What replaced the size field in Qwen Image 2.1?

Two fields. aspect_ratio chooses the shape from fifteen values and resolution chooses the 1k, 1.5k or 2k tier, where Qwen-Image 2.0 carried both in one string such as 1024*1024. A square render is now aspect_ratio 1:1 plus an explicit tier; leave the tier out and the request runs at the 1k default.

What happens if I send Qwen Image 2.1 a field it does not declare?

The request is rejected rather than trimmed. Text-to-image declares prompt, aspect_ratio, resolution, output_format and seed, and nothing else is accepted, so a leftover size or an extra parameter carried over from another image API fails at submit time instead of disappearing quietly on the way through.

Which Qwen Image 2.1 field changes what a render costs?

resolution. The shape, the container and the seed quote the same whatever you set them to, so the tier is the only lever on this endpoint. The estimate shown above the form before you submit is the authoritative number for the request you have assembled.

Does Qwen Image 2.1 return transparent PNGs?

Not on this route. Choosing png changes the container, not the channel count, and what comes back is three-channel RGB. Asking in the prompt for a subject on a clean background does work — it returns the subject on a white field rather than an alpha channel, so keying that white is a step you do downstream.

Do I have to declare an uncensored prompt on Qwen Image 2.1?

There is no field for it. The request carries no maturity flag, there is no separate uncensored endpoint to route to, and no filter sits in front of the model. The rules that do apply are the ordinary ones: everyone depicted is an adult, and the result is not passed off as a real person.

Is Qwen Image 2.1 text-to-image synchronous?

No. Submitting returns a task handle and the image arrives afterwards, so a client needs a polling loop and somewhere to keep that handle. Once the task reaches a terminal state the result carries the image as a platform asset URL.

Qwen Image 2.1 API — Uncensored NSFW Images & Edits

alibaba/qwen-image-2.1/text-to-image

Qwen Image 2.1 text-to-image is Alibaba's open-weight model: a 5,000-character prompt becomes one image at 1k, 1.5k or 2k in fifteen aspect ratios, as JPEG, PNG or WebP. Transparent backgrounds are available, and there is no content filter.

From $0.024/image

Run model View API

Input

Text to image

Prompt337/5000

Aspect ratio

Resolutionaffects price

Prompt expansion

Background

OpaqueTransparent

Output format

JPEGPNGWebP

Seed

Safety checkerOff

SpicyAPI does not run a safety checker on this model. Your prompt and media go to the model as written, and the model itself decides what comes back.

Checking your account before enabling submission and pricing.

Output

Ready for a new run

Example result

Animate

Edit

Split into layers

{
  "model": "alibaba/qwen-image-2.1/text-to-image",
  "input": {
    "aspect_ratio": "3:4",
    "resolution": "1k",
    "enable_prompt_expansion": true,
    "background": "opaque",
    "output_format": "jpeg",
    "prompt": "Photograph of an adult woman in her thirties leaning in a sunlit bedroom doorway in a short ivory silk robe tied loosely at the waist, the robe slipping off one shoulder, one hand resting high on the door frame, looking at the camera with a knowing smile. Late morning light, medium-format film look, honey and cream palette, soft grain."
  }
}

Introduction

Qwen Image 2.1 Text to Image turns a written brief into one image per request. Alibaba's Qwen team open-weighted the model on 21 September 2026 and it went live here the same day, hosted, with no weights to download and no GPU to rent.

What sets it apart on this catalogue is not a feature — it is what is missing. There is no content filter, no maturity flag and no prompt scanner in front of it. In our own acceptance testing it rendered uncensored scenes exactly as written, with no substitution and no quiet softening.

