HelloSun/SANA1.5_1.6B_1024px-OpenVINO-INT4

SANA1.5_1.6B_1024px OpenVINO INT4

Source: Efficient-Large-Model/SANA1.5_1.6B_1024px (export uses Efficient-Large-Model/SANA1.5_1.6B_1024px_diffusers) Pipeline: SanaPipeline / OVDiffusionPipeline — official example guidance_scale=4.5, num_inference_steps=20, 1024x1024. Quant: NNCF weight-only via OVQuantizer + OVPipelineQuantizationConfig — transformer + text_encoder = OVWeightQuantizationConfig(bits=4,sym=False,group_size=128,group_size_fallback="adjust",ratio=1.0), rest INT8, ov_config=OVConfig(quantization_config=...). Size: FP16 ~8.5G → INT4 ~3.1G (-63%). See REPORT.md + outputs/benchmark.json.

每組 prompt 皆有 step=5 vs step=20 並列對比 (1024)，另附 512px 對照在 examples/*_512_*。

01_hanfu — seed42

Young Chinese woman in red Hanfu, intricate embroidery, impeccable makeup, red floral forehead pattern, elaborate high bun, golden phoenix headdress, soft-lit outdoor night background, silhouetted tiered pagoda, blurred colorful distant lights, photorealistic, ultra detailed, 8k

steps=5 (11~14s)	steps=20 (35~36s)

02_astronaut — seed43

Astronaut in a jungle, cold color palette, muted colors, detailed, 8k, photorealistic, cinematic lighting

steps=5	steps=20

03_taipei — seed44

Cyberpunk street in Taipei at night, heavy rain, neon signs with text 'TAIPEI' and Chinese characters '台北', reflections on wet asphalt, crowded night market, cinematic, ultra detailed

steps=5	steps=20

04_shiba — seed45

Cute Shiba Inu wearing a tiny astronaut helmet, sitting in a field of sunflowers under a starry sky, dreamy illustration, vibrant colors, high quality

steps=5	steps=20

05_ink — seed46

Traditional Chinese ink wash landscape, misty mountains, a small pagoda on a cliff, cranes flying, minimalist, elegant, high aesthetic quality

steps=5	steps=20

512px control: examples/*_512_seed*_steps{5,20}.png (same prompt/seed).

Benchmark (Intel Xeon Platinum 8559C, 192 logical CPU, OpenVINO CPU, compile=True, guidance=4.5)

• load+compile: 8.1s, RSS 2.44 GiB → peak ~8.9GiB
• avg step 1.61.9s; step5 total 1114s; step20 total 3539s

file	total (s)	avg step (s)
01_hanfu_1024_seed42_steps5.png	14.28	1.68
01_hanfu_512_seed42_steps5.png	13.41	1.90
01_hanfu_1024_seed42_steps20.png	35.63	1.60
01_hanfu_512_seed42_steps20.png	35.99	1.64
02_astronaut_1024_seed43_steps5.png	11.29	1.62
02_astronaut_512_seed43_steps5.png	11.36	1.63
02_astronaut_1024_seed43_steps20.png	36.21	1.65
02_astronaut_512_seed43_steps20.png	35.64	1.64
03_taipei_1024_seed44_steps5.png	10.97	1.60
03_taipei_512_seed44_steps5.png	11.15	1.60
03_taipei_1024_seed44_steps20.png	36.84	1.69
03_taipei_512_seed44_steps20.png	37.38	1.72
04_shiba_1024_seed45_steps5.png	11.59	1.67
04_shiba_512_seed45_steps5.png	12.30	1.82
04_shiba_1024_seed45_steps20.png	38.83	1.76
04_shiba_512_seed45_steps20.png	36.65	1.66
05_ink_1024_seed46_steps5.png	11.74	1.70
05_ink_512_seed46_steps5.png	11.71	1.68
05_ink_1024_seed46_steps20.png	36.07	1.64
05_ink_512_seed46_steps20.png	36.21	1.65

Full per-step log: outputs/benchmark.json, prompts: outputs/prompts.txt, full report: REPORT.md.

Versions: diffusers 0.37.1, transformers 5.5.4, tokenizers 0.22.2, huggingface-hub 1.21.0, optimum 2.3.0, optimum-intel 2.2.0, openvino 2026.4.0, nncf 3.4.0, torch 2.14.1, pillow 12.3.0, psutil 7.2.2.

Usage

from optimum.intel import OVDiffusionPipeline
import torch
pipe = OVDiffusionPipeline.from_pretrained("HelloSun/SANA1.5_1.6B_1024px-OpenVINO-INT4", compile=True)
image = pipe(prompt="a cyberpunk cat with a neon sign that says \"Sana\"",
             num_inference_steps=20, guidance_scale=4.5,
             height=1024, width=1024,
             generator=torch.Generator().manual_seed(42)).images[0]
image.save("out.png")
# fast preview: num_inference_steps=5

python inference_int4.py --prompt "..." --seed 42 --steps 20 --guidance 4.5 --size 1024 --output out.png
python generate5.py

Reproduce

optimum-cli export openvino -m Efficient-Large-Model/SANA1.5_1.6B_1024px_diffusers --task text-to-image --library diffusers --weight-format fp16 ./sana15_1.6b_1024px-ov-fp16
python quantize_int4.py
python generate5.py

Files: quantize_int4.py, generate5.py, inference_int4.py, REPORT.md, outputs/, examples/.