Qwen Image Explained: Every Variant, the Edit Task Family, and Official Pricing (August 2026)
Last updated August 7, 2026
Qwen Image is Alibaba's image family: a 20B Apache 2.0 open flagship (2512), the Edit line through 2511, the layer-decomposing Qwen-Image-Layered, and the API-only 2.0/2.0 Pro ($0.075 per image). Its task-specific edit variants are LoRA adapters on the open Edit Plus base, later merged upstream into 2511. Text rendering, especially Chinese, is its signature. The full roster runs in the invideo agent's Image tab.
Updated August 2026
Qwen Image is Alibaba's image generation and editing family — the most structurally interesting one in the open-weight world. As of August 2026 it spans a 20B Apache 2.0 flagship (Qwen-Image-2512), an instruction-editing line (through Edit-2511), a one-of-a-kind layer-decomposition model (Qwen-Image-Layered), an API-only next generation (Qwen-Image-2.0 and 2.0 Pro), and a family of task-specific edit variants powered by mostly community-trained LoRA adapters. Its founding differentiator is complex text rendering, especially Chinese, where no other family is close. The entire roster, task variants included, runs inside the invideo agent's Image tab.
How did the Qwen image family evolve?
Dates below come from the official QwenLM/Qwen-Image repository and Alibaba's model cards:
| Date | Release | What it added |
|---|---|---|
| Aug 4, 2025 | Qwen-Image | 20B MMDiT foundation model, open weights, text-rendering focus (arXiv:2508.02324) |
| Aug 18, 2025 | Qwen-Image-Edit | Instruction-based editing built on the base model |
| Sep 22, 2025 | Qwen-Image-Edit-2509 ("Edit Plus") | Multi-image input (1–3 images), identity/product consistency, native ControlNet |
| Dec 2025 | Qwen-Image-Layered | Decomposes images into editable RGBA layers |
| Dec 23, 2025 | Qwen-Image-Edit-2511 | Two-person fusion, geometric reasoning; popular community LoRAs merged into the base |
| Dec 31, 2025 | Qwen-Image-2512 | Refreshed flagship: more realistic humans, finer textures, better text |
| Feb 10, 2026 | Qwen-Image-2.0 / 2.0 Pro | Unified generation+editing, ~7B architecture, native 2K, ~1,000-token prompts — API only (TMTPost) |
Two lines move in parallel — edit iterations (2509, 2511) and base refreshes (2512) — with 2.0 breaking the pattern as a smaller, closed, unified model.
Which Qwen image models are open source?
The split is clean; coverage often blurs it. Everything through December 2025 is open: Qwen-Image, Edit, Edit-2509, Edit-2511, 2512, and Layered all ship Apache 2.0 weights on Hugging Face and ModelScope — commercial use, fine-tuning, and self-hosting included. Qwen-Image-2.0 and 2.0 Pro are not: as of August 2026 they are available only through Alibaba Cloud Model Studio's API and Qwen Chat, with no open-source plan announced. That makes 2512 (generation) and Edit-2511 (editing) the open-weight flagships, while 2.0 Pro is the hosted best — the family's first real open/closed capability gap. (An unannounced qwen-image-3.0-pro identifier surfaced in Model Studio's catalog in late July 2026; Alibaba has published nothing about it, so treat that as a sighting, not a release.)
Which Qwen image variant should you pick?
invideo's model picker carries the core family and the task-specific edit variants. The core entries:
| Picker entry | What it is | Best for |
|---|---|---|
| Qwen Image 2.0 Pro | API-only unified gen+edit flagship (Feb 2026); native 2K, ~1,000-token prompts | Posters, infographics, dense text, top realism |
| Qwen Image 2.0 | Same ~7B architecture, tuned for speed and cost | Everyday 2K generation at half Pro's price |
| Qwen Image Layered | Layer-decomposition model (Dec 2025), Apache 2.0 | Outputs you'll edit layer-by-layer afterwards |
| Qwen Image Edit | Original edit model (Aug 2025) | Single-image semantic and in-image text edits |
| Qwen Image Edit Plus | Edit-2509: 1–3 input images, ControlNet, consistency upgrades | Multi-image, identity-preserving edits — the workhorse |
| Wan 2.7 Image | Alibaba's Wan family in image mode — a sibling, not a Qwen-Image checkpoint | An alternative Alibaba aesthetic for stills |
The short version: 2.0 Pro for text-dense finals, Edit Plus for anything that starts from an existing image.
