ComfyUI-Toggle-Pass Introduction
ComfyUI-Toggle-Pass v1.0.0 is a collection of ComfyUI nodes for building image-generation and editing workflows. It brings common tasks together: loading a model and LoRAs, preparing reference images, choosing a sampling setup, comparing results, and controlling when images are saved or previewed.
The extension is especially useful when your canvas has become crowded. For example, you can load a model’s main components in one node, use a single sampler instead of a chain of several nodes, and let the loader temporarily connect its outputs to empty model inputs when you run a workflow. You can still make manual connections wherever you need a different setup.
Other nodes help you experiment without rebuilding a graph: switch between two canvas sizes, try several prompts with one image, use multiple reference images, or create a side-by-side comparison of your originals and results.
How ComfyUI-Toggle-Pass Works
A ComfyUI workflow passes information from one node to another through connections. ComfyUI-Toggle-Pass adds nodes that either prepare that information or make it easier to route and inspect.
- A model loader prepares the image model, text encoder, and VAE. Think of these as the parts that generate an image, interpret your prompt, and turn the generated result into a viewable picture.
- A reference encoder prepares an image so a compatible editing model can use it alongside your written instructions. Some nodes give each reference its own output; another combines multiple references under one prompt.
- A sampler controls the image-generation process. Its settings include the number of steps, the sampling method, and how strongly guidance is applied.
- Utility nodes handle practical tasks such as image loading, previews, comparisons, and choosing between two output sizes.
These nodes do not make every model interpret references in the same way. The model type, text encoder, VAE, and reference-image arrangement still need to match your intended workflow. The sections below explain the choices that matter most.
ComfyUI-Toggle-Pass Features
Unified model and LoRA loading
ToggleUnifiedLoader outputs MODEL, CLIP, and VAE from one node. It offers two loading modes:
| Mode | Use it when |
|---|---|
| UNet (separate components) | You have separate UNet, CLIP, and VAE files and want to select each one. |
| Checkpoint (all-in-one) | You have a checkpoint containing those components together. The VAE comes from that checkpoint. |
The mode is selected with a two-part switch near the top of the node. In checkpoint mode, the separate-component choices are hidden. Existing saved workflows retain their parameter values, and a workflow without a saved mode defaults to separate components.
You can add up to 10 LoRAs. Selecting a LoRA reveals one strength control; selecting none hides it. LoRAs are applied in sequence. After a run, the Add LoRA button displays a short status, and hovering over it shows more detail. A patch count means the LoRA was applied to model weights, not that its effect must be obvious in the final picture. Warnings can indicate a missing file, zero strength, an incompatible LoRA, or a mismatched text encoder.
The loader also has a global broadcast switch. When enabled, it supplies its MODEL, CLIP, and VAE outputs to matching empty inputs elsewhere in the graph as the prompt is built. Manually connected inputs remain untouched. Violet curves preview the planned connections on the canvas, but these are not permanent workflow wires. Broadcast is off by default.
Broadcast works when you run from the ComfyUI interface or export API-format data from that interface. It does not automatically add connections to a hand-written request sent directly to
/prompt.
Saving, loading, and previewing images
| Node | What it does |
|---|---|
| SaveLoadImage | In Save mode, saves an image to the output directory and provides IMAGE and MASK outputs. In Load mode, reuses the last saved image without saving again. |
| ToggleLoadImage | Loads an image with functionality comparable to ComfyUI’s built-in LoadImage node. |
| TogglePreviewImage | Shows a preview using temporary storage rather than saving to the output directory. |
Each of these nodes has a privacy mask controlled by the eye icon in its title bar. Use it to hide or show the image preview on your canvas.
Switching image size
ToggleEmptyLatent creates an empty latent, the starting canvas for generation. Its switch chooses between two size settings:
- ON: use
widthandheight. - OFF: use
fallback_widthandfallback_height.
This makes it convenient to compare two output sizes without repeatedly editing the same fields. You must also choose a latent_type that matches the base model:
| Latent type | Intended models |
|---|---|
| SD / SDXL | SD 1.5, SDXL, Pony |
| SD3 / Flux / Krea 2 | SD3, Flux.1, Krea 2 |
| Qwen-Image 2.1 | Qwen-Image 2.1 |
| Flux2 | FLUX.2 |
The default is SD / SDXL to preserve older workflows. A mismatched type might not produce an immediate error, so check this setting when changing models.
Working with reference images
DynamicRefImageEncode is for several reference images guided by one shared prompt. Set num_images to show the image inputs you need. It outputs CONDITIONING, a starting LATENT, and a ready-made empty-text NEGATIVE conditioning. Connect that negative output to your sampler’s negative input, particularly if you may increase CFG above 1. Do not substitute ConditioningZeroOut applied to the positive conditioning: the original readme reports corrupted results at higher CFG.
DynamicRefIndependent is for references that need separate prompts and separate outputs. Choose between 1 and 10 images, connect each image, and write its corresponding prompt. The node creates a conditioning and latent output for each reference. Use it when references should remain independently controllable, rather than forming one shared reference sequence.
