# Skill: Speed up ComfyUI on limited hardware and reuse community workflows

> Lets an AI agent configure ComfyUI for weak GPUs, extract embedded workflows from PNG files, install missing nodes with ComfyUI Manager, and avoid complexity overload -- whenever a user hits memory errors, slow generation, or broken red-node workflows from the community.

Source: sodigi·learn -- comfyui/uskorenie-i-resursy · https://sodigi.io/learn/comfyui/uskorenie-i-resursy

## When to use
- User hits an out-of-memory error or generation is very slow.
- User wants to reuse a workflow shared as a PNG file.
- User opens a community workflow and half the nodes are red with errors.
- User feels overwhelmed by a complex node graph and does not know where to start.

## Core rules
- VRAM is the ceiling; if a model does not fit, enable medvram (partial offload to RAM) or lowvram (nearly all in RAM) launch flags.
- Speed levers: smaller image size and fewer steps are the fastest way to cut generation time.
- batch size = how many images are rendered simultaneously (eats VRAM); batch count = how many runs happen in sequence (does not).
- On a weak GPU: always set batch size = 1 and use batch count for multiple generations.
- ComfyUI embeds the full node graph into the PNG metadata when it saves -- drag that PNG into the ComfyUI window to restore the workflow automatically.
- Only download the original PNG (use a download button); right-click saves and messenger reposts strip the embedded metadata.
- Red nodes in a community workflow mean missing custom node packages -- use Manager > Install Missing Custom Nodes.
- Start with built-in templates before community workflows; they need no extra node packages and match your installed version.

## Procedure
1. Test with default settings first. If generation succeeds, no tuning is needed.
2. If an out-of-memory error appears, relaunch ComfyUI with --medvram; if it still fails, use --lowvram.
3. For faster iteration: reduce canvas size (512x512 or 768x768) and set steps to 20; only render the final at full quality.
4. Set batch size = 1; increase batch count to generate multiple variants sequentially without stressing VRAM.
5. To reuse a PNG workflow: download the original file, then drag-and-drop it into the ComfyUI browser window.
6. If nodes are red after loading a workflow: open Manager, click "Install Missing Custom Nodes," install all, restart.
7. Change only the prompt text in someone else's workflow; leave all other settings intact until you understand them.

## Ready-to-use prompt
```
I am having trouble with ComfyUI performance and community workflows. Help me:

1. Speed up generation on my GPU with [X] GB VRAM -- which memory-saving mode to enable and how.
2. Set batch size and batch count correctly so I can generate 10 variants without running out of memory.
3. Extract and open a workflow that is embedded inside a PNG file someone shared with me.
4. Fix red (missing) nodes in a community workflow using ComfyUI Manager.
5. Avoid getting overwhelmed by a complex graph: where to find the prompt fields and the run button.

My GPU VRAM: [X] GB
```

## Pitfalls
- Starting a test generation at maximum resolution on a weak GPU -- always do a small draft first.
- Setting batch size = 10 on a 4-6 GB GPU hoping for 10 simultaneous images -- VRAM will run out; use batch count instead.
- Right-click saving an image from a browser or messenger -- this typically strips the PNG metadata and the embedded workflow is lost.
- Panicking at red nodes in a community workflow -- this is normal; Manager resolves it in a few clicks.
- Opening a complex 100-node community workflow as a first project -- start with a simple built-in template.
- Changing multiple settings at once when exploring an unfamiliar workflow -- always change one thing at a time.