ltdrdata / bsz-cui-extras

Addons for ComfyUI, mainly the Principled SDXL node

bsz-cui-extras

Addons for ComfyUI

Custom Nodes

To install all the nodes, simply put the entire bsz-nodes folder inside your custom_nodes folder in ComfyUI

To install specific nodes, you may put individual .py files from bsz-nodes directly into the ComfyUI custom_nodes folder.

__init__.py simply forwards all nodes within its folder to ComfyUI, and is not necessary if you're putting nodes directly into custom_nodes

bsz-auto-hires.py

Contains 3 nodes each with a different means to the same end result. These nodes are designed to automatically calculate the appropriate latent sizes when performing a "Hi Res Fix" style workflow.

Common:

• Input
  • base_model_res : Resolution of base model being used. SD 1.5 ≅ 512, SD 2.1 ≅ 768, SDXL ≅ 1024
• Output
  • Lo Res Width : Width intended to be used for first/low res pass
  • Lo Res Height : Height intended to be used for first/low res pass
  • Hi Res Width : Width intended to be used for final/high res pass
  • Hi Res Height : Height intended to be used for final/high res pass

BSZAbsoluteHires:

• Input
  • desired_width : Width in pixels for final/high res pass.
  • desired_height : Height in pixels for final/high res pass.

BSZAspectHires:

• Input
  • desired_aspect_x : Horizontal aspect.
  • desired_aspect_Y : Vertical aspect.
  • scale : Hi Res horizontal and vertical scale over Lo Res sizes. Note that because this scales both axes, a scale of 2.0 will actually quadruple the amount of pixels in an image, so use with care.

BSZCombinedHires:

A unique node that functions both as BSZAbsoluteHires and BSZAspectHires with a convenient toggle

• Input
  • use_aspect_scale : Use aspect & scale inputs instead of desired width/height inputs

bsz-principled-sdxl.py

All-in-one solution for SDXL text2img, img2img and scaling/hi res fix. Essentially the sdxl and sdxl-upscale workflows both in one node. Do note that while this node shouldn't be any slower than the regular workflow, due to ComfyUI caching latent results per-node, even changing just a refiner setting on this node will result in sampling starting over from the first base pass. There are at least some minimal internal optimizations to skip passes that aren't needed.

Scaling works by running initial scaling passes before running the final pass at target size

Input fields

• base_model : Model from base checkpoint
• base_clip : CLIP from base checkpoint
• latent_image : Latent image to start from
• refiner_model : Model from refiner checkpoint. Optional
• refiner_clip : CLIP from refiner checkpoint. Optional
• pixel_scale_vae : VAE used for pixel scaling methods. Optional, only needed if they're being used
• positive_prompt_G : Positive prompt for base CLIP G and refiner
• positive_prompt_L : Positive prompt for base CLIP L. Usually either set to the same as CLIP G but sometimes is used for supporting terms
• negative_prompt : Negative prompt
• steps : Steps for non-scaled pass
• denoise : Denoise amount for latent input
• cfg : CFG scale
• refiner_amount : Refiner to base ratio. Requires refiner model and refiner clip to function
• refiner_ascore_positive : Refiner aesthetic score for positive prompt
• refiner_ascore_negative : Refiner aesthetic score for negative prompt
• sampler : Sampler
• scheduler : Scheduler
• scale_method : If set, will scale image to match target sizes using the provided algorithm
• scale_target_width : If scale_method is enabled, image will be resized to this
• scale_target_height : If scale_method is enabled, image will be resized to this
• scale_denoise : Denoise amount for scaled passes
• scale_steps : Steps for scaled passes
• scale_iterations : Amount of scaling passes to run. Experimental and very expensive
• vae_tile : Whether to used tiled vae during pixel scaling
• seed : Seedy.

Recommended settings for various workflows...

• Text2Image: default
• Text2Image w/latent upscale:
  • scale_method:latent bicubic
• Text2Image w/pixel upscale:
  • scale_method : pixel bicubic
  • scale_denoise : 0.15
  • vae_tile : encode if scaling to a very large resolution
• Img2Img w/upscale:
  • Same as text2img upscaling
  • steps : 0

bsz-pixelbuster.py

Node that loads the Pixelbuster library. Requires either pixelbuster.dll (windows) or libpixelbuster.so (linux) to be placed in your custom_nodes folder alongside bsz-pixelbuster.py Input fields

• image : Image[s] to work on
• code : Pixelbuster code. See the help for reference

Workflows

sdxl.json

Personal flair of the SDXL "partial diffusion" workflow. Minimalist node setup with defaults balanced approach to speed/quality

Dependencies

• bsz-auto-hires.py : While this workflow doesn't actually perform any upscaling, it still uses the BSZAutoHiresCombined node for quick aspect ratio changing and easy CLIP detail target adjustments

sdxl-upscale.json

Personal flair of the SDXL "partial diffusion" workflow with added "High res fix". Slightly prioritizes speed as far as upscaling is concerned.

Dependencies

• bsz-auto-hires.py : Workflow is painful without it.

sdxl-principled.json

Demonstration of the bsz-principled-sdxl node

Dependencies

• bsz-auto-hires.py : Principled can use hi res sizes
• bsz-principled-sdxl.py : Yes.

F.A.Q.

Question	Answer
Why is there a separate VAE loader instead of using the VAE directly from the main checkpoint?	I personally find it desireable to have the VAE decoupled from the checkpoint so you can change it without re-baking the models. If this isn't desirable to you yourself, simply remove the Load VAE node and reconnect the traces into the main Load Checkpoint node instead.
Why are the KSAmpler nodes so long?	To show live previews of each stage. I strongly recommend you do the same by launching ComfyUI with --preview-method latent2rgb or similar.
Why is this setting the default instead of that setting?	It just happens to look better on my benchmark images. If you think it's objectively wrong, open an issue with a compelling case on why it should be changed.
You should add the refiner 1 step detail trick	No. That "trick" really just causes the refiner to interpret latent noise as "details" it should refine, which hurts the overall image quality. If you render an image and really really think it needs it, just load the most recent history item and adjust the refiner steps as needed. ComfyUI caches the previous latents so you won't have to re-render the whole image, just the part that changed.