tonera/Nepotism_xii-Nunchaku

Model Card (SVDQuant · Nepotism_xii)

Language: English | 中文

Model name

• Model repo: tonera/Nepotism_xii-Nunchaku
• Source checkpoint (full-precision): Nepotism on Civitai — this quantization is derived from the XII (and Flux.1 D family) release; licensing and usage are also subject to upstream terms and Civitai’s policies.
• Full Diffusers layout (VAE, text encoders, scheduler, etc.): {REPO_ID}
• Quantized Transformer weights (for Nunchaku):
  • {REPO_ID}/svdq-fp4_r32-Nepotism_xii-Nunchaku.safetensors
  • {REPO_ID}/svdq-int4_r32-Nepotism_xii-Nunchaku.safetensors

Quantization / inference

• Inference engine: Nunchaku (https://github.com/nunchaku-ai/nunchaku)

Nunchaku targets 4-bit (FP4/INT4) inference to reduce VRAM and latency while preserving quality. The svdq-*_r32-Nepotism_xii-Nunchaku.safetensors files in this repo are SVDQuant-quantized Flux Transformer weights and should be used with FluxPipeline on supported setups.

Install Nunchaku first

• Official install docs (recommended): https://nunchaku.tech/docs/nunchaku/installation/installation.html

(Recommended) Prebuilt wheel

• Prerequisite: Use a PyTorch version that matches the Nunchaku release notes (newer is often better).
• Install: Pick a wheel for your Python, CUDA, and PyTorch from GitHub Releases / Hugging Face / ModelScope, e.g.:

# Example — replace with the correct wheel URL for your torch/cuda/python
pip install https://github.com/nunchaku-ai/nunchaku/releases/download/vX.Y.Z/nunchaku-X.Y.Z+torch2.9-cp311-cp311-linux_x86_64.whl

• Tip (RTX 50 series): When supported by Nunchaku, FP4 weights often give better compatibility and speed (see Nunchaku docs).

Quality reference (N=25 samples)

Summary metrics (higher is generally closer to reference for PSNR/SSIM; lower LPIPS is better).

FP4

Metric	mean	p50	p90	best	worst
PSNR	21.8159	21.9766	29.447	30.8016	13.0762
SSIM	0.811984	0.835828	0.938092	0.944179	0.582228
LPIPS	0.209448	0.178698	0.400699	0.0461679	0.64835

INT4

Metric	mean	p50	p90	best	worst
PSNR	20.8759	20.8797	25.5093	30.0388	14.8672
SSIM	0.78943	0.812346	0.890699	0.913605	0.557165
LPIPS	0.243332	0.203449	0.419361	0.0868137	0.657203

Usage (Diffusers + Nunchaku Flux Transformer)

Set REPO_ID to your Hugging Face repo id or local root. Load the svdq-{precision}_r32-Nepotism_xii-Nunchaku.safetensors transformer and the full pipeline from {REPO_ID} (alongside model_index.json, transformer/, vae/, etc.).

import torch
from diffusers import FluxPipeline

from nunchaku import NunchakuFluxTransformer2dModel
from nunchaku.utils import get_precision

REPO_ID = "tonera/Nepotism_xii-Nunchaku"
MODEL_STEM = "Nepotism_xii-Nunchaku"

if __name__ == "__main__":
    precision = get_precision()  # 'int4' or 'fp4' from GPU
    transformer = NunchakuFluxTransformer2dModel.from_pretrained(
        f"{REPO_ID}/svdq-{precision}_r32-{MODEL_STEM}.safetensors"
    )
    pipeline = FluxPipeline.from_pretrained(
        f"{REPO_ID}",
        transformer=transformer,
        torch_dtype=torch.bfloat16,
    ).to("cuda")

    image = pipeline(
        "A cat holding a sign that says hello world",
        num_inference_steps=50,
        guidance_scale=3.5,
    ).images[0]
    image.save(f"nepotism_xii-{precision}.png")

Licensing follows LICENSE.md in this repo and upstream model terms.

If your Diffusers files live under a subfolder (e.g. local diffusers/), use f"{REPO_ID}/diffusers" for FluxPipeline.from_pretrained and prefix the transformer path the same way.