Image model comparison
Qwen-Image-2.1 vs Nano Banana 2
Qwen-Image-2.1 (released September 20, 2026) is Alibaba's 7B open-weight model for generation, editing, and native transparency; the 2.1-Turbo checkpoint (October 9, 2026) generates and edits in 8 steps. Nano Banana 2 is Google's model, hosted in Imgo with editing, assets, and video handoff around it. This page compares what each path means for teams shipping commercial image work.
Reviewed October 10, 2026, including the Qwen-Image-2.1-Turbo model card and Alibaba Cloud's Model Studio pricing. License terms, limits, and availability change; the license text on the Qwen repository is the authority on what Qwen-Image-2.1 permits, and Imgo's live credit quote is the authority on price.
Run this comparison yourself
Same prompt, same aspect ratio — only the model changes. Run it on each one and judge the difference on your own screen.
Test prompt · Aspect ratio 1:1
A chalkboard menu on a café wall, hand-lettered in white chalk. Title: "TODAY". Three lines: "OAT LATTE 4.50", "MATCHA TONIC 5.00", "BANANA BREAD 3.20". A small chalk drawing of a coffee cup in the corner. Warm morning light, photographed straight on.
- Run on Nano Banana 2
Qwen Image 2.1 isn't on Imgo — copy the prompt to test it there.
Choose Qwen-Image-2.1 when…
You can host the weights yourself, your use fits the Qwen Research License, and you want one 7B model that edits and generates with native RGBA transparency.
Choose Nano Banana 2 in Imgo when…
The work is commercial today: pick a hosted model, guide it with product references, edit results, and hand them into video — with no GPU stack to run.
What to compare
| Decision factor | Nano Banana 2 in Imgo | Qwen-Image-2.1 (self-hosted) |
|---|---|---|
| Where it runs | Hosted in Imgo's image generator with editor, asset library, and video tools one click away. | You download the weights and serve them on your own GPUs or a partner cloud. |
| Commercial use | Generation is a commercial service; outputs ship under the plan's terms. | The Qwen Research License governs the weights — read it before commercial deployment. |
| Editing | Prompt editing, masked inpainting, outpainting, and text replacement in the dedicated editor. | Editing ships inside the same 7B weights, including native RGBA transparency output. |
| Consistency | Reuse references and asset context across image and video tasks, then inspect identity and labels. | Test identity hold with your own references across edits; inspect every accepted frame. |
| Operations | No inference to run; models, updates, and queues are part of the platform. | Serving, scaling, and model updates are yours to operate and pay for. |
| Cost shape | Per-generation credits quoted live before you submit. | Free weights, real GPU hours — cost scales with your own serving stack. |
New: the Qwen-Image-2.1-Turbo checkpoint
On October 9, 2026 the Qwen team published Qwen-Image-2.1-Turbo: an accelerated checkpoint of the same 7B model that generates and edits in 8 denoising steps instead of the 40 the base card's examples use. The 2K presets, native RGBA transparency and the editing modes carry over unchanged — and so does the license. Here is how the three options line up.
