ComfyUI-Bricolage Introduction
ComfyUI-Bricolage is a personal collection of ComfyUI nodes for experimenting with prompts, image sizes, and local AI models. It includes small tools that can pick an artist name, vary parts of a prompt, combine text, create an SDXL-sized starting point for an image, or send a request to a model running in LM Studio.
The author keeps these utilities together for small workflows, sometimes adapting or repackaging ideas from other custom node packs. This is an experimental repository: the author provides no guarantees, support commitment, or roadmap. Some controls may change as the author continues experimenting.
How ComfyUI-Bricolage Works
In ComfyUI, a workflow is a graph of connected nodes. Each node takes an input, performs one task, and passes its result to another node. ComfyUI-Bricolage adds nodes for several common creative tasks:
- Vary a prompt: Simple Wildcards picks words or phrases from choices you write. RandomArtist picks a name from a bundled SDXL artist list.
- Build a prompt: Text Concat Simple joins text you type with text produced by another node.
- Start an SDXL image: Random SDXL Latent creates the latent input used by a KSampler and reports the corresponding image dimensions.
- Work with a local language model: LM Studio sends a prompt to a locally running LM Studio server. Connecting an image makes it a vision request; leaving the image input disconnected makes it a text-only request.
Think of these as individual tools on a workbench. For example, you could use a wildcard to vary the setting in a prompt, join that result with a fixed description, and use Random SDXL Latent to vary the canvas shape. What each node changes depends on where you connect it in your workflow.
ComfyUI-Bricolage Features
RandomArtist
RandomArtist selects one name from the bundled SDXL artist list each time the node executes. Its filter controls which entries are eligible:
any: choose from the full list.photo: choose entries containingphoto.illustration: choose entries containingillustration.
Use photo or illustration when you want to narrow the kinds of names that appear in an exploratory prompt. The node deliberately reports itself as changed on every execution so it can make a fresh selection, even when its settings stay the same. The selected name is a prompt ingredient; its effect on an image depends on the rest of your workflow and the image model you use.
LM Studio
The LM Studio node connects ComfyUI to a locally running LM Studio (https://lmstudio.ai) server through its OpenAI-compatible API. It can send either a text-only request to a language model (LLM) or an image-and-text request to a vision-capable model (VLM).
| Input | What it does |
|---|---|
model | Identifies the model to use. Copy the identifier shown in LM Studio's model list; the node does not supply a default. |
prompt | Provides the user message sent to the model. |
system_prompt | Optionally provides instructions before the user message, such as the kind of response you want. |
load_model / unload_model | Explicitly loads and unloads the model around each call. These controls may be useful when video memory (VRAM) is limited. |
image | Optionally attaches an image for a vision request. Leave it disconnected for a text-only request. |
For example, you might send a written idea as prompt to a text model, or connect an image and ask a vision-capable model to describe it. Choose a model in LM Studio that can handle the type of request you are sending.
The node expects the LM Studio server at http://localhost:1234 by default. The server address is set by LM_STUDIO_BASE_URL in src/lm_studio.py; pointing it to another host or port requires changing that setting.
Random SDXL Latent
Random SDXL Latent creates a random latent image ready to connect to a KSampler. A latent is the working image data used during generation, rather than a finished picture you can view directly.
The node randomly chooses from nine supported SDXL dimension combinations, including square, portrait, and landscape shapes. Each combination has an equal chance of being selected. It provides three outputs:
LATENT: the starting latent for a KSampler.WIDTH: the chosen image width in pixels. Possible values are 1024, 1152, 1216, 1344, and 1536.HEIGHT: the chosen image height in pixels. Possible values are 1024, 896, 832, 768, and 640.
This is useful when you want to explore how a composition works across different canvas shapes without choosing a size manually each time. Connect the width and height outputs to other nodes that need to know the selected dimensions.
Text Concat Simple
Text Concat Simple joins two pieces of text into one output:
textis the text you type into the node.incoming_textis text supplied by another node.incoming_firstcontrols the order. Turn it on to putincoming_textbeforetext; leave it off to put it after.
