Load Checkpoint

If you've run a single ComfyUI tutorial, this is probably the first loader you ever used. Load Checkpoint takes one file from ComfyUI/models/checkpoints and hands you three wires at once: MODEL (the denoiser), CLIP (the text encoder), and VAE (the pixel codec). It's the complete starter kit in a single node, and for SD 1.5 and SDXL checkpoints it's still exactly right. The catch, and it's a real one: most of the models people run in 2026 don't ship as a single checkpoint anymore, so this node's prominence has quietly shrunk to the classic era.

How it works

One input, ckpt_name - the file. Behind the scenes, ComfyUI reads the checkpoint, guesses its architecture from the weights, and splits out the three components. That auto-detection is the node's superpower and its limit: it works when everything is in one file, and a checkpoint is a complete model that generates on its own - the thing everyone means by "model," the level of the hierarchy above LoRAs and embeddings.

A few things worth knowing about the file before you download it:

• Prefer .safetensors over .ckpt. The .ckpt format uses Python pickle, which can execute arbitrary code on load. The community has seen malicious fake models. Same output quality, far less risk.
• Grab the pruned fp16 build. Full checkpoints carry optimizer states for resuming training and weigh twice as much. Output is identical.
• Sizes are architecture-bound: ~2GB for SD 1.5, ~6.5GB for SDXL, ~12GB for Flux in fp16.

The modern fork in the road

Here's the honest version of why you'll sometimes load a checkpoint and sometimes not. Flux, SD3, and the newer video models ship as separate files - transformer, text encoder, VAE all in their own downloads - and their workflows use Load Diffusion Model + Load CLIP / DualCLIP + Load VAE instead. So the rule is simple:

• SD 1.5 / SDXL workflow with a single .safetensors: this node, every time.
• Flux / SD3 / modular model: don't force it into this node. Use the component loaders.

The output wiring is the same either way: MODEL → sampler, CLIP → CLIP Text Encode → conditioning, VAE → VAE Encode/Decode. If a checkpoint-based workflow produces grey or washed-out images on SD 1.5, the VAE inside the file is often the weak link - load the 840k separately via Load VAE and wire its VAE output over the checkpoint's. Modern checkpoints bake the right VAE in, so leave it alone there.

Common issues

• Wrong architecture add-ons. The checkpoint defines the family; a Flux LoRA on an SDXL checkpoint does nothing. Check the base before stacking.
• Merges inherit their parents' problems. Most community checkpoints are merges, and merge lineages share defects - same VAE quirks, same color biases. It's why some checkpoints "just look better" than others for reasons nobody can quite pin down.
• Don't swap checkpoints to add one character or style. A 10–200MB LoRA on the checkpoint you already like beats a multi-gigabyte download.

It ships with ComfyUI core - no install. For a clean SD/SDXL start it remains the single best node in the category: one file, three wires, done.