Qwen-Image 2.1 and NSFW: what the stock model draws, and where it breaks
380 test renders on the stock weights. Anime style, no LoRA.
Qwen-Image 2.1 came out in mid-September. It is one checkpoint that draws from a text prompt and also edits from up to ten reference pictures, and the release made a point of its typography. I picked it up for comic pages, because a comic page lives or dies on its lettering. It spelled the dialogue right. When a page called for sex, it drew that too, on the stock weights with no LoRA. The release says nothing about adult content, so I wanted to know how far it actually goes.
The short answer is that it's wildly uneven. It letters a twelve-panel page without a typo, then turns a plain "doggystyle" into two people standing next to each other.
I tested four things: sex positions, well-known characters, character references, and layout and text rendering. Every prompt ran on fixed seeds, two or three per prompt, at 896x1152 (1152x896 for wide scenes, 768x1344 for the phone screen), default settings. I judged every result by eye at full size. The seconds under each tile are how long that image took in my runs.
| Test | What I asked | Renders |
|---|---|---|
| Positions by name | 40 positions, the name and nothing else | 120 |
| Positions described | the 34 it missed, rewritten body by body | 68 |
| Characters | 24 characters, clothed and nude | 96 |
| References | 1 to 4 reference pictures of my own characters | 72 |
| Layout and text | pages of 4 to 12 panels, a cover, an app screen, a character sheet | 24 |
1. Positions: it knows about six by name
Each of the 40 prompts named one position and nothing more, with the same cast every time: a silver-haired woman and a dark-haired man, plus a pink-haired woman and a bearded man when the scene needed them.
It knew 5 of the 6 solo poses (M-legs, all fours, face down with hips up, spreading, fingering). A vibrator shows up in her hand, but she doesn't use it. Of the 34 partnered positions it knew one: the handjob, on all three seeds.
Most of the others came back as a naked couple posing side by side. His penis is often drawn. Nothing touches. Some misses were stranger:
- "Wheelbarrow" drew an actual garden wheelbarrow on most seeds.
- "Strap-on" gave the second woman a penis, three seeds out of three.
- "Blowjob" and "facial" turned into kissing or licking his face.
- The pink-haired woman came out as a man twice, and one foursome had five people.
Positions asked by name alone
Describe the bodies instead
So I stopped naming positions. For the 34 misses, I rewrote each prompt to say where each body is, which way each person faces, and what goes where. Same cast, same room, two seeds each.
That fixed a lot:
- 13 of 34 landed on both seeds: vibrator, missionary, mating press, reverse cowgirl, spooning, full nelson, blowjob, deepthroat, cunnilingus, face sitting, cunnilingus between two women, strap-on (now a real harness) and spitroast.
- 12 landed on one seed, or landed with a visible fault: doggystyle, lotus, standing against a wall, standing carry, piledriver, wheelbarrow (the garden tool was gone, at least), standing doggy, paizuri, footjob, mutual fingering, double penetration and a foursome.
- 9 still missed: prone bone, cowgirl (she faced away, so it drew reverse cowgirl), butterfly, 69, anal from behind (the bodies never joined), facial (no cum, it read as a blowjob), scissoring (one seed put both women in swimsuits), a double blowjob (the two women kissed each other instead) and an FFM threesome (a tangle of limbs).
The same four positions, named and then described
Described body by body: positions it drew on both seeds
Described and still missed
The model understands placement far better than porn vocabulary. I would not fight it on 69. An earlier round of mine tried a top-down 69 nine times and got nine broken bodies, and this round's descriptions didn't save it either.
2. Characters: 8 of 24 from the name alone
I prompted 24 well-known adult characters by name and series, with no description, standing on a plain background. Then I asked for each one nude, lying on a bed, same simple pose for all of them.
Eight were recognisable on both seeds: Tifa Lockhart, 2B, Makima, Raiden Shogun, Chun-Li, Lara Croft, Samus Aran and Rangiku Matsumoto.
Characters it drew from the name alone
Nine were half there, usually with the right colours and the wrong details: Boa Hancock, Mercy (brown hair), Tsunade, Jill Valentine, Esdeath, Ahri, Revy, Nico Robin, and Widowmaker without the blue skin.
