ComfyUI Node
Nougan Ultimate Model Loader π
Loads any diffusers/UNET checkpoint (bf16, int8 convrot, w6a8/w4a8 convrot, and new w*a8 formats as ComfyUI adds them), applies attention/efficiency backends, and loads up to two CLIPs plus video/default and audio VAEs.
Nougan Ultimate Model Loader π
- MODEL
- CLIP (main)
- CLIP (enhancer)
- VAE (video/default)
- VAE (audio)
βmodel_nameβΎβΊ
βquant_formatautoβΊ
βmodel_typedefaultβΊ
βkitchentrueβΊ
βsage2falseβΊ
βsage3falseβΊ
βtritonfalseβΊ
βflash2falseβΊ
βpytorchfalseβΊ
βsparsefalseβΊ
βvsafalseβΊ
βclip1_nameβΎβΊ
βclip1_typeautoβΊ
βclip1_devicedefaultβΊ
βclip2_enabledfalseβΊ
βclip2_nameβΎβΊ
βclip2_typeautoβΊ
βclip2_devicedefaultβΊ
βvae_nameβΎβΊ
βaudio_vae_enabledfalseβΊ
βaudio_vae_nameβΎβΊ
βheading_modelModelβΊ
βheading_engineEfficienciesβΊ
βheading_clipsText EncodersβΊ
βheading_vaesVAEsβΊ
βsection_ordermodel,engine,clips,vaesβΊ
CategoryNougan/Loaders
Inputs (26)
| Name | Type | Default | Description |
|---|---|---|---|
| model_name | COMBO | Diffusion model / UNET from the diffusers (diffusion_models) folder. | |
| quant_format | COMBO | auto | Expected weight format. 'auto' trusts the checkpoint's embedded quant metadata (int8 / w6a8 / w4a8 convrot etc. are detected automatically). Pick a format to validate it after load β you'll get a warning if the file doesn't match. |
| model_type | COMBO | default | Preset family β drives 'auto' CLIP types and guards (VSA needs minimax_h3). |
| kitchen | BOOLEAN | true | comfy-kitchen INT8 attention (dense backend candidate). |
| sage2 | BOOLEAN | false | SageAttention 2 (dense backend candidate). Beats comfy-kitchen in priority. |
| sage3 | BOOLEAN | false | SageAttention 3 (Blackwell; dense backend candidate). Highest priority. |
| triton | BOOLEAN | false | Opt into ComfyUI's triton backend for comfy-kitchen quant ops (global session flag, like --enable-triton-backend). |
| flash2 | BOOLEAN | false | FlashAttention 2 (dense backend candidate). |
| pytorch | BOOLEAN | false | Force PyTorch SDPA attention explicitly (dense backend candidate). |
| sparse | BOOLEAN | false | Block-sparse attention (sol-attn) on eligible layers; wraps the dense backend. Long sequences only β short ones stay dense. |
| vsa | BOOLEAN | false | Video Sparse Attention (FastH3) β MiniMax-H3 only; conflicts with sparseattention. |
| clip1_name | COMBO | Main text encoder (from text_encoders folder). | |
| clip1_type | COMBO | auto | 'auto' infers from the model type selector. |
| clip1_device | COMBO | default | 'cpu' pins this encoder to system RAM. |
| clip2_enabled | BOOLEAN | false | Load a second CLIP (e.g. for a prompt enhancer chain). |
| clip2_name | COMBO | Second text encoder β used when its switch is on. | |
| clip2_type | COMBO | auto | 36 options: auto, stable_diffusion, sd3, flux, hunyuan_video, wan, +30 |
| clip2_device | COMBO | default | 2 options: default, cpu |
| vae_name | COMBO | Video/default VAE: the image VAE for image models, the video VAE for MiniMax-H3. | |
| audio_vae_enabled | BOOLEAN | false | Also load an audio VAE (MiniMax-H3 audio, LTX-2, ...). |
| audio_vae_name | COMBO | Audio VAE file β used when its switch is on. | |
| heading_model | STRING | Model | β |
| heading_engine | STRING | Efficiencies | β |
| heading_clips | STRING | Text Encoders | β |
| heading_vaes | STRING | VAEs | β |
| section_order | STRING | model,engine,clips,vaes | β |
Outputs (5)
| Name | Type | Description |
|---|---|---|
| MODEL | MODEL | Diffusion model, patched with the chosen efficiency backends. |
| CLIP (main) | CLIP | Main text encoder for the diffusion model. |
| CLIP (enhancer) | CLIP | Second text encoder (e.g. prompt enhancer) β None while its switch is off. |
| VAE (video/default) | VAE | Video/default VAE (image VAE for image models, video VAE for MiniMax-H3 etc.). |
| VAE (audio) | VAE | Audio VAE β None while its switch is off. |