And the task family — fixed-function edit variants sharing one base:
| Picker entry | What it does |
|---|---|
| Qwen Edit Multiple Angles | Re-shoots the image from a new camera position (rotation, zoom, height) |
| Qwen 2511 Multiple Angles | Newer 96-pose camera control on the Edit-2511 base (details below) |
| Qwen Edit Remove Element | Removes objects, people, or text while keeping the rest consistent |
| Qwen Edit Add Background | Puts a realistic scene behind a white-background product shot |
| Qwen Edit Integrate Product | Blends a product into a background with perspective and lighting correction |
| Qwen Edit Shirt Design | Applies a graphic onto a person's shirt for apparel mockups |
| Qwen Edit Face To Portrait | Generates a full portrait from a cropped face photo |
| Qwen Edit Group Photo | Merges people from separate photos into one group shot |
| Qwen Edit Next Scene | Generates a cinematic next shot — camera moves, framing changes |
| Qwen Edit Remove Lighting | Strips existing lighting to soft, even illumination for compositing |
| Qwen Edit Relight / Qwen Edit Skin | Community-LoRA members of the same pattern: light migration and skin-texture realism |
Eleven entries, one shared base — what they actually are is the next section.
What is the Qwen Edit task family, really?
This is the part almost every article gets wrong. The task variants are not separate Alibaba models. Each one pairs the open Qwen-Image-Edit-2509 weights with a LoRA adapter — mostly community-trained — plus a fixed trigger-prompt template, productized as a fixed-function endpoint (a pattern first packaged commercially in November 2025). The clearest example: the multiple-angles behavior traces to community author dx8152's Qwen-Edit-2509-Multiple-angles LoRA, at about 205,000 Hugging Face downloads; "Relight" traces to the same author's light-migration LoRAs, and "Skin" to community skin-realism adapters — members of the same pattern rather than catalogued official endpoints.
Then comes the step nobody covers: Alibaba closed the loop. The Edit-2511 model card states that popular community LoRAs — lighting enhancement and viewpoint generation — were merged directly into the 2511 base, so those behaviors now work without adapters. Community fine-tune → hosted product → absorbed into the next official release: an open-weight flywheel that closed families structurally cannot run, and the best reason to care about Qwen beyond any single benchmark.
What does Qwen-Image-Layered do?
Qwen-Image-Layered (December 2025, Apache 2.0) decomposes a single image into semantically separate RGBA layers — subject, background, shadows, text — each of which can be moved, resized, recolored, or deleted without regenerating anything else. Layer count is a parameter, and decomposition is recursive: any output layer can be decomposed again (arXiv:2512.15603). The closest mental model is a generative PSD file — and unlike segmentation or matting, it plausibly fills in the content behind whatever it separates out. One honest limit: the current version recommends 640px output (1024 works with trade-offs), so as of August 2026 it's a structure tool, not a final-resolution renderer.
How does the 96-pose camera control work?
Qwen 2511 Multiple Angles is the most concrete camera control shipped on any open image model. The official LoRA (Apache 2.0, trained on 3,000+ Gaussian-Splatting renders for 3D-consistent viewpoints) exposes a controlled vocabulary rather than free text: 8 azimuths (45° steps, 0°–315°) × 4 elevations (−30°, 0°, 30°, 60°) × 3 distances (0.6× close-up, 1.0× medium, 1.8× wide) = 96 discrete camera poses, invoked as <sks> [azimuth] [elevation] [distance] — for example <sks> front view eye-level shot medium shot. It's a grid, not a continuous dial, which is exactly why it's reliable: every pose was trained explicitly.
Can Qwen build a product photography pipeline?
Yes — the task family chains into a genuine e-commerce workflow, which is clearly what much of it was built for:
- Add Background turns a white-background SKU shot into a lifestyle scene.
- Integrate Product composites the product into any setting, perspective and lighting corrected automatically.
- Shirt Design generates apparel mockups from a flat graphic.
- Multiple Angles produces a turntable of views from one photo — no reshoot.
Add Remove Lighting → Relight for catalog-consistent illumination across a product line. Inside invideo this maps onto the product photography workflow, with the same models behind general AI image generation and cleanup in the photo editor.
How do you prompt Qwen image models?
From the official model cards:
- Text-to-image: append the "positive magic" suffix — English
", Ultra HD, 4K, cinematic composition.", Chinese", 超清，4K，电影级构图". For in-image text, quote the exact string. - Edit models: recommended settings are
true_cfg_scale=4.0, 40 inference steps,guidance_scale=1.0, and a single-space negative prompt. - Multi-image edits: describe the spatial layout explicitly — "the magician bear is on the left, the alchemist bear is on the right" — so the model knows which input goes where.
