sharp-y/sana1.5_1.6b_1024px_fp8

license: apache-2.0 tags:

• text-to-image
• sana
• linear-dit
• quantized
• fp8
• comfyui
• low-vram

---

SANA 1.5 1.6B 1024px (Native FP8 Quantized)

This repository contains a native FP8 (float8_e4m3fn) quantized version of the official SANA 1.5 1.6B (1024px) model by NVLabs / Efficient-Large-Model.

The original checkpoint was released in uncompressed FP32 formats (6.43 GB), making it incredibly heavy for budget setups. This version was precision-aware quantized on an NVIDIA T4 GPU down to a sleek **1.7 GB** file. It runs natively inside ComfyUI without requiring extra command-line casting arguments, making it perfect for 4GB and 6GB VRAM graphics cards (like the GTX 1650 Ti / RTX 3050 mobile).

Precision-Aware Quantization Details

Unlike blind compression, this checkpoint uses selective casting:

• FP8 (float8_e4m3fn): Applied strictly to heavy 2D/4D structural weight matrices (linear layers and convolutions).
• FP16 (float16): Applied to critical 1D vectors (biases, normalization layers, embeddings) to prevent math breakdown and eliminate output artifacts or static noise.

---

🚀 ComfyUI Deployment Guide

To get this model running on low-VRAM hardware, you need three distinct puzzle pieces placed in your ComfyUI directory structure.

1. The Diffusion Model (This File)

• Download sana1.5_1.6b_1024px_fp8.safetensors from this repo.
• Place it here: ComfyUI/models/diffusion_models/

2. The Text Encoder (Gemma 2 2B IT)

Sana requires Google’s Gemma-2-2b-it instead of standard CLIP models.

• Download the monolithic single-file version from Efficient-Large-Model/gemma-2-2b-it (gemma-2-2b-it.safetensors).
• Place it here: ComfyUI/models/text_encoders/
• Alternative for extreme low RAM: If using a dedicated transformers loader node, grab the pre-compiled 4-bit variant from unsloth/gemma-2-2b-it-bnb-4bit and keep the directory folder structure intact under ComfyUI/models/text_encoders/gemma-2-2b-it/.

3. The Custom VAE (DC-AE)

Sana does not use SD1.5 or SDXL VAEs. It uses a specialized Deep Compression Autoencoder.

• Download model.safetensors from the official mit-han-lab/dc-ae-f32c32-sana-1.1 repo.
• Rename it to dc-ae-f32c32-sana-1.1.safetensors and place it here: ComfyUI/models/vae/
• (Note: Do not attempt to quantize this VAE to FP8, or your generation colors and fine details will completely break).

---

⚡ Low-VRAM Execution Strategy

If you are running on a 4GB VRAM GPU, always launch ComfyUI with the low-VRAM flag. This forces the heavy text encoder to unload from your VRAM and shift into system RAM immediately after it finishes parsing your prompt, freeing up the entire GPU for the Sana diffusion process:

python main.py --lowvram

Acknowledgements

All credit goes to the original authors at Efficient-Large-Model for the core Sana 1.5 framework and architecture. This repository is simply a community-driven compression effort to keep open-source AI accessible to localized, modest hardware configurations.