Models

Overview

ComfyUI supports various model types for image generation, each serving a specific purpose in the pipeline. The model system handles loading, caching, memory management, and multi-model workflows.

Model Types

Diffusion Models

The core model for denoising latents:

class CheckpointLoaderSimple:
    @classmethod
    def INPUT_TYPES(s):
        return {
            "required": {
                "ckpt_name": (folder_paths.get_filename_list("checkpoints"), {
                    "tooltip": "The name of the checkpoint (model) to load."
                })
            }
        }
    
    RETURN_TYPES = ("MODEL", "CLIP", "VAE")
    FUNCTION = "load_checkpoint"
    CATEGORY = "loaders"
    
    def load_checkpoint(self, ckpt_name):
        ckpt_path = folder_paths.get_full_path_or_raise("checkpoints", ckpt_name)
        out = comfy.sd.load_checkpoint_guess_config(
            ckpt_path, 
            output_vae=True, 
            output_clip=True, 
            embedding_directory=folder_paths.get_folder_paths("embeddings")
        )
        return out[:3]

Auto-detection: load_checkpoint_guess_config automatically detects the model architecture (SD 1.5, SDXL, SD3, Flux, etc.) from the checkpoint file.

VAE (Variational Autoencoder)

Encodes images to latent space and decodes latents back to images:

class VAELoader:
    @classmethod
    def INPUT_TYPES(s):
        return {"required": {"vae_name": (s.vae_list(s), )}}
    
    RETURN_TYPES = ("VAE",)
    FUNCTION = "load_vae"
    CATEGORY = "loaders"
    
    def load_vae(self, vae_name):
        if vae_name == "pixel_space":
            # Special case: no encoding/decoding
            sd = {"pixel_space_vae": torch.tensor(1.0)}
        elif vae_name in self.image_taes:
            # Tiny AutoEncoder for fast preview
            sd = self.load_taesd(vae_name)
        else:
            vae_path = folder_paths.get_full_path_or_raise("vae", vae_name)
            sd, metadata = comfy.utils.load_torch_file(
                vae_path, return_metadata=True
            )
        
        vae = comfy.sd.VAE(sd=sd, metadata=metadata)
        vae.throw_exception_if_invalid()
        return (vae,)

TAESD (Tiny AutoEncoder)

Lightweight VAE alternatives for fast preview:

@staticmethod
def load_taesd(name):
    sd = {}
    encoder = f"{name}_encoder.safetensors"
    decoder = f"{name}_decoder.safetensors"
    
    enc = comfy.utils.load_torch_file(
        folder_paths.get_full_path_or_raise("vae_approx", encoder)
    )
    for k in enc:
        sd[f"taesd_encoder.{k}"] = enc[k]
    
    dec = comfy.utils.load_torch_file(
        folder_paths.get_full_path_or_raise("vae_approx", decoder)
    )
    for k in dec:
        sd[f"taesd_decoder.{k}"] = dec[k]
    
    # Set scale and shift based on model type
    if name == "taesd":
        sd["vae_scale"] = torch.tensor(0.18215)
        sd["vae_shift"] = torch.tensor(0.0)
    elif name == "taesdxl":
        sd["vae_scale"] = torch.tensor(0.13025)
        sd["vae_shift"] = torch.tensor(0.0)
    
    return sd

CLIP (Text Encoder)

Encodes text prompts into conditioning:

class CLIPLoader:
    @classmethod
    def INPUT_TYPES(s):
        return {
            "required": {
                "clip_name": (folder_paths.get_filename_list("text_encoders"),),
                "type": (["stable_diffusion", "stable_cascade", "sd3", 
                         "stable_audio", "flux"], )
            }
        }
    
    RETURN_TYPES = ("CLIP",)
    FUNCTION = "load_clip"
    CATEGORY = "advanced/loaders"
    
    def load_clip(self, clip_name, type="stable_diffusion"):
        clip_type = getattr(comfy.sd.CLIPType, 
                          type.upper(), 
                          comfy.sd.CLIPType.STABLE_DIFFUSION)
        
        clip_path = folder_paths.get_full_path_or_raise(
            "text_encoders", clip_name
        )
        clip = comfy.sd.load_clip(
            ckpt_paths=[clip_path], 
            embedding_directory=folder_paths.get_folder_paths("embeddings"),
            clip_type=clip_type
        )
        return (clip,)

