Benchmarks

nunchaku_lite reduces transformer storage and peak CUDA memory while keeping the normal Diffusers pipeline workflow. The measurements below compare unmodified Diffusers pipelines against nunchaku_lite pipelines loaded with FP4 SVDQ checkpoints.

All runs used 1024x1024 generation on an NVIDIA RTX PRO 6000, BF16 runtime dtype, 5 measured runs, and 2 warmup runs. Source data is stored under outputs/benchmark_*/summary.json.

Results

Model	Steps	Latency	Speedup	Peak CUDA memory	Transformer storage
FLUX.2 Klein 4B	4	0.660s -> 0.429s	1.54x	17.32 GB -> 12.55 GB	7.22 GiB -> 2.39 GiB
FLUX.2 Klein 9B	4	1.213s -> 0.677s	1.79x	34.73 GB -> 23.24 GB	16.91 GiB -> 5.40 GiB
Qwen-Image	50	28.322s -> 17.837s	1.59x	57.94 GB -> 31.21 GB	38.05 GiB -> 11.13 GiB
Z-Image Turbo	8	1.851s -> 1.102s	1.68x	21.67 GB -> 14.07 GB	11.46 GiB -> 3.91 GiB

Across this benchmark set, nunchaku_lite reaches up to 1.79x lower latency, up to 46% lower peak CUDA memory, and up to 71% smaller transformer storage.

FLUX.2 Klein 4B

Prompt:

A cat holding a sign that says This is Nunchaku Lite

Original Diffusers	nunchaku_lite

FLUX.2 Klein 9B

Prompt:

A cat holding a sign that says This is Nunchaku Lite

Original Diffusers	nunchaku_lite

Qwen-Image

Prompt:

A coffee shop entrance features a chalkboard sign reading "This is Nunchaku Lite"

Original Diffusers	nunchaku_lite

Z-Image Turbo

Prompt:

a cinematic photo of a glass greenhouse full of tropical plants during golden hour, detailed, natural light

Original Diffusers	nunchaku_lite

Reproducing

The benchmark scripts live under benchmarks/. Each script writes generated images, a summary.json, and, when both Diffusers and nunchaku_lite are run, a comparison.png chart.

python benchmarks/benchmark_qwen_image.py \
  --model-id Qwen/Qwen-Image \
  --checkpoint nunchaku-tech/nunchaku-qwen-image/svdq-fp4_r32-qwen-image.safetensors \
  --precision fp4 \
  --dtype bf16 \
  --runs 5 \
  --warmup-runs 2