How do I toggle a workflow stage on and off in ComfyUI?

Select the node and press Ctrl+B. The node is removed from execution and wires reconnect through it. Press Ctrl+B again to re-enable.

Does ComfyUI have a Bypass Node in the right-click context menu?

The documented bypass is Ctrl+B. A right-click bypass may exist in recent UI builds but is not in the official shortcuts table.

Modular ComfyUI: One Graph for txt2img, img2img, Upscale

Q: Can one ComfyUI graph handle txt2img, img2img, AND upscale?

Yes. Feed KSampler an EmptyLatentImage for txt2img, a VAEEncode latent with denoise < 1.0 for img2img, and chain UpscaleModelLoader → ImageUpscale after the first decode for upscaling. All three share the same checkpoint, CLIP, and VAE nodes.

Q: Is ComfyUI Modular Node System an official feature name?

No. ComfyUI's tagline is 'the most powerful and modular diffusion model GUI,' referring to its node-graph architecture. There is no product or feature called 'Modular Node System.'

Q: How do I load a workflow JSON into ComfyUI?

Drag the .json file onto the canvas, or use Workflows → Open (Ctrl+O). The 'Load (custom nodes)' menu entry installs Python custom node packages, not workflow files.

Q: Are IPAdapter and ControlNet nodes built into ComfyUI?

ControlNetApply and ControlNetApplyAdvanced are built-in. IPAdapter is a third-party custom node — install it via ComfyUI Manager or GitHub.

Q: What does the KSampler denoise parameter actually do?

At 1.0 (default), the sampler treats the input latent as pure noise. Below 1.0, it blends the input latent's structure with the denoised output — lower values preserve more of the original image. This is the img2img mechanism.

Q: How do I make a truly reusable node group?

ComfyUI core has no sub-graph/module system. Options: (1) save a sub-graph as a .json template and import manually, (2) write a custom Python node that encapsulates the logic, (3) copy-paste nodes between workflow tabs.

Build a single ComfyUI graph for txt2img, img2img, and upscale using bypass, KSampler denoise, and reusable templates.

Updated 2026-07-27 6 sources cited ComfyUI

On this page

TL;DR
Why Modular Workflows Matter
The Toggle Mechanism: Bypass, Don't Delete
One Graph, Three Modes
Optional Conditioning Branches
Grouping: Visual Only, Not Reusable
Settings Reference Table
Common Errors and Fixes
FAQ
Sources
Internal Links

Modular ComfyUI Workflows: One Graph for txt2img, img2img, Upscale

TL;DR

Toggle any stage with Ctrl + B. Bypassed nodes stay on the canvas but are removed from execution.
One graph, three modes: EmptyLatentImage for txt2img, VAEEncode + denoise < 1.0 for img2img, UpscaleModelLoader → ImageUpscale for upscaling.
Group for clarity, not reuse: Ctrl + G collapses visually. For true reuse, use .json templates or a custom node.

Why Modular Workflows Matter

ComfyUI workflows are directed acyclic graphs. Every task — load model, encode prompt, sample, decode, upscale — is a discrete node with typed inputs and outputs. Because the graph is data-flow rather than imperative code, you can rewire, bypass, or duplicate branches without rewriting logic. A modular workflow keeps optional stages (ControlNet, IPAdapter, upscaler) in separate bypassable branches so the same .json serves text-to-image, image-to-image, and high-res fix without duplication.

The Toggle Mechanism: Bypass, Don't Delete

How bypass works

Select any node and press Ctrl + B. ComfyUI removes it from the execution graph and stitches inputs directly to outputs. The node stays on the canvas (dimmed) so you can re-enable it instantly with another Ctrl + B.

Source: ComfyUI README: "Ctrl + B — Bypass selected nodes (acts like the node was removed from the graph and the wires reconnected through)" [^1]

A right-click "Bypass Node" may exist in recent UI builds, but the documented mechanism is Ctrl + B.

When to bypass vs. delete

Scenario	Action
Temporary disable	`Ctrl + B`
Permanent removal	Delete and rewire manually
Swap model	Keep node, change loader input

One Graph, Three Modes

The pivot is KSampler. Its latent_image input and denoise parameter determine the generation mode.

txt2img (pure generation)

EmptyLatentImage → KSampler (denoise = 1.0) → VAEDecode

EmptyLatentImage creates pure noise at your target resolution. denoise = 1.0 means the sampler ignores input latent structure entirely.

img2img (guided transformation)

LoadImage → VAEEncode → KSampler (denoise < 1.0) → VAEDecode

VAEEncode compresses the input image to latent space. The KSampler docs say "lower values will maintain the structure of the initial image allowing for image to image sampling" [^2]. denoise < 1.0 preserves that structure; lower values stay closer to the source, higher values allow more creative deviation.

upscale (high-res fix)

KSampler → VAEDecode → UpscaleModelLoader → ImageUpscale → (optional) VAEDecode again

UpscaleModelLoader reads a model file from models/upscale_models/ (for example 4x-UltraSharp.pth). For a true high-res fix, insert a second KSampler after ImageScale → VAEEncode, matching the two-pass pattern linked in ComfyUI's examples [^1].