Dual CLIP Models

Some models require two CLIP encoders:

class DualCLIPLoader:
    @classmethod
    def INPUT_TYPES(s):
        return {
            "required": {
                "clip_name1": (folder_paths.get_filename_list("text_encoders"),),
                "clip_name2": (folder_paths.get_filename_list("text_encoders"),),
                "type": (["sdxl", "sd3", "flux"], )
            }
        }
    
    RETURN_TYPES = ("CLIP",)
    FUNCTION = "load_clip"
    
    def load_clip(self, clip_name1, clip_name2, type):
        clip_path1 = folder_paths.get_full_path_or_raise(
            "text_encoders", clip_name1
        )
        clip_path2 = folder_paths.get_full_path_or_raise(
            "text_encoders", clip_name2
        )
        
        clip = comfy.sd.load_clip(
            ckpt_paths=[clip_path1, clip_path2],
            clip_type=getattr(comfy.sd.CLIPType, type.upper())
        )
        return (clip,)

SDXL: Requires CLIP-L and CLIP-G
SD3: Supports CLIP-L + CLIP-G or CLIP-L + T5 or CLIP-G + T5
Flux: Requires CLIP-L and T5

Model Paths

From folder_paths.py:22-53, ComfyUI organizes models in standard directories:

models_dir = os.path.join(base_path, "models")

folder_names_and_paths = {
    "checkpoints": ([os.path.join(models_dir, "checkpoints")], 
                    supported_pt_extensions),
    "vae": ([os.path.join(models_dir, "vae")], 
            supported_pt_extensions),
    "loras": ([os.path.join(models_dir, "loras")], 
              supported_pt_extensions),
    "text_encoders": ([os.path.join(models_dir, "text_encoders"),
                       os.path.join(models_dir, "clip")], 
                      supported_pt_extensions),
    "diffusion_models": ([os.path.join(models_dir, "unet"),
                          os.path.join(models_dir, "diffusion_models")],
                         supported_pt_extensions),
    "controlnet": ([os.path.join(models_dir, "controlnet"),
                    os.path.join(models_dir, "t2i_adapter")],
                   supported_pt_extensions),
    "embeddings": ([os.path.join(models_dir, "embeddings")],
                   supported_pt_extensions),
    "upscale_models": ([os.path.join(models_dir, "upscale_models")],
                       supported_pt_extensions),
}

Supported File Extensions

supported_pt_extensions = {
    '.ckpt', '.pt', '.pt2', '.bin', 
    '.pth', '.safetensors', '.pkl', '.sft'
}

SafeTensors Recommended: Use .safetensors format for security and reliability. Pickle-based formats (.ckpt, .pt, .pkl) can execute arbitrary code.

LoRA Models

Low-Rank Adaptation models modify base models:

class LoraLoader:
    def __init__(self):
        self.loaded_lora = None
    
    @classmethod
    def INPUT_TYPES(s):
        return {
            "required": {
                "model": ("MODEL", {
                    "tooltip": "The diffusion model to apply LoRA to"
                }),
                "clip": ("CLIP", {
                    "tooltip": "The CLIP model to apply LoRA to"
                }),
                "lora_name": (folder_paths.get_filename_list("loras"),),
                "strength_model": ("FLOAT", {
                    "default": 1.0, "min": -100.0, "max": 100.0, "step": 0.01
                }),
                "strength_clip": ("FLOAT", {
                    "default": 1.0, "min": -100.0, "max": 100.0, "step": 0.01
                })
            }
        }
    
    RETURN_TYPES = ("MODEL", "CLIP")
    FUNCTION = "load_lora"
    CATEGORY = "loaders"
    
    def load_lora(self, model, clip, lora_name, strength_model, strength_clip):
        if strength_model == 0 and strength_clip == 0:
            return (model, clip)
        
        lora_path = folder_paths.get_full_path_or_raise("loras", lora_name)
        
        # Cache loaded LoRA
        if self.loaded_lora is not None:
            if self.loaded_lora[0] == lora_path:
                lora = self.loaded_lora[1]
            else:
                self.loaded_lora = None
        
        if lora is None:
            lora = comfy.utils.load_torch_file(lora_path, safe_load=True)
            self.loaded_lora = (lora_path, lora)
        
        model_lora, clip_lora = comfy.sd.load_lora_for_models(
            model, clip, lora, strength_model, strength_clip
        )
        return (model_lora, clip_lora)

LoRA Chaining: Multiple LoRA loaders can be chained together to apply multiple adaptations. Each LoRA modifies the model passed to it.