OneImageMultiPrompt takes one image and 2–8 prompts. It produces a separate conditioning for each prompt while sharing one latent. This is useful for trying several edits, viewpoints, or descriptions of the same source image.
The reference nodes show settings appropriate to the selected model type. Supported choices described in the supplied readme include Flux2, Qwen Image 2.1, Qwen Layered, SD / SDXL / Z-Image, and, on applicable nodes, Krea2. For Qwen Image 2.1, additional controls include aspect ratio, output megapixels, reference-image resolution, and whether references keep their proportions, are cropped, or are stretched. Megapixels is an image-area setting: a value around 1.0 corresponds roughly to a 1024 × 1024 image; do not enter 1024 in that field.
DynamicRefLatentInput adds or removes latent input slots for workflows using the dynamic reference nodes. ImageBatcher similarly adds or removes image inputs to batch multiple images.
Z-Image: The supplied readme identifies the generic SD / SDXL / Z-Image choice as appropriate for text-to-image use. Leave reference-image inputs empty for a text-to-image-only model; connecting a reference where the model has no reference path can produce an unwanted mosaic.
Comparing results
ToggleImageCompare combines originals and generated images into one comparison image, previewed inside the node. You can enable 1–10 source slots and 1–10 result slots independently, so one result can be compared with several sources or each source can be paired with a result.
Choose automatic, paired, stacked, or grid layout. Automatic layout pairs images when the connected source and result counts match; otherwise it stacks the groups. You can show or hide labels on the tiles. The layout responds to images actually connected, and images with different dimensions retain their proportions when arranged. The comparison output can go to a save or preview node.
Sampling in one node
ToggleSampler combines the functions of ComfyUI’s noise, guidance, sampler-selection, scheduling, and advanced-sampling nodes. Connect your model, positive and negative conditioning, and starting latent; the node outputs the sampled latent and a denoised latent.
Its ⚡ preset control fills in the usual sigma curve, sampler, scheduler, step count, and CFG for a selected setup. It does not change denoise or seed, and it does not permanently lock your choices. If you adjust a preset-filled value, the control returns to manual mode.
The Sigma curve setting separately controls the noise schedule. You can calculate it automatically from steps and scheduler, select a supplied fixed table, or enter a custom table. For a custom table, values must decrease strictly, end in 0, and contain one more value than the desired number of steps.
| Preset | Steps | CFG |
|---|---|---|
| Qwen-Image 2.1 | 25 | 1.0 |
| Qwen-Image 2.1 turbo | 8 | 1.0 |
| Krea 2 turbo | 8 | 1.0 |
| Krea 2 turbo, detail | 12 | 1.0 |
| FLUX.2 klein | 4 | 1.0 |
| FLUX.2 klein, upscale refine | 4 | 1.0 |
| Z-Image turbo | 8 | 1.0 |
| Qwen-Image Layered | 9 | 2.5 |
| Qwen-Image Edit 2511 AIO | 4 | 1.0 |
| FIBO 2.0 turbo | 4 | 1.0 |
| MOSS-Image | 30 | 4.0 |
The presets also set the appropriate sampler and scheduler listed in the original readme. For distilled turbo models, it recommends euler and CFG 1.0. Treat a preset as a starting configuration for its named model, not as a substitute for matching the model, encoder, and VAE.
Choosing a pose reference
PoseReference lets you browse a pose-sheet library and output the selected sheet as an ordinary IMAGE. It requires no ControlNet or additional pose model. Choose a category, group, pose, and view; the choices narrow as you go, and the node shows a thumbnail without requiring a generation run. You can switch between line-art and depth sheets, set the output’s long edge, or flip the pose horizontally.
The node can use bundled sheets and a personal library. The supplied readme describes ComfyUI/models/pose_reference/ as the location for your own sheets; a TP_POSE_ROOT setting can point to another collection. Your sheets take priority over bundled sheets with matching names. A selected sheet can also be embedded in the workflow so someone opening it can still see and use that selection without having the local library.
For editing a photo of a person with a pose sheet, the placement of images matters:
- With editing encoders such as
TextEncodeQwenImageEditPlus, put the source photo inimage1and the pose sheet inimage2. - With Ref Image Encode set to Krea2, the supplied readme specifies the opposite reference order: pose sheet in
image1, subject photo inimage2.
Describe the desired pose in the prompt rather than writing only “follow the pose sheet.” Specify the subject, body position, limb placement, and viewing angle, then name what should stay unchanged, such as face, hair, clothing, and background. The original readme reports that vague instructions can cause a model to render the pose-sheet figure as another subject.
