Model
Trends
.ai
Model
Trends
.ai
최고의 오픈 소스 AI 모델

ComfyUI의 Load Diffusion Model 노드

Load Diffusion Model (UNETLoader) loads a bare denoising network such as Flux, SD3.5, Wan or HunyuanVideo from models/diffusion_models, with a weight_dtype option for fp8 to save VRAM.
노드 팩: ComfyUI core
카테고리: 로더
이 페이지는 아직 번역되지 않아 영어로 표시됩니다.

Load Diffusion Model이(가) 하는 일

Newer model families are not shipped as one all-in-one checkpoint. Flux, SD3.5, Wan 2.x, HunyuanVideo, LTX and Qwen-Image come as a standalone transformer file plus separate text encoders and a VAE. Load Diffusion Model loads only that transformer and outputs a MODEL; the text encoders come from DualCLIPLoader or Load CLIP and the VAE from Load VAE.
The weight_dtype option decides how the weights sit in memory: default keeps the file precision (bf16/fp16), while fp8_e4m3fn and fp8_e5m2 halve the VRAM needed with a small quality cost. fp8_e4m3fn_fast enables faster fp8 matrix math on RTX 40 series cards. GGUF quantised files need the separate ComfyUI-GGUF loader instead.

입력

이름
타입
의미
unet_name
COMBO
The model file from models/diffusion_models (models/unet on older installs).
weight_dtype
COMBO
default, fp8_e4m3fn, fp8_e4m3fn_fast or fp8_e5m2. fp8 options store the weights in 8 bits to fit smaller cards.

출력

이름
타입
의미
MODEL
MODEL
The denoising network, ready for LoRA loaders, ModelSamplingFlux, guidance nodes and KSampler.
AD
설치는 건너뛰세요: 클라우드에 미리 설치된 ComfyUI 워크플로
BitVector는 완성된 ComfyUI 워크플로를 자체 GPU에서 실행합니다. 설치 없음, 누락 노드 없음, 빨간 박스 없음. 휴대폰이나 노트북에서 열고 1분 안에 생성하세요.

Load Diffusion Model 사용법

  1. Download the transformer file (for example flux1-dev.safetensors) into models/diffusion_models.
  2. Download the text encoders into models/text_encoders (clip_l and t5xxl for Flux) and the VAE into models/vae.
  3. Add Load Diffusion Model and pick the file; choose fp8_e4m3fn if you have 12 GB of VRAM or less.
  4. Add DualCLIPLoader for the text encoders and Load VAE for the VAE.
  5. Wire MODEL to KSampler, CLIP to CLIP Text Encode, VAE to VAE Decode, and use EmptySD3LatentImage for the latent.

설정과 팁

  • Flux Dev wants cfg 1.0 with Flux Guidance around 3.5; the normal negative prompt has no effect at cfg 1.
  • If the file is already an fp8 checkpoint, leave weight_dtype on default; converting again gains nothing.
  • T5 can be loaded in fp8 too (t5xxl_fp8_e4m3fn.safetensors) to save another 5 GB of memory.
  • Keep bf16 weights when you fine-tune or train LoRAs; use fp8 for generation only.
  • The node name was UNETLoader for years; searching "UNET" still finds it.

Load Diffusion Model 문제 해결

ERROR: Could not detect model type

원인
The file is not a diffusion transformer ComfyUI knows (for example a full checkpoint, a LoRA or a GGUF file placed in the diffusion_models folder).

해결 방법
Full checkpoints go through Load Checkpoint, LoRAs through Load LoRA and .gguf files through the Unet Loader (GGUF) node from the ComfyUI-GGUF pack. Update ComfyUI if the model family is newer than your install.

Flux output is a blurry mess or pure noise

원인
The sampler settings come from an SD workflow: cfg above 1 without Flux Guidance, an Empty Latent Image instead of EmptySD3LatentImage, or a wrong VAE (SDXL VAE with Flux).

해결 방법
Use ae.safetensors as the VAE, EmptySD3LatentImage, cfg 1.0, sampler euler with scheduler simple or beta, 20-28 steps and a Flux Guidance node at 3.5.

torch.OutOfMemoryError when loading a 24 GB Flux file

원인
The bf16 file needs about 24 GB of VRAM on its own; the text encoder adds 10 GB in fp16.