| Capability | Qwen-Image-2.1-Turbo | Qwen-Image-2.1 (base) | Nano Banana 2 in Imgo |
|---|---|---|---|
| Transparent (RGBA) output | Inherited from 2.1 — the Turbo model card shows a native transparent-output example alongside generation and editing. | Native: generate regular or transparent (RGBA) images from text, edit transparent layers, extract subjects. | Not native to generation. Generate the full scene, then cut the subject out with the background remover as a second step. |
| Resolution | The same 2K presets as 2.1 — 2048×2048 for 1:1 up to 2752×1536 for 16:9. | 2K presets, 2048×2048 for 1:1 up to 2752×1536 for 16:9. | Nano Banana 2 outputs up to 4K; in Imgo's generator it prices at the 2K tier. |
| Steps and speed | 8 denoising steps, with the recommended schedule saved in the checkpoint, CFG 1 by default, and prefix KV caching that reuses the text and reference-image context. | The model card's examples run 40 denoising steps. | Hosted — the step count is Google's serving problem, not a dial you see. What you see is the credit quote. |
| Hosted API or local | Weights on Hugging Face and ModelScope, or a hosted API: qwen-image-2.1-turbo on Alibaba Cloud Model Studio at $0.016 per image (international list; $0.014133 global). | Weights only — the base 2.1 has no hosted tier of its own; Model Studio's API serves the pro and turbo variants. | Hosted in Imgo — no GPUs to run, and the credit price is quoted before you submit. |
| Local VRAM | Not published on either model card; budget for a large GPU and plan on CPU offload. | Not published; the card's memory-optimization snippet suggests enable_model_cpu_offload() for smaller GPUs. | Not applicable — hosted. |
| License | The same Qwen Research License as the base: non-commercial use only; commercial use needs a separate license from Qwen. | Qwen Research License — non-commercial only. | A commercial service; outputs ship under Imgo's plan terms. |
Turbo's speed does not change its terms. The checkpoint ships under the same Qwen Research License as the base weights, so the commercial-use question below applies to it identically — 8 steps makes iteration cheaper, not commercial.
The license decides who can ship
The Qwen-Image-2.1 weights are public, but they ship under the Qwen Research License, which grants rights for non-commercial purposes only; commercial deployment needs a separate commercial license from Qwen. Nano Banana 2 in Imgo is a hosted commercial service — the images you generate are yours to ship under Imgo's plan terms. If your output ends up in ads, store listings, or client work, that difference comes before any benchmark score.
Commercial deployment: which Qwen-Image weights allow it
Qwen-Image is not one license. The weights the Qwen team has published on Hugging Face fall into two groups (checked 2026-10-07):
Apache-2.0 — commercial use permitted
Qwen-Image (20B, August 2025), Qwen-Image-Edit, Qwen-Image-Edit-2509, Qwen-Image-Edit-2511, Qwen-Image-2512 and Qwen-Image-Layered. Apache-2.0 allows commercial use, modification and redistribution, as long as you keep the license and its notices.
Qwen Research License — non-commercial only
Qwen-Image-2.1, the Qwen-Image-2.1-Turbo accelerated checkpoint (October 2026), and the prompt-enhancer companions (Qwen-Image-2.1-PE-T2I and Qwen-Image-2.1-PE-I2I). The license grants rights for non-commercial purposes only; commercial use needs a separate commercial license from Qwen, which the license says to request at [email protected].
Check before you deploy
- The exact repository and revision you download: the LICENSE file there binds you, not a blog post or a model listing.
- Where the outputs go: images in ads, store listings or client work are commercial use, even when the model runs on an internal server.
- Attribution and naming terms: the Research License requires "Built with Qwen" or "Improved using Qwen" when outputs are used to build or improve another AI model, and restricts "Qwen" in product names.
- Everything stacked on the weights: LoRAs, ControlNets, fine-tunes and workflow nodes carry their own licenses.
- Hosted access: Qwen models called through Alibaba Cloud or Qwen Cloud follow that service's terms, not the open-weight license.
This is a summary for orientation, not legal advice. Read the license text and talk to your counsel before a commercial deployment.
A repeatable test
Use the same owned product photo, the same three-angle brief, and the same acceptance rules on both paths. Record label and logo accuracy, transparency edge quality, edit round-trip reliability, and the true cost per accepted image — GPU time included for the self-hosted path. Keep every output so the comparison stays inspectable; this page does not claim an unrun benchmark winner.
Prompt
Shoot the same matte-black ceramic pour-over kit on a white studio set from three angles, keeping the wooden handle grain and the small 'AROMA 02' label exact in every frame.
Input and controls
Attach one owned packshot of the kit, use the same three angles and acceptance rules in both products, and compare label accuracy, handle material, and shadow realism.
Record these outcomes
- Label and logo accuracy
- Native transparency edges
- Edit round-trip reliability
- Setup time and serving cost