What is the difference between the checkpoints and diffusion_models folders?

The ComfyUI/models/checkpoints/ folder holds complete bundled model files used by the Load Checkpoint node. The ComfyUI/models/diffusion_models/ folder holds UNet-only files used by the Load Diffusion Model node, such as the FLUX.1 UNet. Placing a file in the wrong folder means it will not appear in the dropdown of the node you expect.

What is UNETLoader vs CheckpointLoaderSimple?

CheckpointLoaderSimple is the internal name for the Load Checkpoint node — it loads MODEL, CLIP, and VAE from one file. UNETLoader is the internal name for the Load Diffusion Model node — it loads only the MODEL (UNet) and requires separate CLIPLoader and VAELoader nodes to provide CLIP and VAE.

Load Checkpoint vs Load Diffusion Model in ComfyUI (2026)

Load Checkpoint vs Load Diffusion Model

Quick answer: Load Checkpoint and Load Diffusion Model are two different ways to load an AI model in ComfyUI. Load Checkpoint loads a complete bundled file (MODEL + CLIP + VAE) and works with most SD 1.5 and SDXL models. Load Diffusion Model loads only the UNet from the diffusion_models/ folder and is used for FLUX and other split-architecture models — you must add separate CLIP and VAE loader nodes alongside it.

You opened a workflow and it uses a node you've never seen before, or your downloaded model isn't showing up in the dropdown you expected. That's the Load Checkpoint vs Load Diffusion Model confusion — two nodes that load models in completely different ways, reading from two different folders.

This guide explains exactly what each node loads, which folder each file type goes in, and which node your workflow actually needs.

What You Need

This is a conceptual guide — you don't need to download anything to follow along. If you want to test both nodes yourself, here's what's been used for the screenshots below.

Tested on: ComfyUI v0.3.7, RTX 4080 (16 GB VRAM)
For Load Checkpoint: any SDXL checkpoint, e.g. juggernautXL.safetensors from CivitAI
For Load Diffusion Model: the FLUX.1-dev UNet file from Hugging Face, plus its CLIP and VAE files

Tip: New to checkpoints entirely? Read what a checkpoint model is first, then come back here for the node-by-node comparison.

What Does Load Checkpoint Do?

Load Checkpoint (internal name: CheckpointLoaderSimple) loads a single file that contains everything needed to generate an image: the diffusion model itself (called the UNet — the part that actually generates the image), the CLIP text encoder (this reads your prompt and converts it into something the model understands), and the VAE (this converts the model's internal output into a real pixel image).

All three come from one .safetensors or .ckpt file, and the node gives you three ready-to-use outputs.

ComfyUI Load Checkpoint node showing MODEL, CLIP, and VAE outputs from a single bundled checkpoint file — The Load Checkpoint node. One dropdown, one file, three outputs ready to connect — no extra nodes needed.

Right-click the empty canvas and select Add Node → loaders → Load Checkpoint. The node appears with one dropdown (ckpt_name) and three outputs on the right: MODEL, CLIP, and VAE.

Click the dropdown and select your checkpoint file, for example juggernautXL.safetensors. This list only shows files that are in ComfyUI/models/checkpoints/.

Connect MODEL to your KSampler, CLIP to your text encoding nodes (CLIP Text Encode), and VAE to your VAE Decode node. All three outputs come from the same file, so there's nothing else to configure.

Warning: If a model you downloaded doesn't appear in this dropdown, it's either in the wrong folder or it's a UNet-only file meant for Load Diffusion Model instead. See the folder structure section below.

What Does Load Diffusion Model Do?

Load Diffusion Model (internal name: UNETLoader) loads only the UNet — the diffusion model component itself, with no CLIP and no VAE included. This is used for "split architecture" models, where the developer distributes the UNet, CLIP, and VAE as three separate files instead of bundling them into one checkpoint. FLUX.1 is the most common example.

Because this node only outputs MODEL, you need to add two more nodes to your workflow: a CLIP loader and a VAE loader, each pointing at their own separate file.

ComfyUI Load Diffusion Model node connected to a separate CLIPLoader and VAELoader node for FLUX — Load Diffusion Model only outputs MODEL. CLIPLoader and VAELoader are added separately and point at their own files.

Right-click the empty canvas and select Add Node → loaders → Load Diffusion Model. The node appears with one dropdown and a single output: MODEL.

Select your UNet file from the dropdown, for example flux1-dev.safetensors. This list only shows files in ComfyUI/models/diffusion_models/ — not checkpoints/.

Add a DualCLIPLoader node (Add Node → loaders → DualCLIPLoader) and select the CLIP files FLUX requires, for example clip_l.safetensors and t5xxl_fp16.safetensors. Connect its CLIP output to your text encoding nodes.

Add a VAELoader node (Add Node → loaders → VAELoader) and select the matching VAE, for example ae.safetensors. Connect its VAE output to your VAE Decode node.

This is why a FLUX workflow looks more crowded than a standard SDXL one — three loader nodes instead of one, each reading from a different folder.

Load Checkpoint vs Load Diffusion Model — Side-by-Side

Here's the full comparison in one place. The core difference is simple: Load Checkpoint loads three things from one file, Load Diffusion Model loads one thing from one file and needs two more nodes to fill in the rest.

	Load Checkpoint	Load Diffusion Model
Internal node name	CheckpointLoaderSimple	UNETLoader
What it loads	MODEL + CLIP + VAE — all from one file	MODEL (UNet) only
File format	Single .safetensors or .ckpt file	UNet file, separate CLIP file(s), separate VAE file
Extra nodes needed	None — all 3 outputs ready to connect	CLIPLoader/DualCLIPLoader + VAELoader
Folder	models/checkpoints/	models/diffusion_models/
Typical models	SD 1.5, SDXL, Juggernaut, Pony Diffusion	FLUX.1-dev, FLUX.1-schnell (split releases)
Beginner-friendly	✓ Yes — one node, one dropdown	Needs 3 nodes wired correctly

Which Folder Does Each Model Go In?

