ComfyUI_MaraScott_nodes is an extension set designed to improve the readability of ComfyUI workflows and enhance output images for printing. It offers a Bus manager, an Upscaler/Refiner set of nodes for printing purposes, and an Inpainting set of nodes to fine-tune outputs.
AnyBus Node
- V2 - Dynamic Bus up to 25 inputs/outputs with profile management
- V1 - Static Bus
McBoaty set
- V5 - Improved Per Tile Prompt edition adding Image and Denoise on each tile & allowing the use of McBoaty as the first step in the refining process
- V4 - Introduced Per Tile Prompt Editor via an Upscaler, a TilePrompter, and a refiner node
- V3 - Introduced Dynamic Tiling & assisted prompt generation via LLM
- V2 - Converted @TreeShark initial Upscaler/Refiner workflow to a Node
- V1 - Initially a workflow from @TreeShark, you can watch where everything started
McInpainty set
- V1 - Introducing a 2-node set: to generate an inpainted output and to paste it back to the original image
Installation [method1] via ComfyUI-Manager (recommended)
- Click the "Manager" button on the main menu.
- Click the "Custom Nodes Manager" button in the Menu.
- Search for "MaraScott" and click the "install" button.
- Restart ComfyUI.
To install ComfyUI_MaraScott_Nodes
in addition to an existing installation of ComfyUI, follow these steps:
- Go to
ComfyUI/custom_nodes
directory in the terminal (cmd). git clone https://github.com/MaraScott/ComfyUI_MaraScott_Node.git
- Restart ComfyUI.
- v5.0.0 - McBoaty: Improved Tile prompts editor
- v4.4.1 - McBoaty: Tile prompts editor
- v4.3.0 - McBoaty: Hotfix tiling ksampling issue on non-standard ratio
- v4.2.8 - McBoaty: Added color match feature
- v4.2.3 - McBoaty: McBoaty Dynamic Tiling
- v4.0.0 - McBoaty: Updated workflow PNG images
- v3.3.0 - McBoaty: Added size feature, hidden iteration mechanism, log improvements
- v3.2.9 - McBoaty: Patched add_loop_system_back
- v3.2.8 - McBoaty: Patched node_versioning_plus_upscaler_v2
- v3.2.7 - McBoaty: Patched add_image_size_output
- v3.2.4 - McBoaty: No hotfix for slow node
- v3.2.3 - McBoaty: Added Sigmas_type management
- v3.2.2 - McBoaty: Cleaned code
- v3.2.1 - McBoaty: Implemented VAEEncode/DecodeTiled
- v3.1.0 - AnyBus: Hotfix setNode issue
- v2.2.0 - AnyBus: Added console log if 2 flows have the same names
- v2.1.0 - Updated readme for delivery, updated max input number, updated web flow to hotfix initial load from DnD flow
- v2.0.0 - Release of obsolete Bus_node + UniversalBusNode (py)
- v1.3.0 - AnyBus: Added details to build Detailer Pipe
- v1.2.0 - AnyBus: Updated node description and dynamic name
The AnyBus Node (AKA UniversalBus Node) is designed to provide a Universal Bus Node based on AnyType Input/Output.
Native Support
You can assume that any node input/output works like "UE Nodes / Use Everywhere" Nodes.
BUS Input/Output Support
Bus Nodes can be connected from one AnyBus Node to another, and we support the following Reroute Nodes as long as they are linked from an AnyBus Node before connecting to another.
Set/Get Node Supported:
- "SetNode" from ComfyUI-KJNodes
- "GetNode" from ComfyUI-KJNodes
Reroute Node Supported:
- "Reroute (rgthree)" from RgThree Custom Node
Reroute Node Not Supported:
- "Reroute" from ComfyUI native - reason: does not allow connection on Any Type if the origin Type is not Any Type too
- "ReroutePrimitive|pysssss" from Pysssss Custom Node - reason: does not display the name of the origin slot
- "0246.CastReroute" from 0246 Custom Node - reason: undefined behavior
Here is a very simple workflow:
Any workflow can quickly become messy and difficult to identify easily which output goes to which input. At some point, either you use a custom_node which "hides" the splines for you so you don't feel overwhelmed but will still have some issue identifying which goes where OR you use a Bus node like AnyBus which will clarify your workflow without having to remember the origins of your splines.