Model File Discovery

From folder_paths.py:303-343, ComfyUI recursively scans model directories:

def recursive_search(directory: str, excluded_dir_names=None):
    if not os.path.isdir(directory):
        return [], {}
    
    result = []
    dirs = {}
    
    for dirpath, subdirs, filenames in os.walk(directory, followlinks=True):
        # Exclude .git and other special directories
        subdirs[:] = [d for d in subdirs if d not in excluded_dir_names]
        
        for file_name in filenames:
            relative_path = os.path.relpath(
                os.path.join(dirpath, file_name), 
                directory
            )
            result.append(relative_path)
        
        # Track directory modification times for cache invalidation
        for d in subdirs:
            path = os.path.join(dirpath, d)
            dirs[path] = os.path.getmtime(path)
    
    return result, dirs

def filter_files_extensions(files, extensions):
    return sorted([
        f for f in files 
        if os.path.splitext(f)[-1].lower() in extensions or len(extensions) == 0
    ])

Model Caching

filename_list_cache = {}  # Global cache

def get_filename_list(folder_name: str) -> list[str]:
    folder_name = map_legacy(folder_name)
    out = cached_filename_list_(folder_name)
    
    if out is None:
        # Cache miss - scan directory
        out = get_filename_list_(folder_name)
        filename_list_cache[folder_name] = out
    
    return list(out[0])

def cached_filename_list_(folder_name: str):
    if folder_name not in filename_list_cache:
        return None
    
    out = filename_list_cache[folder_name]
    
    # Validate cache - check if directories changed
    for folder, time_modified in out[1].items():
        if os.path.getmtime(folder) != time_modified:
            return None  # Cache invalidated
    
    return out

ControlNet Models

Guidance models for spatial control:

class ControlNetLoader:
    @classmethod
    def INPUT_TYPES(s):
        return {
            "required": {
                "control_net_name": (
                    folder_paths.get_filename_list("controlnet"),
                )
            }
        }
    
    RETURN_TYPES = ("CONTROL_NET",)
    FUNCTION = "load_controlnet"
    CATEGORY = "loaders"
    
    def load_controlnet(self, control_net_name):
        controlnet_path = folder_paths.get_full_path_or_raise(
            "controlnet", control_net_name
        )
        controlnet = comfy.controlnet.load_controlnet(controlnet_path)
        
        if controlnet is None:
            raise RuntimeError(
                "ERROR: controlnet file is invalid and does not "
                "contain a valid controlnet model."
            )
        
        return (controlnet,)

Applying ControlNet

class ControlNetApplyAdvanced:
    @classmethod
    def INPUT_TYPES(s):
        return {
            "required": {
                "positive": ("CONDITIONING",),
                "negative": ("CONDITIONING",),
                "control_net": ("CONTROL_NET",),
                "image": ("IMAGE",),
                "strength": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 10.0}),
                "start_percent": ("FLOAT", {"default": 0.0, "min": 0.0, "max": 1.0}),
                "end_percent": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 1.0})
            }
        }
    
    RETURN_TYPES = ("CONDITIONING", "CONDITIONING")
    RETURN_NAMES = ("positive", "negative")
    FUNCTION = "apply_controlnet"
    
    def apply_controlnet(self, positive, negative, control_net, image, 
                        strength, start_percent, end_percent):
        if strength == 0:
            return (positive, negative)
        
        control_hint = image.movedim(-1, 1)
        cnets = {}
        
        out = []
        for conditioning in [positive, negative]:
            c = []
            for t in conditioning:
                d = t[1].copy()
                
                prev_cnet = d.get('control', None)
                if prev_cnet in cnets:
                    c_net = cnets[prev_cnet]
                else:
                    c_net = control_net.copy().set_cond_hint(
                        control_hint, strength, 
                        (start_percent, end_percent)
                    )
                    c_net.set_previous_controlnet(prev_cnet)
                    cnets[prev_cnet] = c_net
                
                d['control'] = c_net
                d['control_apply_to_uncond'] = False
                c.append([t[0], d])
            out.append(c)
        
        return (out[0], out[1])

