ComfyUI Sana Custom Node

Support my GPU rental fee for Finetune & Request the desired model

https://coff.ee/kpsss34

A custom node for ComfyUI that supports Sana text-to-image models (600M/1.6B parameters) with advanced features including LoRA support, PAG (Perturbed-Attention Guidance), and optimized VRAM usage.

Installation

1. Clone to ComfyUI custom nodes directory:

   cd ComfyUI/custom_nodes/
   git clone https://github.com/kpsss34/ComfyUI-kpsss34-Sana.git
   cd ComfyUI-kpsss34-Sana
   

2. Install dependencies:

   pip install -r requirements.txt
   

3. Create model directories:

   mkdir -p ComfyUI/models/sana
   mkdir -p ComfyUI/models/loras/sana
   

Model Setup

Download Sana Models

1. Download Sana models from Hugging Face:
   • FINETUNE MODELS

1.1 How to donwload

Ex.repo [kpsss34/SANA600.fp8_illustrious_SFW_V1]

in root/ComfyUI/models/sana

• git clone https://huggingface.co/kpsss34/SANA600.fp8_illustrious_SFW_V1

2. Place model folders in ComfyUI/models/sana/:

   ComfyUI/models/sana/
   ├── SANA600.fp8_illustrious_SFW_V1/
   │   ├── text_encoder/
   │   ├── transformer/
   │   ├── vae/
   │   └── ...
   └── SANA600.fp8_illustrious_SFW_V2/
       ├── text_encoder/
       ├── transformer/
       ├── vae/
       └── ...
   

LoRA Setup (Optional)

Place LoRA files in ComfyUI/models/loras/sana/:

ComfyUI/models/loras/sana/
├── my_lora_1/
│   ├── pytorch_lora_weights.safetensors
│   └── adapter_config.json
└── my_lora_2/
    ├── pytorch_lora_weights.safetensors
    └── adapter_config.json

Usage

The node package provides three main components:

1. Sana Model Loader

• Purpose: Load and configure Sana models
• Options:
  • model_name: Select from available Sana models
  • vram_mode: Choose "low" (2-4GB) or "high" (12GB+)
  • use_pag: Enable Perturbed-Attention Guidance
  • torch_compile: Enable model compilation for performance

2. Sana LoRA Loader

• Purpose: Apply LoRA weights to loaded models
• Options:
  • lora_name: Select LoRA or "None"
  • lora_scale: Adjust LoRA influence (0.0-2.0) Note: Cannot be used with PAG simultaneously

3. Sana Sampler

• Purpose: Generate images using the configured model
• Options:
  • prompt: Text description for generation
  • negative_prompt: What to avoid in generation
  • width/height: Image dimensions (512px-2048px)
  • guidance_scale: Prompt adherence strength (3.0-7.0)
  • pag_scale: PAG strength (0.0-10.0, only with PAG enabled) Note:Cannot be used PAG with LoRA
  • num_inference_steps: Generation steps (1-100)
  • seed: Random seed (-1 for random)

Workflow Example