- Multiple Angles: use the exact
<sks>vocabulary above; paraphrasing the pose names breaks the control.
How much does Qwen Image cost?
Official Alibaba Cloud Model Studio pricing (international region, per image, August 2026), each model with a 100-image free quota:
| Model | Price per image |
|---|---|
| qwen-image-2.0-pro | $0.075 |
| qwen-image-2.0 | $0.035 |
| qwen-image | $0.035 |
| qwen-image-edit | $0.045 |
| qwen-image-edit-plus | $0.03 |
| qwen-image-edit-max | $0.075 |
The line worth a second look: edit-plus at $0.03 is the cheapest entry on the list — the newer, more capable editor undercuts the original edit model at $0.045. Third-party hosts quote conflicting rates; official numbers are the only ones worth repeating. Self-hosting the open checkpoints carries no license fees.
What are Qwen Image's real strengths and weaknesses?
Strengths, documented: best-in-class text rendering — the family was built around it, and its Chinese text performance remains unmatched among mainstream models. Editing precision with identity preservation (person, product, in-image text) is the 2509 card's headline capability. And Apache 2.0 weights at near-frontier quality created the largest editing-LoRA ecosystem anywhere.
Weaknesses, equally documented: the open models are heavy — 20B parameters, roughly 40GB+ in BF16, so consumer GPUs depend on quantized versions. Photorealistic skin drew persistent "waxy" criticism, implicitly conceded when 2512's headline fix was "more realistic human generation, finer textures." Layered tops out at its 640px recommendation. And the ranking claims lean on Alibaba's own AI Arena: the GitHub README calls 2512 the "strongest open-source image model" across 10,000+ blind evaluations — a vendor-run result, worth knowing and worth flagging as such.
Qwen Image: quick answers
Is Qwen Image open source?
The pre-2026 family is: Qwen-Image, 2512, all Edit versions through 2511, and Layered are Apache 2.0 with downloadable weights. Qwen-Image-2.0 and 2.0 Pro are API-only with no released weights as of August 2026.
Can I use Qwen Image commercially?
Yes for every open-weight model — Apache 2.0 permits commercial use, modification, and self-hosting. Qwen-Image-2.0/Pro outputs are commercial via the paid API.
Is Qwen Image Edit Plus a different model from Qwen Image Edit?
Edit Plus is the hosted alias of Qwen-Image-Edit-2509 — a later, more capable iteration of the same line, adding 1–3 input images and consistency upgrades. Use Edit Plus by default.
Which Qwen edit variant should I use for which job?
Object cleanup → Remove Element. Product staging → Add Background, then Integrate Product. Apparel mockups → Shirt Design. New viewpoints → 2511 Multiple Angles. People → Face To Portrait or Group Photo. Storyboarding → Next Scene. Anything else → general Edit Plus with a written instruction.
What's the difference between Qwen-Image-Layered and background removal?
Background removal masks one foreground and discards the rest. Layered decomposes the whole image into editable RGBA layers — subject, background, shadow, text — and generates the hidden content behind each, recursively if you want. You keep everything, editable.
How much does Qwen Image 2.0 Pro cost?
$0.075 per image on Alibaba Cloud Model Studio (about 2.1× the $0.035 base 2.0), each with a 100-image free quota, as of August 2026.
What is the <sks> trigger in Multiple Angles?
The control token that activates the camera LoRA, followed by one each of 8 azimuth, 4 elevation, and 3 distance terms — 96 valid combinations. A fixed vocabulary, not natural language.
Where can you run the Qwen image family?
Every model on this page — the 2.0 generation line, Edit and Edit Plus, Layered, and the full task family from Skin to Group Photo — is live in the Image tab of the invideo agent, the AI video platform that gives serious creatives every major model in one place. That matters here in particular: the task variants are scattered across hosts elsewhere, but in the agent they sit in one picker beside 200+ other models, and a product still flows straight into a video edit. Start at invideo.io, see the Qwen model hub, or begin from product photography or the AI image generator.
Version history
- August 2026 — First published. Covers Qwen-Image through 2512, Edit through 2511, Layered, the API-only 2.0/2.0 Pro, the eleven task-specific edit variants and their LoRA lineage, 96-pose camera control, and official Alibaba pricing.