The AnyBus Node allows you to apply a profile to your Buses to organize paths in the same workflow. One Bus has a maximum of 25 inputs/outputs.
The BusNode profile is a mechanism to synchronize BusNodes in the same flow (connected by the Bus
input/output). All nodes with the same profile name and connected will be synchronized if one of them is modified.
When adding a node, the profile is default. If you have another flow called main, for example, and you connect the bus output of the flow to the input of the default busnode, the main flow will synchronize the main input to the default one and change the profile name to main. This works only for the default node. If you have a node called TextFlow and try to connect main to TextFlow, the connection will be refused. This allows for multiple flows in the same workflow and avoids conflict.
This AnyBus is dyslexia-friendly :D
McBoaty Node Set (AKA Upscaler, Prompter, Refiner Node set) is an upscaler coupled with a refiner to achieve higher rendering results on a per-tile basis.
McBoaty intends to provide the most efficient way to upscale and refine an image for printing purposes. Its core purpose is to upscale an image with minimal to no modification on the upscaled/refined image. For that purpose, using a denoise at 0.27 or maybe 0.35 is the sweet spot. Using multiple x2 upscaler models would also provide a better result than using one x4 model, but that is for you to choose.
It has been tested successfully with SD1.5, SDXL, SD3; it will probably work with any SD version. Feel free to report issues in the issue section.
The McBoaty set of nodes comes with 4 different nodes:
- An Upscaler Node to upscale the input image and slice it into tiles to refine the image on a per-tile approach.
Inputs:
- image: input image
- model, clip, vae, seed: as usual
- positive, negative: required but not used, will be replaced by positive and negative prompts in the near future
- upscale model, output upscale method: as usual with an Upscale (by model) node
- tile size, feather mask: these 2 sizes will be used to slice your upscaled image and define the size of the image the Ksampler will need to refine
- vae encode, tile size vae: do you want to use a tiled vae encoding method and which size
- color match method: do you want to apply a color match from your input image to the final output image after the refining process
- tile prompting active, vision llm model, llm model: currently used to generate the prompt for conditioning, not accurate and will be improved in the future
Outputs:
- McBoaty Pipe: to connect to Refiner input only
- McPrompty Pipe: to connect to TilePrompter input only
- info: Obsolete
- A TilePrompter Node to edit the prompt and denoise level on a per-tile basis.
Input:
- pipe: (McPrompty) Pipe output from Upscaler, Refiner, or LargeRefiner
Output:
- McPrompty Pipe: Pipe to connect to Refiner input pipe_prompty only
- A Refiner Node to refine the image based on the settings provided, either via general settings if you don't use the TilePrompter or on a per-tile basis if you do use the TilePrompter.
Inputs:
- pipe: McBoaty Pipe output from Upscaler, Refiner, or LargeRefiner
- McPrompty Pipe: to connect to TilePrompter input only
- tiles to process: you can specify an empty input if you want to process all the tiles or specify a tile index shown on the TilePrompter to refine only that tile. The input can specify multiple tiles at once as well either by specifying the tile indexes one by one (
1,2,3,4,...
) or using a range notation likex-y
(1-4
,5-8
) or a combination of those1,3,4-6,8,9
- output size type: choose between your Upscale size which will output an image of the size of your Upscale model or use custom size which will use the output size parameter
- output size: this is a coefficient to apply to your input image size as output size if you pick custom size type
- sigmas type: choose between BasicScheduler, SDTurboScheduler, and AlignYourStepScheduler
- ays model type: choose between SD1, SDXL, SD3, SVD (SVD has never been tested) even if you choose a scheduler different from the AYS, you still need to choose your input model type
- sampler name, basic scheduler, steps, cfg, denoise: as usual on a Ksampler node
- control net name, strength, start percent, end percent: as usual using an apply control net node
- low threshold, high threshold: as usual using a canny node
Outputs:
- McBoaty Pipe: to connect to Refiner input only
- McPrompty Pipe: to connect to TilePrompter input only
- image
- image (original)
- tiles
- prompts
- info
- The fourth node is the LargeRefiner Node which is actually the combination of the Upscaler Node + the Refiner Node in one node using the general settings.
