WarpPipe Nodes

WarpPipe

ComfyUI nodes for people who switch models a lot and like their LoRAs in the prompt.

• Prompt + LoRAs: LoRAs are tags in the text, and colours tell you what will load before you queue.
• Prompt Writer: an LLM chat behind every prompt box. Each answer is a version you can compare and send back.
• Warp / Unwarp: a whole generation in one wire.
• Warp Send / Select: switch between model groups without moving a wire.
• Save Image (Civitai): metadata Civitai reads, so an upload credits the right model and LoRAs.

Install

comfy node install warppipe

Or clone into ComfyUI/custom_nodes and restart. Python 3.9+, no extra dependencies.

Prompt + LoRAs

<img src="assets/docs/prompt-node.webp" alt="Prompt + LoRAs with sections, trigger words, a LoRA list and the AI button" width="600">

The prompt goes on top, the LoRAs in the list below, one per line:

//// Subject
a red bicycle leaning on a stone wall by a lake, early morning mist
soft window light, film grain
//// Style
// a note: never reaches the model
35mm photo, muted colours <lora:old-polaroid:0.5>

<lora:film-grain-xl:0.8>
<lora:window-light:0.6>
// <lora:watercolor-wash:0.5>

| Colour | Means | | --- | --- | | cyan | the LoRA was found | | red | no file by that name | | orange | made for another base model | | green | trigger word of a LoRA that is on (dim when it is off) | | grey | a // note, or a LoRA switched off |

//// starts a section you can fold. / completes from your LoRA library, Tab on a tag adds its trigger words, Ctrl+/ switches lines on and off. A second list can feed model_2 for two-pass setups.

Prompt Writer

Click ✨ AI on any prompt box and say what you want, with reference images if you like. Every answer becomes a version. Edit it, write the next one from it, or Send to prompt. Show changes marks what is new and puts the word it replaced small above it.

• Works with NanoGPT out of the box, or any OpenAI-compatible server (LM Studio, Ollama, llama.cpp).
• The API key stays on the ComfyUI server (user/warppipe_llm_keys.json or NANOGPT_API_KEY). It never reaches the browser, a workflow or an image.
• Bring your own system prompts: a folder of .txt files, set under Settings → WarpPipe.
• It knows your LoRAs for the connected model and can suggest some. Nothing lands in your list until you press +.
• Chats are saved in the ComfyUI user folder and keep writing after you close the pop-up.

One wire

A Warp takes models, conditioning, latents, prompts and sampler settings. One link crosses the canvas, and an Unwarp gives it all back. Warps chain: one further down carries everything the first did and adds to it.

Switching model groups

flowchart LR
  subgraph A[Krea 2]
    L1[Loaders] --> W1[Warp] --> S1[Warp Send]
  end
  subgraph B[Z-Image Turbo]
    L2[Loaders] --> W2[Warp] --> S2[Warp Send]
  end
  S1 -. active .-> SEL[Warp Select]
  S2 -. off .-> SEL
  SEL --> U[Unwarp] --> K[Sampler] --> SAVE["Save Image (Civitai)"]

Put a Warp Send at the end of each model group and one Warp Select in the shared part. Press activate on a Send, or pick it in the Select, and only that group runs. Green runs, red doesn't. If the routing is broken (nothing active, the active Send muted, two Sends with one name) the queue says so instead of running.

The active Send shows the live preview while it samples (--preview-method auto) and a before/after slider when it is done. Right-click any image node and Use as before image to choose what goes on the left. The same slider exists as its own node, Warp Compare.

Save Image (Civitai)

Saves PNG or JPEG with the generation info Civitai reads. The Hashes field is what makes an upload link the checkpoint and every LoRA:

a red bicycle by a lake, early morning mist <lora:film-grain-xl:0.8>
Steps: 30, Sampler: dpmpp_2m, Schedule type: karras, CFG scale: 5.0, Seed: 44,
Size: 1024x1024, Model hash: 4c2b8e1f0a, Model: demo-sdxl,
Lora hashes: "film-grain-xl: 3b1f0c2a9d",
Hashes: {"model":"4c2b8e1f0a","lora:film-grain-xl":"3b1f0c2a9d"}, Version: ComfyUI-WarpPipe

It also shows what it saved next to the before image, in the same slider.

If you use Civitai Updater, WarpPipe reads its .civitai.info files for trigger words, base models and titles.

Example workflow

<img src="examples/workflow_example_screenshot.png" alt="The example workflow: two model groups feeding one shared pipeline" width="820">

Krea 2 and Z-Image Turbo in one workflow. The image below was made with it and has the workflow inside: drag it into ComfyUI to open it, or grab the JSON.

<img src="examples/workflow_example.png" alt="A pixel-art warp pipe made with the example workflow" width="260">

All nodes

| Node | | | --- | --- | | 🌀 Warp / Unwarp | Bundle a generation into one link and unpack it again | | 🌀 Warp Provider | Latent and sampler settings, with size presets for current models | | 🌀 Prompt + LoRAs | Prompt and LoRAs in one box | | 🌀 Save Image (Civitai) | Save with metadata Civitai reads | | 🌀 Warp Send / Warp Select | Pick which model group runs | | 🌀 Warp Compare | Before/after slider | | 🌀 Scheduler Adapter for FaceDetailer | KSampler schedulers for FaceDetailer | | 🚫 Dead End | Takes anything, does nothing. Parks a branch |

Each node has its own help behind the ? in ComfyUI. Everything else, every input and the HTTP routes are in the wiki.

Credits

Screenshots use demo LoRAs and stand-in images, not anyone's real models. Developed with AI assistance (Claude Opus 5). MIT License.