How to load a ComfyUI workflow, and read one node by node

To load a workflow in ComfyUI, drag its .json file, or a PNG that ComfyUI saved, onto the canvas, or choose Open in the workflow menu (Ctrl+O). Then check that every loader finds its file and press Run. Below, our face workflow read node by node: what each of the 10 nodes you can see does, and the custom node it needs. The RunPod guide covers getting ComfyUI running first.

How do you load a workflow in ComfyUI?

Three ways: drag and drop the file onto the canvas, open it from the workflow menu, or pick one from ComfyUI’s built-in templates. A workflow file is a .json, but a PNG that ComfyUI saved works too. After it loads, fix any missing models or nodes, then press Run.

1. Open ComfyUI. On a RunPod pod, connect to port 8188.
2. Load the file. Drag the .json (or a ComfyUI PNG) from your computer onto the canvas. Or open the workflow menu, choose Open, and pick the file. ComfyUI’s getting started page describes both (checked 2 Oct 2026). The menu label has moved between versions: look for Workflow or File.
3. Read the warnings. ComfyUI warns you if models are missing and lets you click the warning to see which. Red nodes mean a missing custom node.
4. Check each loader. Click every loader node’s file field and make sure it shows a file you actually have.
5. Set the prompt and seed, then press Run.

ComfyUI also ships ready-made workflows under Workflow Templates in recent versions. Those check for their models when you open them.

What’s the difference between a workflow file and an image with metadata?

A .json file is only the workflow: the nodes, the wires and every value. A PNG that ComfyUI saved is the finished image with that same workflow tucked inside its metadata. Both load the same way. The catch is that the metadata survives only while nobody re-saves the image.

ComfyUI’s workflow metadata docs say a saved image holds two things: the full workflow graph with its layout, and the prompt ComfyUI actually ran. They also warn that a file re-encoded by another app may lose it.

For anything you want to keep, export the .json. It’s small, it doesn’t depend on an image, and it’s what you’ll load on the next pod.

What does each node in a real workflow do?

Read any workflow left to right: loaders bring in the model files, model nodes adjust the model, the prompt is encoded, the sampler makes the image, the decoder turns it into pixels, and the last node saves it. Here is the face workflow from the AI Empire course, read that way. Values appear only where they explain the node.

The wires show the order. The model runs from the loaders through Load LoRA and ModelSamplingAuraFlow into KSampler. The prompt runs from Load CLIP through CLIP Text Encode into KSampler. KSampler’s result runs through VAE Decode into Save Image.

Every value in this workflow, and why it’s set that way, is in our Z-Image Turbo settings. This page is about reading the graph.

Why does a workflow fail on a standard ComfyUI install?

Because it uses something your ComfyUI doesn’t have: a model file, or a custom node. Our face workflow is an example. Its KSampler uses the FlowMatchEulerDiscreteScheduler, which isn’t built into ComfyUI. On an install without a pack that adds it, pressing Run gives “Value not in list: scheduler”, even when every model file is in place.

First open KSampler’s scheduler list: if FlowMatchEulerDiscreteScheduler is already there, there’s nothing to install. If not, the ComfyUI-EulerDiscreteScheduler node pack adds it to that list. Its page says to install it through ComfyUI Manager (search “erosDiffusion”) or by cloning it into the custom_nodes folder. Install it, restart ComfyUI, then load the workflow again. Our face template (the Portrait Generator) already has that pack built in (checked in its Dockerfile, 3 Oct 2026), so on our pod there’s nothing to install.

Which values can you change safely?

The prompt, the seed and the image size are yours to change. The sampler settings (shift, steps, CFG, sampler, scheduler) were set together for the model. Change one at a time, with a fixed seed, so you can see what it did. Changing several at once is how a working workflow turns into a broken one.

For a character with her own LoRA, our prompts start with her trigger word and a hair-and-eyes line, then the scene; the free prompt generator writes them in that format. Keep batch size at 1 and queue more runs: a bigger batch needs more VRAM, and that’s where out-of-memory errors start.

Where should you keep workflows when you use RunPod?

On your own computer. Workflows you save inside ComfyUI live on the pod’s disk, and terminating the pod deletes that disk unless it’s a network volume. Export each workflow you care about as .json and download it, the same way you download your images and LoRAs before you terminate.

1. Export. Open the workflow menu and choose Export. You get a .json file.
2. Name it clearly, for example face-zimage-turbo-1024x1536.json, so you know which is which next month.
3. Keep it with your LoRAs. One folder per character: her LoRA, her workflows, her best prompts.
4. Load it on the next pod by dragging it onto the canvas.

Can you use any workflow you find?

Use workflows from sources you trust and are allowed to use: the course’s templates, ComfyUI’s built-in templates, the official example pages, or ones you built. Don’t load or pass around someone’s paid workflow. Be careful with custom nodes: each one is code that runs on your pod. Ours may need one too: the scheduler pack above.

A workflow from an unknown source can ask you to install several custom nodes in one click. If you don’t know where it comes from, don’t install it, and never on a pod where your LoRA or your passwords live.

Hosted ComfyUI services only run custom nodes from a supported list; see Comfy Cloud vs RunPod.

What can go wrong?

• Red nodes after loading. A custom node type isn’t installed. Install it from a trusted source and restart ComfyUI, or use a workflow that doesn’t need it.
• “Value not in list: scheduler”. No installed pack provides FlowMatchEulerDiscreteScheduler. Install ComfyUI-EulerDiscreteScheduler and restart.
• “Value not in list” on a model field. A model file is missing or named differently. Our guide to ComfyUI missing models has the folder table and the z_image_turbo vs z_image_turbo_bf16 name trap.
• Nothing happens when you drop a PNG. The image has no workflow in it any more. Get the .json or the original file.
• Images save somewhere odd. Check Save Image’s filename_prefix. Ours still shows the folder path from the computer the workflow was built on. Set a plain prefix like z so images land in ComfyUI’s output folder.
• Every image looks the same. The seed is fixed. Set control after generate to randomize.
• The workflow is gone after a new pod. It was saved on the old pod. Export and download next time.