Choose a Qwen Image 2.1 mode by starting material

Starting material	Mode
A written brief and nothing else	Qwen Image 2.1 Text to Image
One to ten source images to work from	Qwen Image 2.1 Image Editing
A brief plus your own LoRA weights	Qwen Image 2.1 LoRA

Key features & highlights

Uncensored prompts come back rendered, not softened

The failure worth planning around in this market is not a refusal — it is an endpoint that accepts the request, bills for it, and returns a tamer picture than the one described. We tested for exactly that, walking our content ladder from its mildest level to its strongest, and checking each returned frame element by element against the wording that asked for it. Every element was present — nothing substituted, nothing draped over it, no retreat to a safer composition.

Fifteen ratios, composed rather than cropped

Fifteen aspect ratios span 3:1 and 21:9 strips, 1:3 and 9:21 columns, and the print-shaped 5:4 and 4:5. The composition is built at the ratio you request, so a panoramic frame is genuinely staged for its width instead of arriving as the middle slice of a square.

Three tiers, and only the tier moves the price

1k, 1.5k and 2k. The tier is the single field that changes what a generation costs, which makes the working pattern obvious: explore at 1k, then re-run the frames worth keeping at 2k for skin and fabric detail.

Request parameters

Request parameters for Qwen Image 2.1 · Text to Image
Parameter	Required	Description
`prompt`	Yes	Up to 5,000 characters. There is no negative prompt on this family, so exclusions belong here too — name what is in the frame, then say plainly what must not be. at least 1 characters · at most 5000 characters
`aspect_ratio`	No	Composes the frame before generation rather than cropping afterwards, so pick the shape you will publish at. Fifteen options, from the 3:1 and 21:9 strips to 1:3 and 9:21 columns. `1:1` · `16:9` · `9:16` · `4:3` · `3:4` · `3:2` · `2:3` · `21:9` · `9:21` · `2:1` · `1:2` · `3:1` · `1:3` · `5:4` · `4:5` · Default: `1:1`
`resolution`	No	The tier that sets output size and price: 1k is the default and the cheapest, 1.5k sits in between, 2k is where skin texture, hair and fabric weave hold together. `1k` · `1.5k` · `2k` · Default: `1k`
`enable_prompt_expansion`	No	On by default: before generating, the prompt is rewritten into a fuller scene description. Turn it off to have your own wording used as written. Default: `true`
`background`	No	opaque by default. `opaque` · `transparent` · Default: `opaque`
`output_format`	No	JPEG by default. PNG when the frame goes into a compositing step, WebP when it ships straight to a page. `jpeg` · `png` · `webp` · Default: `jpeg`
`seed`	No	Steadies a draw so a series stays in the same neighbourhood. It is not a reproduction guarantee — we sent the same prompt and seed twice and got two different frames. `≥ 0` · `≤ 2147483647`

Pricing

Pricing
Specification	Price
`1k`	$0.024 / image
`1.5k`	$0.048 / image
`2k`	$0.048 / image

Use cases

Uncensored work with nothing in the way

Figure studies, couples and bold art direction run on this endpoint with no flag to set and no separate route to request. Your wording is passed through as written.

High-volume concept passes

The 1k tier is priced for running a spread rather than a single hero frame. Run the spread, pick the survivors, promote those to 2k.

Wide banners and phone-first verticals

3:1 and 21:9 for site headers and cinematic plates; 9:21 and 1:3 for full-bleed mobile. Staged at the ratio, not cropped to it.

Character starting frames

Generate the character here, then move to image editing and feed that frame back as a reference to hold identity while you change wardrobe, light or setting.

Explore key use-case tutorials

Prompt guide & pro tips

Pick the ratio and the tier before writing

Both decisions change the picture more than any adjective will. Choose the shape you will publish at, and stay at 1k until the composition is settled.

Put exclusions inside the prompt

There is no negative prompt on this family. With 5,000 characters available, state what belongs in the frame and then state what must not appear — duplicated limbs, extra figures, clothing you did not ask for.

Keep scenes to one or two figures at 2k

Anatomy is where this model breaks first. Single figures and pairs hold together; crowded multi-figure compositions at the top tier can duplicate or merge limbs. If a scene needs a crowd, build it by editing rather than asking for it all at once.

Treat the seed as a nudge

The same seed twice does not return the same file. Use it to keep a set coherent, and archive the asset when a frame is final.

