Building a Local ComfyUI Environment and Model Foundation

Installing ComfyUI and installing an image-generation model are separate tasks. The application supplies the interface and execution framework, but a checkpoint supplies the model used to generate images. A fresh installation can open successfully and still fail when a prompt is queued because no checkpoint is available.

A checkpoint is therefore the foundation of every generation workflow. Different checkpoints are suited to different visual outcomes, including realistic images, anime, watercolor, and stylized character designs. The model selected in the checkpoint loader influences the style and character of the resulting image before any prompt or sampling parameter is considered.

On Windows, the portable installation is distributed as a compressed archive. After downloading and extracting it, the user starts ComfyUI with the batch file that matches the available hardware. A system with a compatible NVIDIA GPU uses the NVIDIA batch file. CPU-only execution and Apple M1 or M2 systems are also supported, but CPU execution is much slower. An NVIDIA GPU is the more practical choice for serious local generation because it provides faster execution and broader compatibility with other open-source AI tools.

The portable installation archive can be approximately 1.4 GB, while a selected checkpoint can be approximately 6.4 GB. These sizes illustrate an operational constraint: local image generation requires both storage capacity and download time. Additional checkpoints, upscalers, ControlNet models, and custom-tool assets increase that requirement.

Place SafeTensors checkpoint files in:

After the file is placed in that directory, ComfyUI must be started or restarted. The checkpoint then appears in the Load Checkpoint node. If only one model is present, it may be selected automatically.

Model selection is not merely a technical prerequisite. Thousands of checkpoints may be available, and their sample outputs can be used to judge whether they fit the required style. A model-ranking list based on accumulated blind comparisons can also help filter a large catalog before downloading multi-gigabyte files.

An SDXL checkpoint provides a higher-quality option than Stable Diffusion 1.5. The workflows below do not require separate SDXL base and refiner files. A single selected community checkpoint is sufficient, avoiding several additional gigabytes of downloads. The refiner is unnecessary in most cases for this setup.

Many apparent ComfyUI failures are actually environment or model-placement problems. If the interface opens but Queue Prompt produces a checkpoint error, rebuilding the graph will not solve the issue. The model must first exist in the correct directory and be selectable by the checkpoint loader.

Model choice also affects all later decisions. Resolution, step count, and even the usefulness of certain extensions depend on the model family. A Lightning checkpoint, for example, is designed to produce an image in far fewer steps than a conventional setup. Selecting the model first prevents the rest of the workflow from being tuned around the wrong assumptions.

Use the following sequence to establish a working local environment:

1. Download the ComfyUI portable archive for the operating system.
2. Extract the archive to a convenient location. The extracted directory can be placed on the desktop or another local folder.
3. Start the application with the batch file that matches the hardware.
4. Confirm that the interface opens. A checkpoint-related error at this stage indicates that the interface is running but no model is available.
5. Select a checkpoint by examining its intended visual style and example outputs.
6. Download the SafeTensors checkpoint file.
7. Place it in ComfyUI/models/checkpoints.
8. Restart ComfyUI and select the checkpoint in the Load Checkpoint node.
9. Enter a minimal positive prompt and negative prompt, then queue a generation to verify that the entire installation works.

The model-hosting catalog can contain substantial not-safe-for-work content. Filters should therefore be enabled when browsing it in a workplace or shared environment.

CPU-only execution makes ComfyUI accessible without an NVIDIA GPU, but it is significantly slower. The ability to run is not the same as a practical generation speed.

A large model catalog creates choice but also introduces storage costs and selection overhead. Downloading every model is neither necessary nor efficient. Sample outputs and comparative rankings help narrow the field, but the selected checkpoint still needs to be tested against the intended prompts.

Using a single checkpoint instead of separate base and refiner files is a workflow-specific simplification, not a reason to assume every model setup uses the same file structure.

A functioning ComfyUI setup requires both the application and a correctly placed checkpoint. Treat hardware, storage, model family, and checkpoint style as foundational decisions before tuning prompts or building advanced graphs.