Inputs:
- inputs from Upscaler + Refiner
Outputs:
- outputs from Refiner
Below is an example of the LargeRefiner being connected to the TilePrompter and refined by a second pass to a Refiner.
Using the Upscaler instead of the LargeRefiner will require using the 1st pass of the refiner with tiles to process at empty value to process all tiles at least once (which the LargeRefiner does by default).
The overall idea is to Upscale (/Refine) your image and slice it to provide each tile to McPrompty the TilePrompter which will allow you to fine-tune your prompt on a per-tile basis as well as the denoise level. Then those parameters are passed to the Refiner which will process those tiles and rebuild your image for the output. Then you can continue the refining on another refiner pass or modify the current TilePrompter parameters to fine-tune the result.
One use case is to push the denoise to 0.76 (maximum denoise recommended with control net activated) in the LargeRefiner then fine-tune the prompt and denoise for the tiles which went too wild and process the tiles you want (specified in the tiles to process settings) again through a second refiner pass.
Below is a functional workflow embedded in the image which allows you to test both techniques.
Learn More about the Name McBoaty
McInpainty Node is a set of 2 nodes.
Img Slider for result comparison
Custom_node | Execution Time from server start to end of queue | output | Type | workflow | OS Name | Processor | Installed Physical Memory (RAM) GB | GPU Name |
---|---|---|---|---|---|---|---|---|
McBoaty LargeRefiner v5 | 98.37 | ![]() |
Upscaler | benchmark_McBoaty.json | Microsoft Windows 10 Home | Intel(R) Xeon(R) W-3235 CPU @ 3.30GHz | 41 | Red Hat QXL controller NVIDIA Quadro RTX 6000 |
Supir | 70.48 | ![]() |
Upscaler | benchmark_SUPIR.json | Microsoft Windows 10 Home | Intel(R) Xeon(R) W-3235 CPU @ 3.30GHz | 41 | Red Hat QXL controller NVIDIA Quadro RTX 6000 |
UltimateSDUpscaler | 47.34 | ![]() |
Upscaler | benchmark_USDU.json | Microsoft Windows 10 Home | Intel(R) Xeon(R) W-3235 CPU @ 3.30GHz | 41 | Red Hat QXL controller NVIDIA Quadro RTX 6000 |
Visit this link to see an Image Slider comparing the 3 results against the original
I would like to thank my AI Classroom Discord buddies with a shoutout to:
- @Fern to whom I address A special thanks for his support in my ComfyUI journey who accepted me into his discord in the first place and tried to put good ideas in my head
- Rob Adams AKA Treeshark who helped improve McBoaty by providing great insights and always pushing further the limits,
- @YouFunnyGuys From the discord channel for his invaluable contribution and exceptional testing skills,
- Comfyanonymous with his ComfyUI
- DrLtData with his Custom nodes Manager,
- WASasquatch with his Was Node Suite
- Kijai with his Get/Set nodes,
- Trung0246 with his Highway node,
- Jags with his support
- Rob Adams AKA Treeshark who provided the workflow in the first place which I converted into a ComfyUI Node, you can find his original workflow presentation in his YouTube video: Advanced UpScaling in ComfyUI,
- DrLtData with his ImpactPack,
- Kijai with his ColorMatch node
- Chflame163 with his LayerStyle Nodes
I might have forgotten some other people, please contact me if you want to appear here and please forgive me.
If anybody is interested in using these nodes, I'm following up on feature requests via the issues. If you have Node ideas, feel free to make a request in the issues.