API quick start

# Generate the idempotency key once; reuse it when retrying after a lost response.
export SPICY_IDEMPOTENCY_KEY="${SPICY_IDEMPOTENCY_KEY:-$(uuidgen)}"
curl -X POST https://api.spicyapi.ai/api/v1/jobs/createTask \
  -H "Authorization: Bearer $SPICY_API_KEY" \
  -H "Content-Type: application/json" \
  -H "Idempotency-Key: $SPICY_IDEMPOTENCY_KEY" \
  -d '{
  "model": "alibaba/qwen-image-2.1/text-to-image",
  "input": {
    "prompt": "A ceramic teapot on a linen cloth, soft window light, muted palette",
    "background": "opaque",
    "resolution": "1k",
    "aspect_ratio": "1:1",
    "output_format": "jpeg"
  }
}'

# Copy data.taskId from the response into TASK_ID, then poll the status.
TASK_ID="TASK_ID_FROM_RESPONSE"
curl "https://api.spicyapi.ai/api/v1/jobs/recordInfo?taskId=$TASK_ID" \
  -H "Authorization: Bearer $SPICY_API_KEY"

See the API reference for authentication, the complete schema, polling, and error handling.

Related models and next tasks

Missing the model you need?

Tell us what to add and where you would use it. Requests go to the people who decide what gets onboarded next.

Shape and tier are two fields here, not one size string

Qwen Image 2.1 Text to Image turns one written brief into one image per request, and the first field to get right is the one that moved. Qwen-Image 2.0 asked for output size as a single string, size set to something like 1024*1024. Here that decision is two fields: aspect_ratio picks one of fifteen shapes, resolution picks the 1k, 1.5k or 2k tier. The request schema is closed, so a field it does not declare is rejected rather than quietly dropped, and generation is asynchronous — the call returns a handle and the image arrives after polling.

Open the playground

Shape and tier are two fields here, not one size string

Request body: prompt, aspect_ratio, resolution, output_format and seed — prompt is the only required one
Frame shape: aspect_ratio, fifteen values; the frame is composed at that shape rather than cropped to it
Priced axis: resolution alone — 1k by default, then 1.5k and 2k
Unknown fields: Rejected, not ignored: the schema is closed, so a leftover size field fails the call
Content declaration: None: no maturity field in the request and no platform filter in front of the model
Delivery: Asynchronous — one image per request, read from platform storage once the task is terminal

Migration: One decision became two fieldsQwen-Image 2.0 carried output size as a single string. Here the shape and the tier are separate: aspect_ratio takes fifteen values and resolution takes three. Splitting them is what lets you move a frame from square to panoramic without also changing how much detail you paid for.
Quote: Only the tier moves the numberresolution is the one field on this endpoint that changes what a render costs. aspect_ratio, output_format and seed quote the same whatever you set them to, so sweeping a brief across every shape is the cheap kind of experiment and raising the tier is the expensive one. The live estimate above the form is what settles.
Declaration: Nothing in the body labels the workThere is no maturity field to set, no separate uncensored endpoint to request and no filter standing in front of the model. For an integration that means one code path and one request shape: what changes from one render to the next is the sentence in prompt, not the call.

Working with the Qwen Image 2.1 text-to-image API

Porting a body from the 2.0 generation

1024*1024 becomes aspect_ratio 1:1, and the tier becomes a decision you now make out loud. Drop the old string and set nothing in its place and the request still runs — at the 1k default — which is the quiet way a clean-looking port ends up rendering smaller than the code it replaced.

Unknown fields fail loudly, which is the useful behaviour

The schema declares five fields and rejects everything else, so a leftover size, a negative prompt field or a batch count is not silently dropped on the way upstream. The port surfaces as a rejected request at submit time instead of as a picture that is subtly not the one you asked for.

