Qwen Image 2.1 vs Qwen Image 3.0: What Changes and When to Upgrade

Qwen Image 2.1 earned its place in a lot of image pipelines: solid text-to-image quality, multi-reference editing, and custom LoRA support. Qwen Image 3.0 is the next generation, and it changes the shape of the family. Instead of one model with many knobs, 3.0 ships as two quality tiers, Standard and Pro, each with a text-to-image and an edit endpoint, plus built-in prompt expansion.

This guide compares the two generations parameter by parameter and explains which workloads should move to Qwen Image 3.0 now. All four Qwen Image 3.0 endpoints are available together in the Qwen Image 3.0 collection.

At a Glance

	Qwen Image 2.1	Qwen Image 3.0
Lineup	Text-to-image, edit, and LoRA variants	Standard and Pro tiers, each with text-to-image and edit
Resolution tiers	`1k`, `1.5k`, `2k`	`1k`, `2k`
Aspect ratios	15 presets, `1:3` to `3:1`, including `9:21` and `21:9`	The same 15 presets
Edit references	1 to 10 images	1 to 3 images (384 to 2048 px per side)
Prompt expansion	Not available	Optional `enable_prompt_expansion` (off by default)
Prompt language	Natural-language prompts	Chinese and English, up to 800 characters
Custom LoRAs	Up to 3 LoRAs on LoRA variants	Not available
Output format	`jpeg`, `png`, `webp`	Standard image output
Seed control	Yes	Yes, `-1` for random or `0` to `2147483647`

What Qwen Image 3.0 Adds

Two quality tiers instead of one

The biggest change is the Standard/Pro split. Both tiers share the same request shape, so switching between them is a one-line change in your integration:

Qwen Image 3.0 Text-to-Image and Qwen Image 3.0 Edit are the everyday tier for high-volume generation, drafts, and production assets where throughput and cost matter most.
Qwen Image 3.0 Pro Text-to-Image and Qwen Image 3.0 Pro Edit target polished final output, with stronger detail, texture, and visual fidelity for portraits, fashion, product hero shots, and campaign visuals.

A common pattern is to iterate on Standard and render the approved concept on Pro with the same prompt, aspect ratio, and seed.

Built-in prompt expansion

Qwen Image 3.0 can enrich a short prompt before generation. Set enable_prompt_expansion to true when you want the model to fill in lighting, composition, and detail from a brief idea, and leave it off when you need the output to follow your wording exactly. Qwen Image 2.1 had no equivalent, so teams usually wrote long prompts by hand or ran a separate rewriting step.

Stronger prompt understanding in two languages

Qwen Image 3.0 accepts prompts in both Chinese and English, up to 800 characters, with strong adherence to subjects, composition, style, lighting, and attributes. That makes it a natural fit for bilingual products and for prompts that mix languages.

Simpler resolution choices

Qwen Image 3.0 offers 1k and 2k tiers across all 15 aspect-ratio presets. The intermediate 1.5k tier from 2.1 is gone. In practice, most teams used 1k for previews and 2k for finals anyway. On Standard, choosing 2k does not change the price; Pro is priced by resolution tier, and edit requests on both tiers also scale with the number of reference images. Current rates are listed on each model page.

What Qwen Image 2.1 Still Does Differently

An honest comparison also covers what 3.0 does not carry over:

Many-reference editing. Qwen Image 2.1 Edit accepts up to 10 reference images; Qwen Image 3.0 Edit accepts up to 3. For most edits (a subject, a style reference, and a background), three references are enough. Workflows that composite many separate inputs in a single call will need to restructure into several passes.
Custom LoRAs. Qwen Image 2.1 has LoRA variants that apply up to 3 custom LoRAs. Qwen Image 3.0 does not take LoRAs, so identity or style locking relies on reference images in the edit endpoints instead.
Output format selection. Qwen Image 2.1 lets you pick jpeg, png, or webp. Qwen Image 3.0 returns a standard image output, so convert downstream if your pipeline requires a specific format.

Which One Should You Use?

Your workload	Recommendation
High-volume text-to-image for apps, social content, and drafts	Qwen Image 3.0 Standard
Final marketing, product, fashion, or portrait renders	Qwen Image 3.0 Pro
Short or casual user prompts in consumer apps	Qwen Image 3.0 with `enable_prompt_expansion` on
Strict prompt control for templated generation	Qwen Image 3.0 with prompt expansion off and a fixed seed
Instruction-based edits with 1 to 3 references	Qwen Image 3.0 Edit or Pro Edit
Bilingual (Chinese and English) prompting	Qwen Image 3.0
Pipelines built around custom LoRAs or 4+ references per edit	Keep these on Qwen Image 2.1 for now, and test whether 3.0 reference-based editing covers the use case

For most new projects, Qwen Image 3.0 is the better default: two clear quality tiers, better prompt handling, and the same aspect-ratio coverage, with no new parameters to learn beyond prompt expansion.

Migrating from 2.1 to 3.0

Moving an existing integration takes a few small changes:

Swap the model path to the Qwen Image 3.0 or 3.0 Pro endpoint of the same task (text-to-image or edit).
Map resolution tiers. Keep 1k and 2k. Map 1.5k to 2k when quality matters or to 1k when cost matters.
Trim edit references to at most 3 images, each between 384 and 2048 pixels per side.
Drop output_format and convert downstream if needed.
Keep prompts under 800 characters, and decide per use case whether to enable prompt expansion.
Re-check seeds. Seeds are not portable across model generations, so re-pick seeds for any looks you want to reproduce.

aspect_ratio and seed keep the same names and presets, so the rest of a typical request stays as it is.

Try Qwen Image 3.0 on WaveSpeedAI

Every Qwen Image 3.0 endpoint runs on WaveSpeedAI with a ready-to-use REST API, no cold starts, and pay-per-image pricing. Compare Standard and Pro side by side, try prompt expansion on your own prompts, and pick the tier that fits each workload.

Explore the Qwen Image 3.0 collection on WaveSpeedAI