Checkpoints and Models: What They Are and Where to Get Them

What a Checkpoint Actually Contains

The term "checkpoint" comes from machine learning training — it is a saved snapshot of model weights at a point during training. In ComfyUI, a checkpoint file is a single large file (typically 2–24 GB) that contains three bundled components:

• The UNet: The core diffusion model. This is the network that actually denoises the latent image step by step. It is the largest component and the one that determines the model's visual style and capability.
• The CLIP text encoder: Converts your text prompt into a numerical vector the UNet can use as guidance. Different architectures use different CLIP variants.
• The VAE: The variational autoencoder, which encodes pixel images into compressed latent space for the UNet to work with, and decodes them back to pixels after generation. Sometimes the checkpoint's built-in VAE produces slightly washed-out colors, and you load a separate VAE file to replace it — covered in Lesson 7.

When you select a model in the Load Checkpoint node, all three components are loaded from that single file. The three colored outputs (MODEL, CLIP, VAE) wire these components to the rest of the workflow.

The Three Major Architectures

Most checkpoint models fall into one of three families. Each family has a different internal structure, which means the node wiring for the workflow looks different — and the optimal settings differ too.

SD 1.5 (Stable Diffusion 1.5): The original architecture, released in 2022. Uses a single CLIP text encoder. Generates images natively at 512x512. Extremely well-supported with thousands of LoRAs, ControlNet models, and fine-tuned checkpoints on CivitAI. Slower than SDXL and FLUX but very lightweight — runs on GPUs with as little as 4 GB VRAM. Good starting model for learning.

SDXL (Stable Diffusion XL): Released in 2023. Uses two CLIP text encoders (clip_g and clip_l from OpenCLIP). Generates at 1024x1024 natively. Significantly better quality than SD 1.5, especially for text legibility and complex compositions. Requires 8 GB+ VRAM. Includes a refiner model for optional quality improvement in a second pass. Most active fine-tune community after SD 1.5.

FLUX.1: Released in 2024 by Black Forest Labs, the original Stable Diffusion team. Uses dual CLIP (T5-XXL + CLIP-L). Generates at any aspect ratio without a fixed native resolution. Dramatically better photorealism, text rendering, and prompt adherence than SD 1.5 or SDXL. Requires 12 GB+ VRAM for quantized versions; 24 GB+ for full precision. The current state-of-the-art for most image generation tasks.

This track uses SD 1.5 examples for most lessons because it runs on the widest range of hardware. The image generation track (T49) covers FLUX.1 workflows in depth.

Where to Download Models

Two main sources:

• CivitAI (civitai.com): The primary community hub for fine-tuned checkpoints, LoRAs, and other model types. Filters by base model (SD 1.5, SDXL, FLUX) and model type. Create a free account to download. Be aware that some models are NSFW — use the Safe Mode filter.
• Hugging Face (huggingface.co): The authoritative source for base models (v1-5-pruned-emaonly.safetensors, stabilityai/stable-diffusion-xl-base-1.0, black-forest-labs/FLUX.1-dev). Download individual files, not entire repositories, unless you need everything.

For your first model if you don't have one: search Hugging Face for runwayml/stable-diffusion-v1-5 and download v1-5-pruned-emaonly.safetensors. It is 4 GB, runs on most GPUs with 4 GB+ VRAM, and works with the default workflow immediately.

The ComfyUI Folder Structure for Models

Model files must go in the correct subfolder or ComfyUI will not detect them. The base directory is your ComfyUI installation folder (or ~/comfyui/ for RunPod). Inside it:

• models/checkpoints/ — full checkpoint models (.safetensors, .ckpt)
• models/vae/ — standalone VAE files
• models/loras/ — LoRA files
• models/controlnet/ — ControlNet models
• models/clip/ — standalone CLIP models (used in FLUX.1 workflows with separate loaders)
• models/unet/ or models/diffusion_models/ — standalone UNet files (used in some FLUX.1 setups)
• models/upscale_models/ — upscaler models like 4x-UltraSharp
• models/embeddings/ — textual inversion files

After placing a file in the correct folder, press R on the canvas (or use the Model panel in the left sidebar → refresh) to update the dropdown lists. The new model will appear.

SafeTensors vs. CKPT Files

You will see two file formats for checkpoint models: .safetensors and .ckpt (or .pt). Always prefer .safetensors when both are available. SafeTensors is a safer format that cannot execute arbitrary code when loaded — unlike .ckpt files, which are Python pickle files and can theoretically contain malicious code. Reputable sources publish both; newer models are usually SafeTensors only.