Each node reads from a different folder inside your ComfyUI installation. A file placed in the wrong folder simply won't appear in that node's dropdown — it isn't broken, ComfyUI just isn't looking there.

ComfyUI/
└── models/
├── checkpoints/       ← Load Checkpoint reads from here (SD1.5, SDXL files)
├── diffusion_models/  ← Load Diffusion Model reads from here (FLUX UNet files)
├── clip/              ← CLIPLoader / DualCLIPLoader reads from here
├── vae/               ← VAELoader reads from here
├── loras/
└── controlnet/

ComfyUI models folder structure showing checkpoints, diffusion_models, clip, and vae subfolders — The ComfyUI/models/ folder. Bundled checkpoints go in checkpoints/, FLUX UNet files go in diffusion_models/.

For a standard checkpoint, drop the file here:

ComfyUI/models/checkpoints/

For a FLUX UNet file, drop it here instead:

ComfyUI/models/diffusion_models/

Tip: After moving a file, press F5 to refresh your browser tab. ComfyUI re-scans the model folders on page load — no full restart needed.

When Should You Use Each Node?

Use Load Checkpoint for SD 1.5, SDXL, Juggernaut XL, Pony Diffusion, and the vast majority of models on CivitAI. These are distributed as single bundled files and this is the only node you need.

Use Load Diffusion Model for FLUX.1-dev, FLUX.1-schnell, and other newer releases that distribute their UNet, CLIP, and VAE as separate files. If you're installing FLUX in ComfyUI for the first time, expect to be wiring up all three loader nodes.

Warning: Not every FLUX release is split. Some community re-packagers bundle FLUX into a single checkpoint file for use with Load Checkpoint. Always check the model's download page — it will tell you which node and folder to use.

Troubleshooting

"Model not showing up in Load Checkpoint dropdown"

The file is either in the wrong folder or it's a UNet-only file meant for Load Diffusion Model.

Confirm the file is in ComfyUI/models/checkpoints/, not diffusion_models/.

Press F5 to refresh the browser tab and reopen the dropdown.

If it's a FLUX UNet file (often named with "unet" or "fp8" in the filename), move it to diffusion_models/ and use Load Diffusion Model instead.

ComfyUI red node error when a required model file cannot be found in the expected folder — A red node usually means ComfyUI can't find the selected file, or the node type itself isn't installed.

"Required input is missing: clip" or similar CLIP error

This happens when Load Diffusion Model is used without a connected CLIP loader.

Add a DualCLIPLoader node via Add Node → loaders → DualCLIPLoader.

Select the correct CLIP files for your model (FLUX needs both clip_l.safetensors and a T5 file).

Connect its CLIP output to every CLIP Text Encode node in the workflow.

Black or blank image output with no errors

Generation finishes but the image is blank — almost always a missing or wrong VAE.

Add a VAELoader node via Add Node → loaders → VAELoader.

Select the VAE that matches your model — FLUX uses ae.safetensors, not the SDXL VAE.

Connect its VAE output to your VAE Decode node, replacing any connection from a Load Checkpoint node's VAE output.

ComfyUI node with an unconnected CLIP or VAE input highlighted in red — An unconnected CLIP or VAE input shows as a red dot or red outline on the node — connect the matching loader's output to it.

Tip: If a node itself appears red and says "This node type does not exist", that's a missing custom node, not a model problem. See how to install missing nodes in ComfyUI for the fix.

Frequently Asked Questions

Load Checkpoint loads a single bundled file containing the MODEL, CLIP, and VAE all at once. Load Diffusion Model loads only the UNet (the diffusion model itself), and you need to add separate CLIPLoader and VAELoader nodes to complete the setup. Load Checkpoint is used for SD 1.5, SDXL, and most CivitAI models. Load Diffusion Model is used for FLUX.1 and other split-architecture models.

The checkpoints/ folder holds complete bundled model files used by Load Checkpoint. The diffusion_models/ folder holds UNet-only files used by Load Diffusion Model, such as the FLUX.1 UNet. Placing a file in the wrong folder means it won't appear in the dropdown of the node you expect.

Only if the FLUX release you downloaded is packaged as a single bundled checkpoint file. Most FLUX.1 releases on Hugging Face distribute the UNet, CLIP, and VAE as separate files, which require Load Diffusion Model plus separate CLIPLoader and VAELoader nodes. Check the model's download page to confirm which format you have.

The file is likely in the wrong folder. Load Diffusion Model reads from ComfyUI/models/diffusion_models/, not models/checkpoints/. Move the file to the correct folder and refresh your browser tab — no restart needed.

No. Load Checkpoint outputs MODEL, CLIP, and VAE directly from the single bundled file, so you connect these outputs straight to your sampler and conditioning nodes. Separate CLIPLoader and VAELoader nodes are only needed alongside Load Diffusion Model.

CheckpointLoaderSimple is the internal name for Load Checkpoint — it loads MODEL, CLIP, and VAE from one file. UNETLoader is the internal name for Load Diffusion Model — it loads only MODEL and requires separate CLIPLoader and VAELoader nodes to provide CLIP and VAE.

What to Do Next

If your workflow uses Load Checkpoint, pick any SDXL model from CivitAI, drop it in ComfyUI/models/checkpoints/, and run it. If your workflow uses Load Diffusion Model, follow how to install FLUX in ComfyUI for the exact file list and folder placement for all three loader nodes.