The quick answer
The first ComfyUI workflow should be boring: one model, one prompt, one negative prompt, one resolution, one seed, one output folder, and one notes file. If that baseline is not repeatable, every advanced graph you add will be harder to debug.
Choose the GPU around the workflow
Run the baseline first. Upgrade only when resolution, batching, LoRAs, or upscale chains prove you need more VRAM.
Baseline workflow
| Stage | What to fix | Why it matters | Do not change yet |
|---|---|---|---|
| Model | One checkpoint or model file | Changing models changes speed, memory use, and output style. | Do not test multiple models at once. |
| Prompt | One prompt and one negative prompt | Keeps visual changes tied to settings instead of wording drift. | Do not keep rewriting while benchmarking. |
| Canvas | One image size | Resolution is one of the easiest ways to trigger VRAM pressure. | Do not add upscaling yet. |
| Sampling | Fixed steps, sampler, scheduler, and seed | Turns the run into a comparable baseline. | Do not randomize the seed for the first test. |
Folder layout
Keep workflow files and outputs near the notes that explain them. If you ever publish a benchmark, readers need the exact context.
| Folder | Contents | Reason |
|---|---|---|
| /workflows | Saved ComfyUI workflow JSON files | Lets you rerun the exact graph later. |
| /outputs/baseline | Images from the fixed baseline run | Separates baseline outputs from experiments. |
| /outputs/failed | Broken or partial outputs | Failure cases are useful evidence. |
| /notes | Hardware, settings, render notes, and follow-up tests | Prevents vague memory-based claims. |
Settings to capture
ComfyUI evidence checklist
- Hardware: GPU, VRAM, system RAM, OS, driver or platform details.
- Software: ComfyUI version, model or checkpoint, custom nodes, workflow file.
- Run settings: prompt, negative prompt, resolution, steps, sampler, scheduler, seed, batch size.
- Result: render time, VRAM pressure if available, output sample, failure notes, and whether the workflow is pleasant to repeat.
Upgrade path
Once the baseline works, add complexity one layer at a time. The goal is to learn what breaks first, not to build an impressive graph you cannot explain.
Add next
- One LoRA or style layer.
- One upscale pass.
- Batch size test.
- One alternate resolution.
Delay until later
- Video workflows.
- Many custom nodes at once.
- Multi-stage routing.
- Buying another GPU before the failure is clear.
Where the gear advice fits
ComfyUI buying advice should start from workflow failure. If the baseline works but upscaling fails, you may need more VRAM. If outputs are scattered, you need storage discipline. If long runs crash, check cooling and power before shopping for a new card.
GPU choice check
Do not let one card become the default. Start with 16GB if you are learning, consider used 24GB when VRAM-per-dollar matters, and look at RTX 4090/5090-class builds when speed, warranty, or 32GB VRAM are worth paying for.
ComfyUI workflow FAQ
What is the best first ComfyUI workflow to test?
The best first workflow is a simple text-to-image run with fixed model, prompt, negative prompt, size, steps, sampler, seed, and output folder. Keep it repeatable before adding upscale chains, ControlNet-style inputs, LoRAs, or batching.
What should I record for a ComfyUI benchmark?
Record the GPU, VRAM, driver or platform, ComfyUI version, model or checkpoint, workflow file, prompt, seed, resolution, steps, sampler, batch size, render time, failure notes, and output samples.
Do I need 24GB VRAM for ComfyUI?
You do not need 24GB VRAM for every ComfyUI workflow, but 24GB gives more room for higher resolutions, upscale chains, LoRAs, ControlNet-style inputs, and experimentation. Lower VRAM can work if the workflow stays smaller.
When should I add advanced ComfyUI nodes?
Add advanced nodes only after the baseline workflow is stable and documented. If a simple run is not repeatable, adding upscale chains, routing, batching, or custom nodes makes debugging harder.
Final advice
Do not let ComfyUI become a folder of mystery graphs. Start with one repeatable baseline, record the evidence, then add one new variable at a time. That turns a workflow into a useful home-lab test instead of a screenshot you cannot reproduce.