해결 방법
Switch weight_dtype to fp8_e4m3fn and load t5xxl in fp8. On 8 GB cards use a GGUF Q4 or Q5 file with the GGUF loader and add --lowvram.

fp8_e4m3fn_fast is slower or errors on my GPU

원인
The fast path needs an RTX 40 series (Ada) or newer GPU and a recent PyTorch; older cards fall back or fail.

해결 방법
Choose fp8_e4m3fn without _fast on RTX 30 and older cards.

AD
이 워크플로를 휴대폰에서 실행하세요
이 페이지의 모든 워크플로는 모델과 커스텀 노드까지 갖춘 채 BitVector에 미리 설치되어 있습니다. 하나 고르고 프롬프트를 입력하면 끝. 설정 제로, 다운로드 없음.

Load Diffusion Model에 관한 질문

Where do I put the file, models/unet or models/diffusion_models?

Both work. diffusion_models is the current name; unet is kept for old installs and both folders are scanned.

Does this node load the text encoder?

No. Pair it with DualCLIPLoader (two encoders, Flux and SD3) or Load CLIP (one encoder, Wan and Hunyuan).

Can I use Load LoRA with it?

Yes. Load LoRA takes the MODEL from this node. For Flux LoRAs that only patch the model, LoraLoaderModelOnly avoids needing a CLIP input.

관련 노드

Load Checkpoint
ComfyUI core
Load Checkpoint (CheckpointLoaderSimple) opens a .safetensors or .ckpt model file and hands out the three parts every workflow needs: the diffusion MODEL, the CLIP text encoder and the VAE.
DualCLIPLoader
ComfyUI core
DualCLIPLoader loads two text encoders at once, for example clip_l plus t5xxl for Flux or clip_g plus t5xxl for SD3, and outputs one CLIP object for the prompt nodes.
Load VAE
ComfyUI core
Load VAE (VAELoader) loads a standalone VAE file from models/vae so a checkpoint with a missing or weak VAE decodes clean colours, and so split models like Flux and Wan get their decoder.
EmptySD3LatentImage
ComfyUI core
EmptySD3LatentImage creates the 16-channel empty latent that Flux, SD3 and SD3.5 need; it replaces Empty Latent Image in those workflows.
Flux Guidance
ComfyUI core
Flux Guidance (FluxGuidance) writes the guidance value into the conditioning that Flux Dev was distilled to expect, replacing cfg; 3.5 is the default and 2 to 5 the useful range.
AD
노드 고치기에 지치셨나요? 클라우드에 맡기세요
BitVector는 수백 개의 ComfyUI 워크플로를 빠른 클라우드 GPU에 설치하고 업데이트하고 테스트해 둡니다. Python도, CUDA 오류도, VRAM 한계도 없습니다. 어떤 브라우저에서든 됩니다.

더 많은 ComfyUI 노드

CLIP Text Encode (Prompt)
ComfyUI core
CLIP Text Encode turns a text prompt into CONDITIONING using the model text encoder. One node holds the positive prompt, a second one the negative prompt.
ComfyUI Manager
ComfyUI-Manager
ComfyUI Manager is the extension that installs, updates and fixes custom node packs and models from inside the interface, resolves missing nodes in imported workflows and snapshots your setup.
KSampler
ComfyUI core
KSampler runs the denoising loop: it takes the model, prompts and a latent and produces the finished latent image, controlled by seed, steps, cfg, sampler, scheduler and denoise.
Save Image
ComfyUI core
Save Image writes the IMAGE tensor to ComfyUI/output as a PNG, with the whole workflow embedded in the file metadata so the picture can be dragged back into ComfyUI to restore the graph.
VAE Decode
ComfyUI core
VAE Decode converts the sampled LATENT into a pixel IMAGE with the VAE; VAE Decode (Tiled) does the same in tiles for very large images.
Empty Latent Image
ComfyUI core
Empty Latent Image creates the blank latent canvas (width, height, batch_size) that text-to-image sampling starts from; dimensions must be multiples of 8 and match the model family.