SeargeSDXL

Custom nodes for easier use of SDXL in ComfyUI including an img2img workflow that utilizes both the base and refiner checkpoints.

Install:

Recommended Installation:

• Navigate to your ComfyUI/custom_nodes/ folder
• Run git clone https://github.com/SeargeDP/SeargeSDXL.git
• Restart ComfyUI

Alternative Installation:

• Drop the SeargeSDXL folder into the ComfyUI/custom_nodes directory and restart ComfyUI.

Custom Nodes

SDXL Sampler Node

Inputs

• base_model - connect the SDXL base model here, provided via a Load Checkpoint node
• base_positive - recommended to use a CLIPTextEncodeSDXL with 4096 for width, height, target_width, and target_height
• base_negative - recommended to use a CLIPTextEncodeSDXL with 4096 for width, height, target_width, and target_height
• refiner_model - connect the SDXL refiner model here, provided via a Load Checkpoint node
• refiner_positive - recommended to use a CLIPTextEncodeSDXLRefiner with 2048 for width, and height
• refiner_negative - recommended to use a CLIPTextEncodeSDXLRefiner with 2048 for width, and height
• latent_image - either an empty latent image or a VAE-encoded latent from a source image for img2img
• noise_seed - the random seed for generating the image
• steps - total steps for the sampler, it will internally be split into base steps and refiner steps
• cfg - CFG scale (classifier free guidance), values between 3.0 and 12.0 are most commonly used
• sampler_name - the noise sampler (I prefer dpmpp_2m with the karras scheduler, sometimes ddim with the ddim_uniform scheduler)
• scheduler - the scheduler to use with the sampler selected in sampler_name
• base_ratio - the ratio between base model steps and refiner model steps (0.8 = 80% base model and 20% refiner model, with 30 total steps that's 24 base steps and 6 refiner steps)
• denoise - denoising factor, keep this at 1.0 when creating new images from an empty latent and between 0.0-1.0 in the img2img workflow

Outputs

• LATENT - the generated latent image

SDXL Prompt Node

Inputs

• base_clip - connect the SDXL base CLIP here, provided via a Load Checkpoint node
• refiner_clip - connect the SDXL refiner CLIP here, provided via a Load Checkpoint node
• pos_g - the text for the positive base prompt G
• pos_l - the text for the positive base prompt L
• pos_r - the text for the positive refiner prompt
• neg_g - the text for the negative base prompt G
• neg_l - the text for the negative base prompt L
• neg_r - the text for the negative refiner prompt
• base_width - the width for the base conditioning
• base_height - the height for the base conditioning
• crop_w - crop width for the base conditioning
• crop_h - crop height for the base conditioning
• target_width - the target width for the base conditioning
• target_height - the target height for the base conditioning
• pos_ascore - the positive aesthetic score for the refiner conditioning
• neg_ascore - the negative aesthetic score for the refiner conditioning
• refiner_width - the width for the refiner conditioning
• refiner_height - the height for the refiner conditioning

Outputs

• CONDITIONING 1 - the positive base prompt conditioning
• CONDITIONING 2 - the negative base prompt conditioning
• CONDITIONING 3 - the positive refiner prompt conditioning
• CONDITIONING 4 - the negative refiner prompt conditioning

Examples

Workflow

Result