Explore across shape, commit across tier
Because the shape is free and the tier is not, the cheap sweep is one brief rendered at several aspect ratios on one tier, and the expensive step is re-running the survivor higher. Ordering the work that way keeps the part you pay for to the frames that earned it.
One string carries the whole brief
prompt holds up to 5,000 characters and there is no second field for exclusions, so a template that used to write a subject block and a negative block has to concatenate them. State what is in the frame, then state plainly what must not be.
Uncensored work is the same call
Uncensored prompts go to this endpoint with nothing added: no flag to thread through your client, no second route to keep alive, no approval step to build a queue around. The only thing that differs from any other render is the wording.

Who calls the Qwen Image 2.1 text-to-image API

Teams porting from the 2.0 interface
Re-decide the tier you used to imply
The old size string carried both the shape and how big the render was. Splitting it means the second half of that decision now has a default, and defaults are the part of a migration nobody reviews. Set resolution explicitly on every ported call, even when the value you pick is the default one.
Design and marketing systems
Render each placement instead of cropping one
A banner, a square and a phone vertical are three calls with one brief and three aspect_ratio values, quoting the same on one tier. The composition is built for the shape each time, which is the difference between a wide frame and the middle of a square.
Studios producing uncensored work
One integration, not a second one
Because nothing in the request declares maturity, the uncensored half of a catalogue does not need its own client, its own credentials path or its own review queue. The same builder, the same polling loop and the same storage handling cover both.

How to integrate the Qwen Image 2.1 text-to-image API

Split the old size value first
Map the shape to aspect_ratio and the size to a resolution tier. 1024*1024 is 1:1 plus a tier you choose, and writing the tier down is what stops the default from choosing for you.
Strip anything the schema does not declare
Five fields go up: prompt, aspect_ratio, resolution, output_format and seed. A leftover from an older body is not dropped in transit; the request comes back rejected, which is where you want to find it.
Fold exclusions into the prompt
There is no separate field for what must stay out of frame, so the same string carries both halves. Name the subject, the light and the lens, then name what must not appear.
Submit, then poll for a terminal state
The call returns a task handle rather than pixels. Poll until the task is terminal and read the image from the result, which is served from platform storage.
Settle the shape before you raise the tier
Composition is the free variable and detail is the paid one. Get the frame right at the default tier, then re-run only the frames worth the higher one.

Qwen Image 2.1 text-to-image patterns

One brief, three placements
Hold the prompt constant and send it three times with different aspect_ratio values — a 21:9 header, a 1:1 card and a 9:21 phone vertical. Each frame is staged for its shape, so the columns fall differently in the wide one than in the tall one, and all three quote the same on one tier.
```
A woman in a floor-length oxblood coat walking away down an empty marble colonnade at dusk, low sun raking between the columns, long shadows across the stone floor, shot on a 35mm lens, no text anywhere in the frame
```
Exclusions written into the only field there is
With no negative field available, the clause that would have lived there becomes the tail of the prompt. Keep the exclusions concrete — objects, materials and lettering — rather than abstract quality words, which give the model nothing to act on.
```
A close portrait of a woman with wet swept-back hair lit by one low warm lamp against a charcoal background, bare shoulders, sharp catchlights in the eyes, no jewellery, no visible logos, no lettering of any kind
```

Qwen Image 2.1 text-to-image questions

Porting a request body

Two things changed shape and one thing did not. The size string became aspect_ratio plus resolution, and unknown fields became rejections rather than silent drops; the prompt is still one string carrying the entire brief, exclusions included. Set the tier explicitly on every ported call so the default never quietly stands in for a decision, sweep shapes while they are free, and keep a polling loop between the call and the picture.

All Qwen Image models11

Related comparisons

Qwen Image 2.1 vs Qwen Image 3.0 Qwen Image 2.1 vs Seedream 5.0 Pro Qwen Image 2.1 vs Qwen Image Edit Spicy Qwen Image 2.1 vs Seedream 5.0 Flash Qwen Image 2.1 vs Qwen Image 3.0 Pro Qwen Image 2.1 vs Z-Image Spicy Pro

Workbench