For instance, you could connect a varying subject description as incoming_text and type a fixed scene description into text. Change incoming_first depending on which part you want to appear first in the resulting prompt.
Simple Wildcards
Simple Wildcards adds chance-based variation to text you write. Put alternatives inside braces and separate them with vertical bars:
a { red | blue } ball in a { big | small } room
Each time the text is expanded, the node chooses one option from each wildcard. One possible result is:
a blue ball in a big room
You can use multiple wildcards in multiline text. Each wildcard is resolved independently, so changing one choice does not determine the others. Nested wildcards are not supported; keep each set of alternatives inside its own single pair of braces.
DisplayString Status
The original readme lists DisplayString as a crossed-out control. It was intended as a lightweight, read-only way to show string output in the ComfyUI graph, but it should not be treated as a currently documented feature of this pack. The author noted that ComfyUI's built-in Preview As Text node largely covers that use case, while also observing that its text did not remain visible after reloading a PNG in the author's test.
ComfyUI-Bricolage Models
ComfyUI-Bricolage does not list or bundle a set of LLM or VLM models. For the LM Studio node, you choose a model available in LM Studio and enter its identifier in model. Use a text-capable model for a request without an image, or a vision-capable model when you connect the image input.
RandomArtist instead draws names from a bundled SDXL artist list. That list supplies text choices; it is not a collection of image-generation models.
What's New with ComfyUI-Bricolage
The provided readme records two dated notes about the crossed-out DisplayString control. It does not provide a broader version history for the other nodes.
- May 21, 2026: The author discovered that ComfyUI already includes Preview As Text, which largely covers DisplayString's original purpose. This helps you identify an existing ComfyUI option for displaying text.
- May 22, 2026: The author reported that Preview As Text also did not retain its displayed text after reloading a PNG in the author's test. The readme does not document a completed solution to that behavior.
Troubleshooting ComfyUI-Bricolage
The LM Studio node cannot reach a model
- Check that the LM Studio server is running.
- Check that
modelexactly matches the identifier shown in LM Studio's model list. - Check whether the server is using the node's default address,
http://localhost:1234. If it is running elsewhere, the address insrc/lm_studio.pymust match. - If you connected an image, check that the selected model supports vision requests. Otherwise, disconnect
imageand try a text-only request.
The LM Studio request uses too much video memory
Try the load_model and unload_model controls, which are provided for workflows that need to manage limited VRAM around each call. Their usefulness will depend on the model and the resources available on your computer.
A wildcard does not produce the variation you expected
Check that the alternatives are inside one pair of braces and separated by |, as in { red | blue }. Each block makes one choice independently. Do not place one wildcard inside another.
RandomArtist changes even though you did not edit the node
That is intentional. RandomArtist reports itself as changed on every execution so it can select a new name.
Frequently Asked Questions
Does Random SDXL Latent output a finished image?
No. It outputs a latent for a KSampler, plus the width and height of the selected size.
Can I choose an exact canvas size with Random SDXL Latent?
This node is described as choosing randomly from nine supported SDXL dimension combinations. The provided controls do not describe an exact-size setting.
Do I need to connect an image to the LM Studio node?
No. With image disconnected, it sends a text-only request. Connect an image when you want to send a vision request to a suitable model.
Can Simple Wildcards handle nested choices?
No. It resolves multiple separate wildcard blocks, but not wildcards nested inside other wildcards.
Will displayed text remain visible when I reload a PNG?
The author's May 22, 2026 note says that ComfyUI's built-in Preview As Text did not retain the text in the author's test. The provided readme does not document a fix in ComfyUI-Bricolage.
Learn More about ComfyUI-Bricolage
For the LM Studio node, LM Studio (https://lmstudio.ai) is the relevant resource for the local model application it communicates with. You can also look at ComfyUI's built-in Preview As Text node when you need to display text in a graph, bearing in mind the PNG-reload limitation recorded by the author.