Optional Conditioning Branches

ControlNet (built-in)

ControlNetApply and ControlNetApplyAdvanced are built-in nodes. They take:

positive / negative conditioning
control_net model (loaded via ControlNetLoader)
preprocessed image: Canny, OpenPose, Depth, etc.
strength (0–10, recommended 0.5–1.5)

Output is modified conditioning. To disable: bypass ControlNetApply.

Source: ControlNetApply docs: "Controls the strength of network adjustments, value range 0~10. Recommended values between 0.5~1.5 are reasonable." [^3]

IPAdapter (third-party)

IPAdapter is not a built-in node. It comes from community custom node packs (for example comfyui-ipadapter). The typical chain:

LoadImage → IPAdapterEncoder → IPAdapterApply → conditioning

To disable: bypass IPAdapterApply. IPAdapter has no built-in on/off switch separate from bypass.

Grouping: Visual Only, Not Reusable

Ctrl + G creates a collapsible visual group. It helps readability but does not create a callable sub-graph, allow drag-and-drop reuse, or export a standalone component.

For true reuse:

Approach	How it works	Trade-off
Workflow templates	Save a `.json` with only the sub-graph; import via Workflows → Open	Manual wiring each time
Custom node (Python)	Write a node that internally builds the sub-graph	Requires Python, cleanest API
Copy-paste + regroup	`Ctrl + C` / `Ctrl + V` across tabs	Fastest for ad-hoc reuse

Source: ComfyUI README: "Ctrl + G — Group selected nodes" — grouping is a UI shortcut, not a module system [^1].

Settings Reference Table

Parameter	txt2img	img2img	Upscale (HR fix)	Source
KSampler `latent_image`	`EmptyLatentImage`	`VAEEncode(LoadImage)`	First pass: `EmptyLatentImage`; second pass: `VAEEncode(ImageScale)`	[^2]
KSampler `denoise`	`1.0`	any value below `1.0`	First pass: `1.0`; second pass: below `1.0`	[^2]
ControlNet `strength`	`0.5–1.5` if used	`0.5–1.5`	Usually off for upscale pass	[^3]
Upscale model	—	—	example `4x-UltraSharp.pth` in `models/upscale_models/`	[^4]
Checkpoint loader	`CheckpointLoaderSimple`	`CheckpointLoaderSimple`	shared across all passes	[^5]

Common Errors and Fixes

Symptom	Cause	Fix
"Load > Load (custom nodes)" not found	Custom-node installation is not workflow loading	Use Workflows → Open (Ctrl+O) or drag `.json` onto the canvas [^6]
IPAdapter node missing	IPAdapter is not built-in	Install a custom node pack via ComfyUI Manager
ControlNet does nothing	Preprocessor missing or wrong image format	Install `comfyui-controlnet-preprocessors`; match preprocessor to input type
UpscaleModelLoader shows no models	Model file not in `models/upscale_models/`	Place `.pth`/`.onnx` there, then refresh or restart [^4]
img2img ignores input image	`denoise` left at `1.0`	Set `denoise` below `1.0` [^2]
Bypassed node still runs	Node not actually bypassed	Re-select node, press `Ctrl + B`, verify dimming/wire pass-through

FAQ

How do I toggle a workflow stage on and off? Select the node and press Ctrl + B. Press again to re-enable.

Can one ComfyUI graph handle txt2img, img2img, AND upscale? Yes. Feed KSampler an EmptyLatentImage for txt2img, a VAEEncode latent with denoise < 1.0 for img2img, and chain UpscaleModelLoader → ImageUpscale after decode for upscaling. All modes share the same checkpoint, CLIP, and VAE.

Is "ComfyUI Modular Node System" an official feature name? No. ComfyUI's tagline is "the most powerful and modular diffusion model GUI," referring to its node-graph architecture. There is no product or API called "Modular Node System."

How do I load a workflow JSON into ComfyUI? Drag the .json onto the canvas, or use Workflows → Open (Ctrl+O). The "Load (custom nodes)" menu installs Python custom-node packages, not workflows.

Does ComfyUI have a right-click "Bypass Node"? The documented bypass is Ctrl + B. A right-click option may exist in recent UI builds but is not in the official shortcuts table.

Are IPAdapter and ControlNet nodes built into ComfyUI? ControlNetApply and ControlNetApplyAdvanced are built-in. IPAdapter is a third-party custom node.

What does the KSampler denoise parameter actually do? At 1.0 (default), the sampler treats the input latent as pure noise. Below 1.0, it blends the input latent's structure with the denoised output — lower values preserve more of the original image.

How do I make a truly reusable node group? Save a sub-graph as a .json template and import manually, write a custom Python node, or copy-paste nodes across workflow tabs.

Sources

[^1]: ComfyUI GitHub README — shortcuts and architecture — https://github.com/Comfy-Org/ComfyUI [^2]: KSampler built-in node documentation — https://docs.comfy.org/built-in-nodes/KSampler.md [^3]: ControlNetApply built-in node documentation — https://docs.comfy.org/built-in-nodes/ControlNetApply.md [^4]: UpscaleModelLoader built-in node documentation — https://docs.comfy.org/built-in-nodes/UpscaleModelLoader.md [^5]: CheckpointLoaderSimple built-in node documentation — https://docs.comfy.org/built-in-nodes/CheckpointLoaderSimple.md [^6]: ComfyUI text-to-image tutorial (workflow loading) — https://docs.comfy.org/tutorials/basic/text-to-image