Seven missed, and two of them are worth a warning. Kafka from Honkai: Star Rail came out as a dark-haired young man, and Albedo from Overlord as a blond young man in white, with the series named in both prompts. Both names also belong to men in other series (Kafka Hibino in Kaiju No. 8, Albedo in Genshin Impact), which is likely what the model reached for. The rest just didn't come through: Yor Forger became a generic green-haired woman, Yoruichi got red hair, and Lady Dimitrescu, Motoko Kusanagi and Misato Katsuragi came out as nobody in particular.
Names it missed
Taking the clothes off costs some identity. Hair colour survived in the nude version for all eight known characters. The face and the costume pieces that make a character recognisable drift.
Same name, clothed and nude
If your character is on the miss list, describe them: hair, eyes, outfit, one or two signature items. For a name clash, the description is the only thing telling the model which one you mean.
3. References: one picture holds, a group needs the right prompt
For this part I used my own AI-made characters, never photos of real people. Mara has long wavy black hair, green eyes and a beauty mark. Sofia has copper hair and freckles. Nadia has short pink hair and dark skin, and Kai is a dark-haired man.
One reference
One front-facing picture in, four new scenes out: a cafe in a new outfit, a rooftop at night, nude alone, and sex with a man who isn't in the reference. The face and hair held on 16 of 16. All 16 explicit requests came out explicit, solo and couple, whichever kind of reference I used.
One reference picture in, the same woman out
Clothes that the prompt doesn't mention tend to carry over from the reference, so name the outfit you want. The shape of the reference matters too. When I used a wide three-view character sheet instead of a single front view, 4 of 8 results came back landscape even though I asked for 3:4. A single front-view crop never did that.
Two to four references
Each person went in as a separate picture, one per character. My first round named them in plain prose, along the lines of "the first picture is Mara, the second is Kai". Then I read Qwen's own prompt guide, which says that with two or more pictures the prompt has to call them <image1>, <image2> and so on, and never "the first image". So I reran every multi-reference prompt with tags, same pictures, same seeds.
Whoever shows up keeps the face and hair from their picture, prose or tags. Getting the right people into the scene, each of them exactly once, is the hard part.
Dressed scenes were clean on 5 of 8 either way. Tags fixed a doubled Kai at the night market, then doubled Kai at dinner instead. The four-person group photo failed on both seeds in both rounds. Nadia got drawn twice, or the frame held six people, and she was nude in all four photos although I asked for everyone fully dressed. Her reference picture is a nude one, and the picture beat the prompt.
Two and three references, dressed scenes
For the explicit scenes I wrote every prompt four ways: names in prose or tags, and the position either named or described body by body, the way section 1 does it. Four scenes (two women, a woman and a man, three people, four people), two seeds each, so eight renders per way.
| How the prompt was written | Right people, nobody extra (of 8) |
|---|---|
| Names in prose, position named | 3 |
| Tags, position named | 1 |
| Names in prose, bodies described | 3 |
| Tags, bodies described | 7 |
Tags alone made things worse. With the position only named, 5 of the 8 tagged renders had a clone or a stranger in them, and one four-person scene came out with seven people. The act was mostly wrong as well, a kiss where I asked for sex. Describing the bodies with names in prose had its own failures: Mara took Nadia's skin, a second Sofia replaced Mara, and one render added a blowjob I never asked for.
Doing both got 7 of 8 right. The two women and the threesome came out as asked on both seeds, and the four-person scene put each pair on the side of the bed I asked for, both times.
Two women, four ways of writing the prompt
Three people, four ways of writing the prompt
Tags plus the act described
One position never worked with references: missionary. Sixteen renders asked for it, half of them with the bodies described, and every time the woman ended up sitting in his lap facing the viewer. From text alone in section 1, the described missionary landed on both seeds. My guess is that the upright, front-facing reference pictures pull the pose upright, but I haven't tested that.
What tags did not fix
4. Layout and text rendering: the thing it's best at
The release made a point of typography, so this part is all about text: pages with dialogue and sound effects, and designed layouts where words and elements have to land in set places. One prompt per image, every word written into the prompt.
Pages
Six page prompts, two seeds each, plus the two reference pages again with tags:
- 4 panels, two characters, four lines of lettering, explicit: right on both seeds.
