Which loader for which file

<!-- lesson: loaders -->

Lesson "loaders", in the group Pictures: models of the lesson list. Steps: 8.

#What you will be able to do

You find the right loader for four kinds of model files and learn where ComfyUI looks for each.

#What you need

• It goes on from A first picture from text. The card says so; it is a hint, not a lock.
• Model files the lesson checks:
  • a picture model (SD 1.5, SDXL, Pony, Illustrious, Flux and the like), not a video model: models/checkpoints, for example juggernautXL_v9.safetensors;
  • a full VAE, not a small preview decoder (taesd, vae_approx): models/vae, for example sdxl_vae.safetensors.
• Nodes: Load Checkpoint, Load Diffusion Model, Load CLIP, Load VAE.
• The app downloads and installs nothing. The step texts below say which files and folders, and which ComfyUI template names them.

#The steps

The words of each step, as the lesson shows them. Node names are in bold.

1. Every loader has its folder (reading)

   A model file comes in one of a few kinds, and each kind has its loader node and its folder in ComfyUI. A loader lists only the files of its own folder. So when a file is not in the list, it lies in another folder or is another kind of file.

   In this lesson you put four loaders on the canvas and choose a file in each. They do not need to be linked: the aim is to know which node takes which file.

2. A checkpoint: Load Checkpoint

   Add a Load Checkpoint node and choose a file in ckpt_name.

   A checkpoint holds three parts in one file, so the node has three outputs: the model, the text encoder and the VAE. ComfyUI looks for it in models/checkpoints. SD 1.5, SDXL, Pony, Illustrious and SD 3.5 usually come this way.

3. A diffusion model alone: Load Diffusion Model

   Add a Load Diffusion Model node and choose a file in unet_name.

   This is only the part that draws, without the text encoder and the VAE. ComfyUI looks for it in models/diffusion_models (and in the old folder models/unet). Flux, Flux.2, Z-Image, Qwen-Image, HiDream and the video models come this way. The weight_dtype field can squeeze a full file into fp8 to save memory.

4. A text encoder: Load CLIP

   Add a Load CLIP node, choose a file in clip_name and the family in type.

   The text encoder turns the prompt into numbers for the model. ComfyUI looks for it in models/text_encoders (and in the old models/clip). The type must match the family: lumina2 for Z-Image, qwen_image for Qwen-Image, flux2 for Flux.2, sd3 for a single T5 of SD 3.5. Families with two encoders take the node for two files (Flux.1: flux), SD 3.5 the one for three, HiDream the one for four.

5. A VAE: Load VAE

   Add a Load VAE node and choose a file in vae_name.

   The VAE turns the latent into pixels and back. ComfyUI looks for it in models/vae. Each family has its own: ae.safetensors for Flux.1 and Z-Image, flux2-vae.safetensors for Flux.2, qwen_image_vae.safetensors for Qwen-Image.

   The list may also show small preview files (taesd, taesdxl, taef1 and others) from models/vae_approx. They are quick and rough: ComfyUI uses them to show the picture while it is being drawn. For the final picture take the full VAE.

6. How to tell what a file is (reading)

   Ways to tell, from the surest:

   • the official template of the model names each file and its folder;
   • the model's page says "checkpoint", "diffusion model", "text encoder" or "VAE";
   • the size: a VAE is up to a few hundred MB, an SD 1.5 checkpoint 2 to 4 GB, an SDXL one about 6.5 GB, a large diffusion model alone 10 to 40 GB;
   • the card of the file in the "Models" tab of this app names its family.

   A checkpoint in Load Diffusion Model or a diffusion model in Load Checkpoint ends in an error. A .gguf file needs the loaders of the GGUF pack. The old formats .ckpt, .pt, .pth and .bin can run code when loaded: prefer .safetensors.

7. Other files of picture graphs (reading)

   Other kinds you will meet, with their folder:

   • LoRA: models/loras, the Load LoRA node;
   • ControlNet: models/controlnet;
   • upscale models: models/upscale_models;
   • CLIP Vision, for a picture as a hint: models/clip_vision;
   • embeddings: models/embeddings; they have no node and are called in the prompt as embedding:name.

8. What next (reading)

   Now you know which node takes which file. The lessons "Flux.1: a picture from separate files" and "Z-Image Turbo: a picture in eight steps" build whole graphs from such loaders.

   These four loaders can stay on the canvas or be deleted: a loader whose output leads nowhere is not run by ComfyUI.

#What the lesson checks

A reading step is done when you press Got it, next. A doing step ticks itself off when the canvas holds:

• Step 2, A checkpoint: Load Checkpoint:
  • a Load Checkpoint node.
  • ckpt_name of Load Checkpoint holds a picture model (SD 1.5, SDXL, Pony, Illustrious, Flux and the like), not a video model; such a file lies in models/checkpoints, for example juggernautXL_v9.safetensors.
• Step 3, A diffusion model alone: Load Diffusion Model:
  • a Load Diffusion Model node.
  • unet_name of Load Diffusion Model: a value is chosen.
• Step 4, A text encoder: Load CLIP:
  • a Load CLIP node.
  • clip_name of Load CLIP: a value is chosen.
  • type of Load CLIP: a value is chosen.
• Step 5, A VAE: Load VAE:
  • a Load VAE node.
  • vae_name of Load VAE holds a full VAE, not a small preview decoder (taesd, vae_approx); such a file lies in models/vae, for example sdxl_vae.safetensors.

#What next

These lessons go on from this one: Flux.1: a picture from separate files, Z-Image Turbo: a picture in eight steps.

#See also

• Separate loaders: diffusion model, text encoder, VAE
• Checkpoint
• Text encoders: CLIP, T5 and others
• VAE
• Where ComfyUI files live
• How lessons work