ComfyUI-Toggle-Pass Models
ComfyUI-Toggle-Pass does not introduce a single new image-generation model. Its nodes provide settings and connections for different model families. The following distinctions from the supplied readme are particularly important:
| Model or family | What to check |
|---|---|
| Qwen-Image 2.1 | Select its model type in a compatible reference node. It has dedicated output-size and reference-size controls; changing reference resolution does not change output resolution. In separate-component loading, use the appropriate qwen_image CLIP type so reference images reach the intended text encoder. |
| Krea2 | Select Krea2 for a compatible reference encoder and use a matching VAE. Its reference-image workflow has specific input-order and sizing rules. |
| FLUX.2, MOSS-Image, FIBO 2.0 | The supplied readme says their reference workflows can use the Flux2 latent choice, but MOSS-Image and FIBO 2.0 have specialized editing controls available through their native encoders. |
| Qwen Layered | Set the number of layers when using a reference node that offers this choice. It produces a layered starting latent rather than an ordinary single-image latent. |
| SD / SDXL / Z-Image | Use the appropriate generic choice for the workflow. For text-to-image-only Z-Image use, leave reference-image inputs unconnected. |
VAE compatibility deserves special attention. The supplied readme distinguishes Qwen-Image 2.1’s 64-channel VAE from Krea2’s 16-channel VAE. Mixing them can let sampling proceed but cause decoding to fail. If you switch model families, recheck the VAE as well as the model and CLIP selections.
What’s New with ComfyUI-Toggle-Pass
The provided readme identifies the extension as v1.0.0, but does not provide a dated version history or identify which features arrived in separate releases. The features documented for this version include unified loading, LoRA status reporting, optional global broadcast, dynamic reference-image nodes, an all-in-one sampler with model presets, image comparison, privacy-masked image nodes, and pose-reference browsing.
Troubleshooting ComfyUI-Toggle-Pass
A LoRA appears to have no effect
- Run the workflow and check the short status on Add LoRA.
- Hover over the button for the per-LoRA breakdown.
- If it reports a missing file, check the selected LoRA. If it reports zero strength, raise the strength.
- If no weights matched, check that the LoRA was made for your base model.
- If weights were applied but the picture barely changes, remember that patching confirms the LoRA was loaded, not that its visual effect will be strong in every image.
Automatic model connections do not appear
- Confirm global broadcast is on in ToggleUnifiedLoader.
- Check whether the target
MODEL,CLIP, orVAEinputs are already connected. Broadcast fills empty inputs only. - Run or export the workflow through the ComfyUI interface. A hand-written direct
/promptrequest does not pass through the front-end broadcast step. - If multiple loaders have broadcast enabled, use only the one you intend to supply each input type; the first enabled loader wins for that type.
A reference image is ignored or produces an unwanted result
- Confirm that the selected reference-node model type matches your base model.
- For Qwen-Image 2.1 in separate-component mode, check the CLIP type; choose
qwen_imageexplicitly if you suspect automatic detection failed. - For a text-to-image-only Z-Image workflow, disconnect reference images.
- For a pose edit, check which image input should contain the photo and which should contain the pose sheet for your chosen encoder.
- Rewrite vague pose instructions to describe the actual body position and what must remain unchanged.
The sampler finishes, but VAE decoding fails
- Check the selected VAE against the base model.
- In particular, do not interchange the Qwen-Image 2.1 and Krea2 VAEs.
- In separate-component mode, also check that the CLIP type and files belong to the same model family. The loader’s compatibility warnings can help identify a mismatch.
Increasing CFG corrupts a reference-image result
- If you use DynamicRefImageEncode, connect its
NEGATIVEoutput to the sampler’s negative input. - Do not create the negative by applying
ConditioningZeroOutto the positive reference conditioning. - For DynamicRefIndependent or OneImageMultiPrompt, use a plain empty
CLIPTextEncodefor the negative input when CFG is above 1.
Frequently asked questions
Do I need a separate pose model to use PoseReference?
No. PoseReference outputs an image for a compatible editing workflow; it does not require ControlNet or an additional pose model.
Will global broadcast overwrite the connections I drew myself?
No. It supplies only matching inputs that are empty.
Does hiding a preview remove or change the image?
The eye control is described as a privacy mask for showing or hiding the preview. It is not an image-editing setting.
Can I compare different numbers of originals and results?
Yes. ToggleImageCompare lets you set source and result counts independently, from 1 to 10 each.
Why are some sampler settings greyed out?
A fixed sigma preset already determines its step count and noise schedule. Switch back to automatic or manual settings when you need to control those values yourself.
Learn More about ComfyUI-Toggle-Pass
The supplied readme names a ready-made workflow at example_workflows/Pose Reference Editing.json. It demonstrates pose-reference editing with Qwen-Image 2.1 turbo and carries its selected pose sheet inside the workflow, making it a practical example to inspect on the ComfyUI canvas.
For familiar points of comparison within ComfyUI, the extension’s checkpoint mode corresponds to CheckpointLoaderSimple, while ToggleSampler combines operations otherwise handled by nodes such as RandomNoise, CFGGuider, KSamplerSelect, BasicScheduler, and SamplerCustomAdvanced. Comparing their connections with this extension’s nodes can help you understand what each consolidated control is doing.
