Modelpublic

QuantFunc/Qwen-Image-Edit-Series

sourceHugging Faceotherupdated 3mo agoView on Hugging Face

5likes17downloads

Model Card

License & Attribution

These are quantized derivative weights of `Qwen/Qwen-Image-Edit-2511` (Qwen-Image-Edit-2511).

—Modifications: the original weights were quantized (e.g. W4A4 / FP4 / INT4 / FP8) and repackaged for the QuantFunc inference engine — a "modification" under Apache-2.0 §4(b).
—Upstream license: the base model is licensed under the Apache License 2.0, included here as `LICENSE-APACHE`; the upstream copyright and attribution notices are retained.
—This derivative: the QuantFunc quantization & packaging are additionally provided under the QuantFunc Model License (see `LICENSE`).
—This repository is not affiliated with or endorsed by the upstream model authors.
—Text/vision encoder: the bundled Qwen2.5-VL-7B text/vision encoder is licensed under the Apache License 2.0 (as distributed with `Qwen/Qwen-Image-Edit-2511`).

QuantFunc

<p align="center"> 🤗 <a href="https://huggingface.co/QuantFunc">Hugging Face</a>  |  🤖 <a href="https://www.modelscope.cn/profile/QuantFunc">ModelScope</a>  |  💻 <a href="https://github.com/RealJonathanYip/ComfyUI-QuantFunc">GitHub</a>  |  💬 <a href="#wechat">WeChat (微信)</a>  |  🎮 <a href="https://discord.gg/jCp9TpFWcn">Discord</a> </p>

⚡ Qwen-Image-Edit-2511 — Lighting-backend pre-quantized image editing. Instruction-based edits 2x–11x faster with the QuantFunc plugin.

Pre-quantized Qwen-Image-Edit-2511 for the Lighting engine — edit input images from text instructions (e.g. a 1K edit in ~1s on RTX 4090), for RTX 30-series and above.

Powered by the [QuantFunc ComfyUI plugin](https://github.com/RealJonathanYip/ComfyUI-QuantFunc) — the fastest diffusion inference engine:

—🚀 2x–11x speedup over standard BF16/FP16 Python pipelines (pre-exported → even faster loading).
—⚙️ Native C++/CUDA (libquantfunc.so / quantfunc.dll) with zero Python model dependencies.
—🧩 Dual engine (SVDQ offline + Lighting runtime 4-bit), zero-cost LoRA stacking, reference-image editing & inpainting.
—🟢 Full GPU coverage — RTX 20/30/40/50 · A100/H100/H200/B100/B200/GB300 · RTX 6000 Ada / PRO Blackwell (CUDA 12 & 13); native FP4 on Blackwell.

👉 Install the plugin: https://github.com/RealJonathanYip/ComfyUI-QuantFunc

Qwen-Image-Edit-Series

Pre-quantized Qwen-Image-Edit-2511 image editing model series by QuantFunc, with both Lighting and SVDQ backend inference support.

Overview

Qwen-Image-Edit-2511 is an image editing diffusion model distilled from Alibaba Qwen team's image editing model. It can edit input images according to text instructions and supports multi-reference image inputs.

With the latest QuantFunc ComfyUI plugin, inference achieves 2x–11x speedup over mainstream frameworks — e.g. Qwen-Image-Edit 1K image inference in ComfyUI reduced from 9.6s to 1.6s (tested on RTX 4090).

Hardware Requirements

—Supports NVIDIA RTX 30 series and above
—RTX 20 series does not support BF16, which causes significant precision loss in Qwen series model quantization scenarios. Therefore, the 20 series currently only supports Z-Image models.

Compatibility

—The base models in this repository are compatible with any version of Qwen-Image-Edit transformer weights
—The QuantFunc code plugin and ComfyUI plugin are 100% compatible with previous versions of Qwen-Image-Edit models

Directory Structure

Qwen-Image-Edit-Series/
├── qwen-image-edit-series-50x-above-base-model/    # Base model, optimized for RTX 50 series and above
│   ├── text_encoder/          # Qwen2.5-VL text encoder (pre-quantized)
│   ├── vision_encoder/        # Qwen2.5-VL vision encoder (pre-quantized)
│   ├── vae/                   # VAE encoder + decoder (~242MB)
│   ├── tokenizer/             # Tokenizer
│   ├── processor/             # Image preprocessor
│   ├── scheduler/             # Scheduler config
│   ├── model_index.json
│   └── quantfunc_config.json
├── qwen-image-edit-series-50x-below-base-model/    # Base model, optimized for RTX 50 series and below
│   └── (same structure as above)
├── transformer/
│   ├── config.json
│   ├── qwen-image-2511-50x-above-lighting-4steps.safetensors           # RTX 50+ Lighting 4-step
│   ├── qwen-image-2511-50x-above-lighting-4steps-prequant.safetensors  # RTX 50+ Lighting pre-quantized
│   ├── qwen-image-2511-50x-above-svdq-4steps.safetensors               # RTX 50+ SVDQ 4-step
│   ├── qwen-image-2511-50x-above-svdq.safetensors                      # RTX 50+ SVDQ full-step
│   ├── qwen-image-2511-50x-below-lighting-4steps.safetensors           # RTX 30/40 Lighting 4-step
│   └── qwen-image-2511-50x-below-lighting-4steps-prequant.safetensors  # RTX 30/40 Lighting pre-quantized
├── prequant/                                                # Pre-quantized modulation weights
│   ├── qwen-image-edit-2511-50x-above.safetensors           # RTX 50+ mod weights
│   ├── qwen-image-edit-2511-50x-below.safetensors           # RTX 30/40 mod weights
│   └── qwen-image-edit-2509-50x-above.safetensors           # Legacy 2509 mod weights
└── precision-config/                                        # Lighting precision config samples
    ├── 50x-above-fp4-sample.json                            # FP4 config for RTX 50+
    └── 50x-below-int4-sample.json                           # INT4 config for RTX 30/40