- 6 panels, including a text message on a phone screen: right on both, the phone text included.
- 12 panels, twelve lettered beats: every panel drawn and every word spelled right on both seeds. It still made mistakes. One caption landed a panel early, and the undressing panel undressed the wrong woman, on both seeds. The one oral beat became a kiss on one seed. On the other, a second copy of Sofia replaced Mara.
- Japanese lettering: every line right on both seeds. I rendered each seed twice, and one of the four pages dropped a voicing mark (すこく for すごく).
- 4 panels from 2 reference pictures: both usable, prose or tags. From 3 references: with prose names one seed drew Sofia twice. With tags both seeds kept three different people, though the anatomy in the sex panel is a little off.
Pages took 22 to 49 seconds, about the same as a single picture.
Pages from one prompt
An earlier round of 53 pages on the same model adds a few things. Panel counts were right on every layout from 2 to 12 panels, 18 of 18. Japanese dialogue came out right there too, but the scene descriptions got lettered onto the page as extra captions. "Noir" and "manga" did not make it black and white. Asking it to redraw one panel gave me a whole new page every time, 0 of 6. And a "page 2" repeated page 1 on 2 of 12 tries.
Covers, screens and sheets
A page is one kind of layout, so I also tried three designed ones:
- A magazine cover for Mara: a masthead with her head over it, three cover lines, issue number and date, a price and a barcode.
- A dating app profile for Sofia, phone-shaped: the app name, her photo, name and age, distance, a one-line bio, three interest chips, and the two buttons.
- A character sheet for Nadia: front and back views, a small table with height, age and likes, and an 18+ stamp.
All 6 renders came out right. Every word was spelled right and every element sat where I put it: her head over the masthead, the X on the left and the heart on the right, the stamp in the top corner. The nudity held too: Mara nude on the cover, Nadia nude on the sheet, Sofia topless in lace on the app photo, all as asked. Each took 33 to 44 seconds.
These prompts described the women in words, with no reference pictures, which is why Nadia is paler here than in her picture.
A cover, an app screen and a character sheet
One comparison
This is the one place I compared models: the same prompts and seeds on PornMaster Krea 2, an explicit checkpoint that also letters pages.
- Japanese: PornMaster set the balloons vertically, the way manga does, and got 3 of its 4 pieces of text right on both seeds. The fourth broke both times: キレイ came out as キレキレイ on one seed and すレキレイ on the other. Qwen set that balloon horizontally and got it right on both.
- English: as good as Qwen, with its own small slips. On the 12-panel page one seed printed stray labels from the prompt in the margins and put a caption in the wrong panel, and the other lettered "FINALLY" twice.
- Explicit beats: PornMaster draws them more readily. It drew the sex panel on both 4-panel pages and the oral panel on both 12-panel pages. Qwen drew that oral panel on neither.
- But it keeps clothes on. Kai never took off his T-shirt, and Sofia wore her jacket through the whole 12-panel page. In the oral panel it drew Sofia twice, on both seeds.
Same prompts on PornMaster Krea 2
The Japanese balloon up close
Settings worth knowing
These come from an earlier round on the same model.
- The negative prompt does nothing at default settings. It only works with CFG above 1.
- CFG 4 with an anatomy negative cut fused bodies from 11 to 5 in that round. It also drew the man with a woman's body on 4 pages (none at default), and it ran about 1.6 times slower, 49s against 31s. I leave it off.
- 2K took about four times as long as the default size (134s against 33s a page), fixed little, and made title lettering worse.
Keep mirrors out of the scene. The reflection became a second person on 3 of 3 tries.
Which one should you use
Use Qwen-Image 2.1 when the text has to be right: anime pages with real dialogue in English or Japanese, covers, app screens, character sheets. It's also good for keeping one character consistent from a single reference, and for scenes you're willing to describe body by body. With two to four of your own characters, tag every picture and describe every body. Skip it when you want to type a position name and get the position, or when you need missionary with reference pictures in the prompt.
Everything here is anime style, and I haven't run the photo-style version of these tests yet. If you'd like to try your own prompts without setting anything up, our free anime porn generator runs in the browser with no signup.