Supported models

ComfyUI ships with built-in support for ~30 model families across image, video, audio, and 3D generation. This page is a navigation aid — for each architecture, where to find the network code, the text encoder, and the user-facing nodes.

Image diffusion

Family	LDM code	Text encoder(s)	Nodes
SD 1.x / 2.x	comfy/ldm/modules/diffusionmodules/ (UNet)	comfy/sd1_clip.py, comfy/text_encoders/sd2_clip.py	Built-in (nodes.py)
SDXL / SDXL Turbo	comfy/ldm/modules/diffusionmodules/ (UNet)	comfy/sdxl_clip.py	Built-in
Stable Cascade	comfy/ldm/cascade/	comfy/sdxl_clip.py (CLIP-G)	comfy_extras/nodes_stable_cascade.py
SD3 / SD3.5	comfy/ldm/modules/diffusionmodules/mmdit.py	comfy/text_encoders/sd3_clip.py	comfy_extras/nodes_sd3.py
Pixart Alpha/Sigma	comfy/ldm/pixart/	comfy/text_encoders/pixart_t5.py	comfy_extras/nodes_pixart.py
AuraFlow	comfy/ldm/aura/	comfy/text_encoders/aura_t5.py	(built-in via nodes.py)
HunyuanDiT	comfy/ldm/hydit/	comfy/text_encoders/hydit.py	comfy_extras/nodes_hunyuan.py
Flux / Flux 2 / Kontext / Krea	comfy/ldm/flux/	comfy/text_encoders/flux.py	comfy_extras/nodes_flux.py
Lumina 2	comfy/ldm/lumina/	comfy/text_encoders/lumina2.py	comfy_extras/nodes_lumina2.py
HiDream	comfy/ldm/hidream/	comfy/text_encoders/hidream.py	comfy_extras/nodes_hidream.py
Qwen Image / Edit	comfy/ldm/qwen_image/	comfy/text_encoders/qwen_image.py, qwen_vl.py	comfy_extras/nodes_qwen.py
Hunyuan Image 2.1	comfy/ldm/	comfy/text_encoders/hunyuan_image.py	comfy_extras/nodes_hunyuan.py
Z Image	comfy/ldm/	comfy/text_encoders/z_image.py	comfy_extras/nodes_zimage.py
Chroma Radiance	comfy/ldm/chroma_radiance/	(uses standard CLIP)	comfy_extras/nodes_chroma_radiance.py
Omnigen 2	comfy/ldm/omnigen/	comfy/text_encoders/omnigen2.py	(in nodes.py and edit nodes)
Anima	comfy/ldm/anima/	comfy/text_encoders/anima.py	(in nodes.py)
Ernie	comfy/ldm/ernie/	comfy/text_encoders/ernie.py	(in nodes.py)
LongCat Image	comfy/ldm/	comfy/text_encoders/longcat_image.py	(in edit nodes)

Video diffusion

Family	LDM code	Text encoder(s)	Nodes
Stable Video Diffusion	comfy/ldm/modules/diffusionmodules/	(uses CLIP-Vision)	comfy_extras/nodes_video_model.py
Mochi (Genmo)	comfy/ldm/genmo/	comfy/text_encoders/genmo.py	comfy_extras/nodes_mochi.py
LTX-Video (Lightricks)	comfy/ldm/lightricks/	comfy/text_encoders/lt.py	comfy_extras/nodes_lt.py, nodes_lt_audio.py, nodes_lt_upsampler.py
Hunyuan Video	comfy/ldm/hunyuan_video/	comfy/text_encoders/hunyuan_video.py	comfy_extras/nodes_hunyuan.py
Wan 2.1 / 2.2	comfy/ldm/wan/	comfy/text_encoders/wan.py (UMT5)	comfy_extras/nodes_wan.py, nodes_wanmove.py
Cosmos	comfy/ldm/cosmos/	comfy/text_encoders/cosmos.py	comfy_extras/nodes_cosmos.py
Kandinsky 5	comfy/ldm/kandinsky5/	comfy/text_encoders/kandinsky5.py	comfy_extras/nodes_kandinsky5.py

Audio

Family	LDM code	Text encoder(s)	Nodes
Stable Audio	comfy/ldm/audio/ (in audio subdir)	(its own conditioning)	comfy_extras/nodes_audio.py
ACE Step	comfy/ldm/ace/	comfy/text_encoders/ace.py, ace15.py	comfy_extras/nodes_ace.py

3D

Family	LDM code	Text encoder(s)	Nodes
Hunyuan3D 2.0/2.1	comfy/ldm/hunyuan3d/, comfy/ldm/hunyuan3dv2_1/	(uses CLIP-Vision)	comfy_extras/nodes_hunyuan3d.py
Stable Zero123	comfy/ldm/	(uses CLIP-Vision)	comfy_extras/nodes_stable3d.py

Auxiliary

Family	Code	Notes
ControlNet	comfy/cldm/, comfy/controlnet.py	Multiple variants — classic ControlNet, Union ControlNet (comfy/cldm/control_types.py), MMDiT ControlNet
T2I-Adapter	comfy/t2i_adapter/
GLIGEN	comfy/gligen.py
Upscalers (ESRGAN, Swin*)	comfy_extras/chainner_models/	Vendored chaiNNer port
TAESD (preview)	comfy/taesd/	Tiny autoencoders for fast latent → RGB previews
RT-DETR	comfy/ldm/rt_detr/	Detection
SAM 3	comfy/ldm/sam3/	Segmentation + tracking
SDPose	comfy/ldm/sdpose.py	Human pose
SUPIR	comfy/ldm/supir/	Photo restoration
Frame interpolation	comfy_extras/frame_interpolation_models/	RIFE etc.

How architectures are added

The pattern is consistent (see Lore):

1. Drop the network code under comfy/ldm/<model>/ — usually model.py plus mmdit.py / vae.py / etc.
2. Register a Model<Name> entry in comfy/supported_models.py with the unet key/shape fingerprints.
3. Add a BaseModel subclass in comfy/model_base.py.
4. Drop a text encoder under comfy/text_encoders/ with the right tokenizer.
5. Add user-facing nodes in comfy_extras/nodes_<model>.py (or build off nodes.py if it's just SD-style).
6. Update comfy.sd.load_clip and comfy.sd.load_checkpoint_guess_config if a new encoder combination is needed.

Recent additions are good templates. comfy_extras/nodes_zimage.py (96 lines) is a small one; comfy/ldm/wan/ and comfy_extras/nodes_wan.py are heavyweights that show every step.

Latent format normalization

Each architecture has its own latent statistics (shift/scale and channel count). comfy/latent_formats.py registers per-architecture LatentFormat classes used by previewers and any cross-model conversion. Every BaseModel subclass declares its latent_format.

Cross-references

• The actual loader entry: comfy/sd.py:load_checkpoint_guess_config.
• Detection: comfy/model_detection.py.
• Memory and dtype: see Model management.
• The sampling math: see Sampling pipeline and comfy/model_sampling.py.