Model Variants

By GPU Generation

Variant	Target GPU	Description
50x-above	RTX 50 series and above (Blackwell: RTX 5090, B200, GB200)	Optimized for Blackwell architecture
50x-below	RTX 30/40 series + pre-Blackwell datacenter (A100, H100, H200)	Broadly compatible

By Inference Backend

Backend	File Suffix	Features
Lighting 4-step	`*-lighting-4steps.safetensors`	Fastest inference with fused operators
SVDQ 4-step	`*-svdq-4steps.safetensors`	4-step distilled + SVDQ quantization, runtime LoRA support
SVDQ full-step	`*-svdq.safetensors`	Default step count inference, runtime LoRA support

The base-model and transformer must use the same variant (both above or both below).

Quick Start

Download

bash

pip install huggingface_hub

python

from huggingface_hub import snapshot_download
model_dir = snapshot_download('QuantFunc/Qwen-Image-Edit-Series')

Lighting Backend Inference

bash

quantfunc \
  --model-dir Qwen-Image-Edit-Series/qwen-image-edit-series-50x-above-base-model \
  --transformer Qwen-Image-Edit-Series/transformer/qwen-image-2511-50x-above-lighting-4steps.safetensors \
  --auto-optimize --model-backend lighting \
  --ref-image input.png \
  --prompt "make the sky more purple and add stars" \
  --output output.png --steps 4

SVDQ Backend Inference

bash

quantfunc \
  --model-dir Qwen-Image-Edit-Series/qwen-image-edit-series-50x-above-base-model \
  --transformer Qwen-Image-Edit-Series/transformer/qwen-image-2511-50x-above-svdq-4steps.safetensors \
  --auto-optimize --model-backend svdq \
  --ref-image input.png \
  --prompt "change the background to a beach scene" \
  --output output.png --steps 4

SVDQ + LoRA

bash

quantfunc \
  --model-dir Qwen-Image-Edit-Series/qwen-image-edit-series-50x-above-base-model \
  --transformer Qwen-Image-Edit-Series/transformer/qwen-image-2511-50x-above-svdq-4steps.safetensors \
  --auto-optimize --model-backend svdq \
  --lora /path/to/style_lora.safetensors:0.8 \
  --ref-image input.png \
  --prompt "apply anime style to the image" \
  --output output.png --steps 4

SVDQ && Lighting Backend

This repository provides both Lighting and SVDQ backend pre-quantized models:

Feature	Lighting	SVDQ
Quantization	Per-layer mixed precision (FP4/INT4/FP8/INT8)	Nunchaku-based holistic pre-quantization + Rotation quantization
LoRA Integration	Real-time quantization — build a custom model in 5 minutes with zero speed loss, integrating any number of LoRAs	Runtime low-rank pathway
Ecosystem	QuantFunc native	Compatible with the widely-adopted Nunchaku ecosystem, enhanced with Rotation quantization and Auto Rank dynamic rank optimization
Flexibility	Per-layer precision control	Precision fixed at export time
Use Cases	Rapid personal model customization, batch LoRA integration	Leverage Nunchaku ecosystem, runtime dynamic LoRA

Pre-quantized Modulation Weights (prequant/)

The prequant/ directory contains pre-quantized modulation (mod) weights extracted from SVDQ models. These are used with the Lighting backend to provide high-quality modulation without runtime quantization overhead.

Usage with Lighting backend:

bash

quantfunc \
  --model-dir Qwen-Image-Edit-Series/qwen-image-edit-series-50x-above-base-model \
  --model-backend lighting \
  --precision-config Qwen-Image-Edit-Series/precision-config/50x-above-fp4-sample.json \
  --mod-weights Qwen-Image-Edit-Series/prequant/qwen-image-edit-2511-50x-above.safetensors \
  --rotation-block-size 256 \
  --ref-image input.png --prompt "edit instruction" \
  --steps 4 --auto-optimize

Alternatively, use the pre-quantized Lighting transformer for instant loading (no runtime quantization):

bash

quantfunc \
  --model-dir Qwen-Image-Edit-Series/qwen-image-edit-series-50x-above-base-model \
  --transformer Qwen-Image-Edit-Series/transformer/qwen-image-2511-50x-above-lighting-4steps-prequant.safetensors \
  --model-backend lighting \
  --ref-image input.png --prompt "edit instruction" \
  --steps 4 --auto-optimize

Precision Config (precision-config/)

Sample per-layer precision configurations for the Lighting backend:

File	Target GPU	Precision
`50x-above-fp4-sample.json`	RTX 50+	FP4 attention + AF8WF4 MLP fc2 + INT8 modulation
`50x-below-int4-sample.json`	RTX 30/40	INT4 all layers + INT8 modulation

These configs control the quantization precision of each transformer sub-layer. Customize them for your speed/quality trade-off.

Related Repositories

—QuantFunc/Qwen-Image-Series — Qwen-Image text-to-image (60 layers)
—QuantFunc/Z-Image-Series — Z-Image-Turbo text-to-image (lightweight, fast)

License

The pre-quantized model weights in this repository are derived from the original models. Users must comply with the original model's license agreement. The QuantFunc inference engine and its plugins (including the ComfyUI plugin) are licensed separately — see official QuantFunc channels for details.

For models quantized from commercially licensed models, users are responsible for obtaining the necessary commercial licenses from the original model providers.

Community

Join our community for support, updates, and discussions:

—🎮 Discord server
—💬 Scan the QR code below to join our WeChat group: