Alibaba’s Qwen group has launched Qwen-Image-2.1, a unified text-to-image technology and picture modifying mannequin. Its visible technology part has 7B parameters throughout 32 single-stream DiT layers. One checkpoint covers text-to-image, multi-reference modifying, native edits, and clear RGBA output.
Is it deployable? Sure, for analysis and analysis. Day 0 assist covers Diffusers, ComfyUI, vLLM-Omni, SGLang, and LightX2V. Industrial deployment wants a separate license from Qwen.
From 20B to 7B
The unique Qwen-Image shipped in August 2025 as a 20B mannequin beneath Apache 2.0. Enhancing lived in a separate Qwen-Picture-Edit checkpoint. Qwen-Picture-2.1 folds each jobs into one mannequin at a few third of the dimensions. Qwen group calls it probably the most balanced and cost-effective mannequin within the Qwen-Picture collection. One vital factor to notice right here for capability planning: the 7B determine covers the diffusion transformer solely. The pipeline additionally masses an 8B Qwen3-VL encoder.
Structure
The GitHub Repo lists 4 elements:
- Transformer: 32 layers, 7B parameters, single-stream design with block-causal consideration.
- Textual content encoder: Qwen3-VL 8B, which encodes textual content directions and situation pictures into one illustration.
- VAE: 64-channel RGBA autoencoder with 16x spatial compression and native transparency.
- Scheduler: Circulation Matching with Euler discrete scheduling and dynamic shifting.
The eye masks is the place the velocity comes from. Textual content tokens use a token-level causal masks. Picture tokens use a chunk-level bidirectional masks inside every picture. Qwen calls this mixed-granularity consideration. The situation prefix sits earlier than the noisy latent, so it by no means attends to it. Its keys and values subsequently keep mounted throughout denoising steps. The mannequin computes textual content and enter pictures as soon as, at step one. It reuses that prefix KV cache for each remaining step. Financial savings develop with the variety of reference pictures, which explains the multi-image velocity declare.
What It Can Do
- Native transparency: Generates RGBA pictures from textual content, edits clear layers, and extracts topics from images. Qwen recommends a set immediate template for clear output.
- Multi-reference modifying: Accepts as much as 10 reference pictures. README examples embrace a gaggle picture from 6 portraits and an outfit from 5 references.
- Native management: Edits can goal areas utilizing circles, painted annotations, or separate masks. Identification is preserved for folks and merchandise.
- Native 2K: Defaults to 2048 x 2048, with 7 supported facet ratios as much as 2752 x 1536.
- Aesthetics: Improved typography, portrait lighting, and effective element. Qwen highlights panoramas, infographics, storyboards, and digital try-ons.
Benchmark: Qwen’s Personal Chart
The research team compares fashions on Qwen-Picture-Bench, Qwen’s in-house benchmark. On that chart, Qwen-Picture-2.1 scores 60.28 general. That locations it above Nano Banana 2.0 at 59.82 and each listed open-weight mannequin. FLUX 2 Max, a 32B open mannequin, sits at 55.33. 6 closed fashions rating larger, led by GPT Picture 2.5 Sunburst at 67.01.
Interactive Explainer
“;toks.forEach(perform(t){h+=’
‘+t[0]+’
‘});
toks.forEach(perform(tq,q){h+=’
‘+tq[0]+’
‘;
toks.forEach(perform(tk,ok){var a=okay(q,ok);h+=”})});
$(“#grid”).innerHTML=h;
$$(“.rowl”).forEach(perform(r){var f=perform(){selq=+r.dataset.q;drawMask()};r.addEventListener(“click”,f);r.addEventListener(“keydown”,perform(e){if(e.key===”Enter”||e.key===” “){e.preventDefault();f()}})});
var t=toks[selq],seen=toks.filter(perform(_,ok){return okay(selq,ok)}).map(perform(x){return x[0]}),seesX=seen.some(perform(s){return s[0]===”X”});
var msg;
if(mode===”full”){msg=”
Everything sees everything
“+t[0]+” attends to all 9 tokens, including the noisy latent X. Because X changes at every step, the prefix keys and values change too, so nothing can be cached.
“}
else if(t[2]===”x”){msg=”
“+t[0]+” is the image being denoised
It attends to “+seen.be a part of(“, “)+”. The latent reads the full prefix and every patch of its own block, so it gets complete context from text and references.
“}
else{msg=”
“+t[0]+” is part of the prefix
It attends to “+seen.be a part of(“, “)+”. It “+(seesX?”sees”:”never sees”)+” the noisy latent X, so its keys and values stay identical across all denoising steps. That is what makes the prefix KV cache valid.
“}
$(“#mexp”).innerHTML=msg;measurement()}
$$(“#seg button”).forEach(perform(b){b.addEventListener(“click”,perform(){$$(“#seg button”).forEach(perform(x){x.classList.take away(“on”)});b.classList.add(“on”);mode=b.dataset.m;drawMask()})});
drawMask();
/* ———- panel 4: decision + rgba ———- */
var ars=[[“1:1”,2048,2048],[“4:3”,2400,1792],[“3:4”,1792,2400],[“3:2”,2528,1696],[“2:3”,1696,2528],[“16:9”,2752,1536],[“9:16”,1536,2752]];
$(“#ar”).innerHTML=ars.map(perform(a,i){return ‘‘}).be a part of(“”);
perform setAr(i){var a=ars[i],w=a[1],h=a[2],m=210,s=m/Math.max(w,h);var f=$(“#frame”);f.fashion.width=Math.spherical(w*s)+”px”;f.fashion.peak=Math.spherical(h*s)+”px”;
$(“#px”).textContent=w+” x “+h;$(“#mp”).textContent=(w*h/1e6).toFixed(2)+” MP”;$(“#lat”).textContent=(w/16)+” x “+(h/16)}
$$(“#ar .chip”).forEach(perform(c){c.addEventListener(“click”,perform(){$$(“#ar .chip”).forEach(perform(x){x.classList.take away(“on”)});c.classList.add(“on”);setAr(+c.dataset.i)})});
setAr(0);
var star=$(“#frame svg path”),dot=$(“#frame svg circle”);
$$(“#bg button”).forEach(perform(b){b.addEventListener(“click”,perform(){$$(“#bg button”).forEach(perform(x){x.classList.take away(“on”)});b.classList.add(“on”);var f=$(“#frame”),v=b.dataset.b;
f.classList.toggle(“checker”,v===”checker”);f.fashion.background=(v===”checker”)?””:(v===”alpha”?”#000″:v);
if(v===”alpha”){star.setAttribute(“fill”,”#fff”);star.setAttribute(“stroke”,”#fff”);dot.setAttribute(“fill”,”#fff”)}else{star.setAttribute(“fill”,”url(#g)”);star.setAttribute(“stroke”,”#2B2A7A”);dot.setAttribute(“fill”,”#fff”)}})});
$(“#copy”).addEventListener(“click”,perform(){var t=”This is an RGBA image with transparency.
var accomplished=perform(){$(“#copy”).textContent=”Copied”;setTimeout(perform(){$(“#copy”).textContent=”Copy”},1400)};
if(navigator.clipboard&&navigator.clipboard.writeText){navigator.clipboard.writeText(t).then(accomplished,accomplished)}else{accomplished()}});
/* ———- panel 5: ship ———- */
var V=[
[“ok”,”Allowed”,”The license grants a royalty-free, worldwide right to use, copy, modify, and distribute the model for non-commercial purposes, which it defines as research or evaluation.”],
[“stop”,”Needs a separate commercial license”,”Commercial use is not covered. Qwen asks teams to request a license at model-business@notice.qwencloud.com before shipping.”],
[“warn”,”Allowed for non-commercial use, with conditions”,”If you use the model or its outputs to build and release another AI model, show u201cBuilt with Qwenu201d or u201cImproved using Qwenu201d in its docs. You cannot use u201cQwenu201d as the primary product name.”],
[“warn”,”Allowed, with notices”,”Give recipients a copy of the license, mark files you changed, and keep the Qwen attribution notice in a Notice file. The non-commercial limit still applies.”]];
perform verdict(i){var v=V[i];$(“#verdict”).innerHTML=”;measurement()}
$$(“.use”).forEach(perform(u){u.addEventListener(“click”,perform(){$$(“.use”).forEach(perform(x){x.classList.take away(“on”)});u.classList.add(“on”);verdict(+u.dataset.u)})});
verdict(0);
measurement();
})();

