable of generating, as different checkpoints are trained on different datasets and may specialize in various artistic styles or subject domains.
EmptyLatentImage
Purpose: This node generates an empty latent image, which serves as the initial canvas for the image generation process. The latent image is a lower-dimensional representation of the image that the diffusion model operates on.
Customizable Settings: You can specify the width and height of the latent image. These dimensions influence the final image resolution. A larger width and height will result in a higher resolution image, but will also require more processing power and time. The batch size determines how many images are generated at once. A higher batch size will generate more images, but will also increase memory usage.
PrimitiveNode (Positive Prompt)
Purpose: This node provides the positive prompt, which is a text description of the image you want to generate. It guides the image generation process by describing the desired content and style.
Customizable Settings: You can enter any text prompt. The more descriptive and specific the prompt, the better the model can understand what you want. Varying the prompt will directly change the subject, style, and composition of the generated image.
PrimitiveNode (Negative Prompt)
Purpose: This node provides the negative prompt, which is a text description of elements you want to exclude from the generated image. It helps refine the generation by specifying unwanted features or artifacts.
Customizable Settings: You can enter any text prompt. Common negative prompts include terms like “blurry,” “distorted,” or specific unwanted objects. Changing the negative prompt allows you to fine-tune the image by removing undesirable elements.
CLIPTextEncode (Positive Prompt)
Purpose: This node takes the positive prompt text as input and encodes it into a format that the Stable Diffusion model can understand. It uses the CLIP text encoder to convert the text into a conditioning vector.
Customizable Settings: You can input the text prompt. Changing the prompt will alter the desired content and style of the generated image.
CLIPTextEncode (Negative Prompt)
Purpose: This node takes the negative prompt text as input and encodes it into a format that the Stable Diffusion model can understand. It uses the CLIP text encoder to convert the text into a conditioning vector representing what should be excluded from the image.
Customizable Settings: You can input the text prompt. Changing this prompt will refine the image generation by specifying what to avoid.
KSampler
Purpose: This is the core image generation node. It performs the iterative denoising process of the diffusion model, guided by the positive and negative conditioning vectors, and starting from the initial latent image.
Customizable Settings: You can customize the seed, steps, cfg scale, sampler name, scheduler, and denoise. The seed controls the random number generation used in the process, so different seeds will produce different images even with the same prompts. The steps setting determines how many denoising steps are performed; more steps usually lead to higher quality but take longer. The cfg scale controls how closely the generated image adheres to the prompt; higher values mean stronger adherence, but can also lead to artifacts. The sampler name selects the algorithm used for the denoising process. The scheduler determines how the noise is added and removed during sampling. The denoise setting controls the amount of denoising to apply; a value of 1 means full denoising.
VAEDecode
Purpose: This node takes the latent image generated by the KSampler and decodes it into a pixel-based image using the VAE decoder. This converts the latent representation into a viewable image.
Customizable Settings: This node requires the latent image and the VAE model. The VAE model is loaded by the Checkpoint Loader node, and it is responsible for the decoding process. Changing the VAE model will affect the image quality and details.
SaveImage
Purpose: This node saves the generated image to a file.
Customizable Settings: You can specify the filename prefix. This allows for organizing and identifying generated images. Changing the prefix allows for batch saving and easier identification.
Credits
This workflow has been modified from the ComfyUI workflow:
https://comfyanonymous.github.io/ComfyUI_examples/sdxl/
