Artifact	Size	vs FP16	What it is
FP16 transformer	3.21 GB	1× (100%)	reference
Clark Air (packed)	374 MB	8.6× (≈12%)	packed ternary (`clark-air-sana-1.6b-packed.safetensors`)
Clark Air (unpacked)	3.21 GB	compatibility	this repo's `transformer/`, dequantized bf16, drop-in `diffusers`

Measured ~1.85 bits/weight → 8.6× smaller (374 MB packed ÷ 3.21 GB FP16).

ℹ️ The transformer/ here is the unpacked compatibility format: it loads in stock diffusers with no custom code, so it is the same 3.21 GB as FP16. The actual compressed artifact is the 374 MB packed safetensors in this repo.

Usage (drop-in `diffusers`)

python

import torch
from diffusers import SanaPipeline, SanaTransformer2DModel

transformer = SanaTransformer2DModel.from_pretrained(
    "clark-labs/clark-air-sana-1.6b-1.58bit", subfolder="transformer", torch_dtype=torch.bfloat16)
pipe = SanaPipeline.from_pretrained(
    "Efficient-Large-Model/Sana_1600M_512px_diffusers", transformer=transformer, torch_dtype=torch.bfloat16)
pipe.to("cuda")
pipe("a half empty bottle of red wine", num_inference_steps=20, guidance_scale=4.5).images[0].save("out.png")

About

The transformer weights are quantized to ternary with group-wise scales; a small high-precision tail (~5% of parameters, the conditioning and projection layers) is kept at higher precision.

Base: Sana 1.6B, 512px

License

Quality vs. bit-width

Same prompts and seed across columns: fp16 base | naive 3-bit PTQ | naive 2-bit PTQ | Clark Air 1.58-bit (trained ternary). Naive post-training quantization collapses by 2 bits, while the trained ternary model uses fewer bits yet tracks fp16.

clark-labs/clark-air-sana-1.6b-1.58bit

Clark Air · Sana 1.6B

Footprint (measured)

Usage (drop-in diffusers)

About

License

Quality vs. bit-width

clark-labs/
clark-air-sana-1.6b-1.58bit

Usage (drop-in `diffusers`)