Repost From Huggingface <a target="_blank" rel="ugc" href="https://huggingface.co/Qwen/Qwen-Image-2.1">https://huggingface.co/Qwen/Qwen-Image-2.1</a><h2 id="introduction">Introduction</h2>We are excited to open-source Qwen-Image-2.1, a unified text-to-image generation and image editing model in the Qwen family. With just 7B parameters in its visual generation component (32 Single-Stream DiT layers), Qwen-Image-2.1 balances generation quality, inference efficiency, and versatility.Four key improvements define this release:<ul><li>Compact and Efficient — A lightweight architecture with mixed-granularity attention and prefix KV cache reuse delivers strong image quality at low computational cost.</li><li>Native Transparency, Unified Creation and Editing — Generate regular or transparent (RGBA) images from text, edit transparent layers, and extract subjects from photographs—all in one model.</li><li>Versatile Editing — Support up to 10 reference images, specify local edits via circles, painted annotations, or separate masks, and preserve identity for people and products.</li><li>Realistic Textures and Refined Aesthetics — Improved typography, portrait lighting, and fine details for more visually compelling results.</li></ul>

Qwen Image 2.1

qwenImage21_21_txt_bf16.safetensors

qwenImage21_21_txt_int8.safetensors

qwenImage21_21_bf16.safetensors

qwenImage21_21_int8.safetensors

qwen_image_2.1_vae_bf16.safetensors

<h3 id="about-these-variants">About these variants</h3>Four precision levels are available for Qwen-Image-2.1, trading VRAM and speed against generation quality:- BF16 — full-precision reference (~14 GB). Highest quality, largest footprint. - INT8 (W8A8) — 8-bit weights + 8-bit activations (~7 GB). Near-lossless quality, ~2× smaller than BF16, runs on INT8 tensor cores (RTX 30-series and up). - INT6 (W6A8) — 6-bit weights + 8-bit activations (~5.5 GB). Middle-ground: smaller than INT8 with only a small quality trade-off. - INT4 (W4A8) — 4-bit weights + 8-bit activations (~4 GB). Smallest footprint; uses ConvRot with a per-tensor codebook that decodes to INT8 for compute, so it runs on the same INT8 hardware as W8A8. Larger quality trade-off than INT6. Qwen-Image-2.1 7B — INT4 (W4A8) &amp; INT6 (W6A8) ConvRot for ComfyUI INT4 and INT6 quantized weights of Qwen-Image-2.1 for fast, low-VRAM inference in ComfyUI. This is a modified (quantized) version of the Qwen-Image-2.1 model. It is not an official Qwen release and is not endorsed by the Qwen team. About Qwen-Image-2.1 A unified text-to-image generation and image editing model in the Qwen family. With just 7B parameters in its visual generation component (32 Single-Stream DiT layers), Qwen-Image-2.1 balances generation quality, inference efficiency, and versatility. Highlights - Efficient Image Generation — combines strong visual performance with fast inference and a compact design, making high-quality image creation accessible across a wide range of creative workflows. - Flexible Creative Control — supports diverse inputs, outputs, and localized edits, giving creators the flexibility to explore ideas and refine details within a unified workflow. Key improvements in 2.1 - Compact and Efficient — lightweight architecture with mixed-granularity attention and prefix KV cache reuse delivers strong image quality at low computational cost. - Native Transparency, Unified Creation and Editing — generate regular or transparent (RGBA) images from text, edit transparent layers, and extract subjects from photographs — all in one model. - Versatile Editing — support up to 10 reference images, specify local edits via circles, painted annotations, or separate masks, and preserve identity for people and products. - Realistic Textures and Refined Aesthetics — improved typography, portrait lighting, and fine details for more visually compelling results. License Qwen-Image-2.1 is licensed under the [<a target="_blank" rel="ugc" href="https://huggingface.co/Qwen/Qwen-Image-2.1/blob/main/LICENSE">Qwen Research License</a>]<a target="_blank" rel="ugc" href="https://huggingface.co/Qwen/Qwen-Image-2.1/blob/main/LICENSE"> </a>These files are a quantized derivative and are distributed under the same license — see the original repository for full terms, permitted uses, and any commercial-use restrictions.

Qwen Image 2.1/Turbo INT4(W4A8)/INT6(W6A8)/INT8

qwenImage21INT8INT4_int8_txt.safetensors

qwenImage21INT8INT4_int8.safetensors

INT8

Turbo INT4(W4A8)

Turbo INT6(W6A8)

Turbo INT8

Base INT6(W6A8)

Base INT4 (W4A8)

Originally released by <a rel="ugc" href="https://huggingface.co/Qwen/Qwen-Image-2.1">Qwen on Hugging Face</a>. All credit goes to the Qwen team.
Qwen Image 2.1 combines text-to-image generation and image editing in a single model. Its visual generation component has 7 billion parameters across 32 single-stream DiT layers, balancing image quality, efficiency and versatility.
The release introduces four main improvements:
<ul>
 <li>Efficient architecture: Mixed-granularity attention and reuse of cached prefixes reduce computation while maintaining image quality.</li>
 <li>Generation and editing with transparency: Create regular or transparent RGBA images, edit transparent layers and isolate subjects from photographs.</li>
 <li>Flexible editing: Work with up to ten reference images, guide local changes using circles, painted annotations or separate masks, and preserve the identity of people and products.</li>
 <li>Improved visual detail: Refined typography, portrait lighting, realistic textures and finer details.</li>
</ul>
Released under the <a rel="ugc" href="https://huggingface.co/Qwen/Qwen-Image-2.1/blob/main/LICENSE">Qwen Research License Agreement</a>. See the <a rel="ugc" href="https://huggingface.co/Qwen/Qwen-Image-2.1">original model card</a> for examples and usage instructions.

Files with SHA256: cb74113cb03faecd79611b01fd7fd642f0aa60d6f0b95086abee214d75eaa57d