Model Configuration

Model Paths

Configure custom model directories in extra_model_paths.yaml:

comfyui:
  checkpoints: /path/to/checkpoints
  vae: /path/to/vae
  loras: /path/to/loras

Memory Management

Models are automatically loaded/unloaded based on VRAM:

--lowvram: Aggressive offloading
--normalvram: Moderate offloading
--highvram: Keep models in VRAM

Model Precision

Control memory usage with precision:

fp32: Full precision
fp16: Half precision
fp8_e4m3fn: 8-bit float

Safe Loading

Use safe_load=True to prevent code execution in pickle files

Advanced Model Loading

UNET-Only Loading

For separate component loading:

class UNETLoader:
    @classmethod
    def INPUT_TYPES(s):
        return {
            "required": {
                "unet_name": (folder_paths.get_filename_list("diffusion_models"),),
                "weight_dtype": (["default", "fp8_e4m3fn", 
                                 "fp8_e4m3fn_fast", "fp8_e5m2"],)
            }
        }
    
    RETURN_TYPES = ("MODEL",)
    FUNCTION = "load_unet"
    
    def load_unet(self, unet_name, weight_dtype):
        model_options = {}
        if weight_dtype == "fp8_e4m3fn":
            model_options["dtype"] = torch.float8_e4m3fn
        elif weight_dtype == "fp8_e4m3fn_fast":
            model_options["dtype"] = torch.float8_e4m3fn
            model_options["fp8_optimizations"] = True
        
        unet_path = folder_paths.get_full_path_or_raise(
            "diffusion_models", unet_name
        )
        model = comfy.sd.load_diffusion_model(
            unet_path, model_options=model_options
        )
        return (model,)

Model Validation

class VAELoader:
    def load_vae(self, vae_name):
        # ... load vae ...
        vae = comfy.sd.VAE(sd=sd, metadata=metadata)
        vae.throw_exception_if_invalid()  # Validate before returning
        return (vae,)

Best Practices

Model Organization

Use subdirectories within model folders
Name models descriptively
Keep model versions separate
Document custom model paths

Performance

Use fp16 or fp8 for large models
Enable LoRA caching for repeated use
Consider TAESD for preview workflows
Batch similar operations together

Compatibility

Check model architecture compatibility
Match VAE to checkpoint type
Verify CLIP encoder requirements
Test LoRA with base model

Security

Prefer SafeTensors format
Use safe_load for pickle files
Verify model sources
Scan for malicious code

Nodes

Using models in node workflows

Execution

Model loading during execution

Workflows

Building workflows with models

Get Started

Core Concepts

Supported Models

Advanced Features

Configuration

Overview

Model Types

Diffusion Models

VAE (Variational Autoencoder)

TAESD (Tiny AutoEncoder)

CLIP (Text Encoder)

Dual CLIP Models

Model Paths

Supported File Extensions

LoRA Models

Model File Discovery

Model Caching

ControlNet Models

Applying ControlNet

Model Configuration

Model Paths

Memory Management

Model Precision

Safe Loading

Advanced Model Loading

UNET-Only Loading

Model Validation

Best Practices

See Also

Nodes

Execution

Workflows

Build docs developers (and LLMs) love

Get Started

Core Concepts

Supported Models

Advanced Features

Configuration

​Overview

​Model Types

​Diffusion Models

​VAE (Variational Autoencoder)

​TAESD (Tiny AutoEncoder)

​CLIP (Text Encoder)

​Dual CLIP Models

​Model Paths

​Supported File Extensions

​LoRA Models

​Model File Discovery

​Model Caching

​ControlNet Models

​Applying ControlNet

​Model Configuration

Model Paths

Memory Management

Model Precision

Safe Loading

​Advanced Model Loading

​UNET-Only Loading

​Model Validation

​Best Practices

​See Also

Nodes

Execution

Workflows

Build docs developers (and LLMs) love

Overview

Model Types

Diffusion Models

VAE (Variational Autoencoder)

TAESD (Tiny AutoEncoder)

CLIP (Text Encoder)

Dual CLIP Models

Model Paths

Supported File Extensions

LoRA Models

Model File Discovery

Model Caching

ControlNet Models

Applying ControlNet

Model Configuration

Advanced Model Loading

UNET-Only Loading

Model Validation

Best Practices

See Also