ixim/Qwen-Image-2512-Quanto-INT8-Full

Qwen-Image-2512 — Quanto INT8 (Full)

This repository provides a directly-loadable Diffusers model folder for Qwen/Qwen-Image-2512, with the DiT transformer quantized to INT8 using Diffusers QuantoConfig / Optimum-Quanto.

• Repo type: Full model directory (ready for DiffusionPipeline.from_pretrained())
• Quantized component: transformer/ (INT8)
• Other components (text_encoder/, vae/, scheduler/, tokenizer/) are included for convenience.

Base model license: Apache-2.0 (see base model page).
This is a quantized derivative of the base model.

Install

pip install -U diffusers transformers accelerate safetensors optimum-quanto

Usage (Text-to-Image)

import torch
from diffusers import DiffusionPipeline

model_id = "ixim/Qwen-Image-2512-Quanto-INT8-Full"

pipe = DiffusionPipeline.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
)

# Recommended for desktop GPUs / limited VRAM
pipe.enable_attention_slicing()
pipe.vae.enable_tiling()
pipe.vae.enable_slicing()

# More stable on Windows desktop GPUs (requires accelerate)
try:
    pipe.enable_model_cpu_offload()
except Exception:
    pipe.to("cuda")

image = pipe(
    "a clean product poster, studio lighting, sharp, high quality",
    height=512,
    width=512,
    num_inference_steps=10,
).images[0]

image.save("out.png")

Notes

• Start from 512px and 10 steps, then scale up gradually.
• If you enable CFG, provide negative_prompt to activate it (pipeline-dependent).

Acknowledgements

• Base model: Qwen/Qwen-Image-2512 (Apache-2.0)
• Quantization: Diffusers QuantoConfig / Optimum-Quanto