Z-Image-Turbo-MXFP8-AutoRound

An MXFP8 quantization of Tongyi-MAI/Z-Image-Turbo, made with AutoRound using SignRound with 200 optimization iterations.

Quantization

Setting	Value
Method	AutoRound SignRound
Iterations	200
Format	MXFP8 W8A8, group size 32
Quantized component	Diffusion transformer
Excluded layers	adaLN_modulation
Text encoder and VAE	BF16, not quantized
Calibration data	128 prompts from OpenS2V
Calibration steps	8
Calibration guidance scale	7.5
AutoRound version recorded in export	0.17.0

Evaluation

DrawBench, 200 prompts, one image per prompt. Images were generated with official vLLM 0.32.0 and vLLM-Omni 0.32.0 on an NVIDIA B200: 1024×1024, 8 steps, guidance scale 0.0, seed 42 + prompt_id.

Model	CLIP ↑	CLIP-IQA ↑	ImageReward ↑
BF16	31.7831	70.6414	1.0073
MXFP8 SignRound, 200 iterations	31.7907	70.3976	0.9916

CLIP uses openai/clip-vit-base-patch32. CLIP-IQA averages three positive/negative prompt pairs. Both are reported on a 0–100 scale. ImageReward uses ImageReward-v1.0.

These results cover this DrawBench run, not all prompts or benchmarks. GenEval has not been evaluated for this checkpoint.

Usage

Tested with vLLM 0.32.0 and vLLM-Omni 0.32.0.

from vllm_omni.entrypoints.omni import Omni
from vllm_omni.inputs.data import OmniDiffusionSamplingParams

omni = Omni(model="INCModel3/Z-Image-Turbo-MXFP8-AutoRound")
params = OmniDiffusionSamplingParams(
    height=1024,
    width=1024,
    seed=42,
    guidance_scale=0.0,
    num_inference_steps=8,
    num_outputs_per_prompt=1,
)
outputs = omni.generate(
    {"prompt": "A red bench in a park", "modalities": ["image"]},
    sampling_params_list=[params],
)

Quantization settings are stored in the transformer checkpoint. Do not pass a global MXFP8 override for the whole pipeline: the text encoder and VAE remain BF16.

License

Apache 2.0, following the base model.