Image and video generation
ComfyUI first image: install, load a model and fix common errors
ComfyUI first presents a web of connections rather than a picture. Begin with one model, one short prompt and one image. This setup is a small installation check, not a claim about the latest model's visual quality.
Key takeaways
- 01
- The workflow uses core nodes: SD 1.5, 512×512, 20 steps and one image.
- 02
- The model and workflow JSON are different downloads.
- 03
- The illustrations explain concepts; they are not measured outputs from this workflow.
1. Install Desktop for your operating system
Choose Comfy Desktop for Windows, Apple silicon Mac or Linux from the official download. Create a ComfyUI installation and let setup finish. An existing portable or manual installation does not need replacing for this guide.
Do not install several custom-node packs at the start. They make it harder to separate installation errors from extension errors. This workflow uses only built-in nodes and no cloud API nodes.
2. Add the model and open the workflow
Use v1-5-pruned-emaonly-fp16.safetensors. Check its distribution terms and place it in the models/checkpoints directory used by your ComfyUI installation, or its configured shared model directory. The JSON alone cannot generate an image.
Save the JSON below, then drag it onto ComfyUI or use Workflows → Open. Select the downloaded file in Load Checkpoint. If absent, check the directory and refresh or restart. Renaming a file extension does not fix an incompatible model.
3. Run without changing size or batch
The setup is 512×512, batch 1, seed 42, 20 steps, CFG 7, euler/normal and denoise 1. The English prompt describes a ceramic cup on a wooden table in window light. Keep these values for the first Run; they are a simple SD 1.5 smoke test.
The first run may include model loading. Confirm that Save Image completes and an output file exists. A preview alone is not the finish line; open the saved file to retain an original for later comparisons.
4. Change one thing on the second run
Keep the seed and change only one lighting or background detail. Changing model, resolution and steps at once hides the cause of a difference. A fixed seed does not guarantee pixel-identical output across different models or execution environments.
An unattractive first image is not necessarily an installation failure. Model choice, prompt specificity and post-processing are separate decisions. Once the basic path works, move to a model suited to your style or editing task.
5. If it stops, check in this order
A missing checkpoint calls for checking the path and filename. Missing-node warnings may mean you opened a different workflow: this file requires no custom nodes. For CUDA out of memory, close competing GPU jobs and return to batch 1 at 512×512.
If output saves but is unusually slow, check the execution log for the selected device. This ComfyUI path uses PyTorch MPS on Apple silicon, not the same path as an MLX LLM. Record a successful baseline before increasing resolution.
6. Keep conditions with the time
The download includes a measurement template for GPU, RAM, app version, model file, resolution and seed. Keep cold loading separate. Repeated Run clicks can reuse cached output; use seeds 42, 43 and 44 and verify sampling really runs.
This is a structurally checked example, not a new hardware benchmark. Timing and memory fields are intentionally empty. Keep output images and logs so others can verify the workload. Hardware comparisons become useful when your real task exceeds this small example.
