Need smarter insights in your inbox? Join our weekly newsletters to get solely what issues to enterprise AI, information, and safety leaders. Subscribe Now
Adobe Photoshop is among the many most recognizable items of software program ever created, utilized by greater than 90% of the world’s inventive professionals, in response to Photutorial.
So the truth that a new open supply AI mannequin — Qwen-Picture Edit, launched yesterday by Chinese language e-commerce big Alibaba’s Qwen Group of AI researchers — is now capable of accomplish an enormous variety of Photoshop-like enhancing jobs with textual content inputs alone, is a notable achievement.
Constructed on the 20-billion-parameter Qwen-Picture basis mannequin launched earlier this month, Qwen-Picture-Edit extends the system’s distinctive strengths in textual content rendering to cowl a large spectrum of enhancing duties, from refined look adjustments to broader semantic transformations.
Merely add a beginning picture — I attempted one in every of myself from VentureBeat’s final annual Rework convention in San Francisco — after which sort directions of what you need to change, and Qwen-Picture-Edit will return a brand new picture with these edits utilized.
AI Scaling Hits Its Limits
Energy caps, rising token prices, and inference delays are reshaping enterprise AI. Be a part of our unique salon to find how prime groups are:
- Turning vitality right into a strategic benefit
- Architecting environment friendly inference for actual throughput beneficial properties
- Unlocking aggressive ROI with sustainable AI techniques
Safe your spot to remain forward: https://bit.ly/4mwGngO
Enter picture instance:
Output picture instance with immediate: “Make the person sporting a tuxedo.”
The mannequin is offered now throughout a number of platforms, together with Qwen Chat, Hugging Face, ModelScope, GitHub, and thru the Alibaba Cloud software programming interface (API), the latter which permits any third-party developer or enterprise to combine this new mannequin into their very own purposes and workflows.
I created my examples above on Qwen Chat, the Qwen Group’s rival to OpenAI’s ChatGPT, nonetheless, it ought to be famous for any aspiring customers that generations are restricted to about 8 free jobs (enter/outputs) per 12 hour interval earlier than it resets. Paying customers can have entry to extra jobs.
With help for each English and Chinese language inputs, and a twin give attention to each semantic which means and visible constancy, Qwen-Picture-Edit goals to decrease boundaries to professional-grade visible content material creation.
And on condition that the mannequin is offered as an open supply code beneath an Apache 2.0 license, it’s protected for enterprises to take, obtain and arrange free of charge on their very own {hardware} or digital clouds/machines, probably leading to an enormous price financial savings from proprietary software program like Photoshop.
As Junyang Lin, a Qwen Group researcher wrote on X, “it could take away a strand of hair, very delicate picture modification.”
The staff’s announcement echoes this sentiment, presenting Qwen-Picture-Edit not as a wholly new system, however as a pure extension of Qwen-Picture that applies its distinctive textual content rendering and dual-encoding strategy on to enhancing duties.
Twin encodings permit for edits preserving type and content material of authentic picture
Qwen-Picture-Edit builds on the inspiration established by Qwen-Picture, which was launched earlier this 12 months as a large-scale mannequin specializing in each picture era and textual content rendering.
Qwen-Picture’s technical report highlighted its skill to deal with advanced duties like paragraph-level textual content rendering, Chinese language and English characters, and multi-line layouts with accuracy.
The report additionally emphasised a dual-encoding mechanism, feeding photos concurrently into Qwen2.5-VL for semantic management and a variational autoencoder (VAE) for reconstructive element. This strategy permits edits that stay trustworthy to each the intent of the immediate and the look of the unique picture.
Those self same architectural selections underpin Qwen-Picture-Edit. By leveraging twin encodings, the mannequin can alter at two ranges: semantic edits that change the which means or construction of a scene, and look edits that introduce or take away components whereas conserving the remainder untouched.
Semantic enhancing consists of creating new mental property, rotating objects 90 or 180 levels to disclose completely different views, or reworking an enter into one other type equivalent to Studio Ghibli-inspired artwork. These edits sometimes modify many pixels however protect the underlying id of objects.
Right here’s an instance of semantic enhancing from Shridhar Athinarayanan, an engineer at AI purposes platform Replicate, who used a Replicate-hosted implementation or “inference” of Qwen to reskin a photograph of Manhattan to appear to be a toy Lego set.
Look enhancing focuses on exact, native adjustments. In these circumstances, many of the picture stays unchanged whereas particular objects are altered. Demonstrations embody including a signboard that generates a mirrored image in water, eradicating stray hair strands from a portrait, and altering the colour of a single letter in a textual content picture.
One good instance of look enhancing with Qwen-Picture Edit comes from AnswerAI co-founder and CEO Thomas Hill who posted a side-by-side on X displaying his spouse in her wedding ceremony costume beneath an archway and one other with the identical archway coated with graffiti:
Mixed with Qwen’s established power in rendering Chinese language and English textual content, the editing-focused system is positioned as a versatile software for creators who want greater than easy generative imagery.
The twin management over semantic scope and look constancy means the identical software can serve very completely different wants, from inventive IP growth to production-level picture retouching.
Including or eradicating textual content to pictures
One other standout functionality is bilingual textual content enhancing. Qwen-Picture-Edit permits customers so as to add, take away, or modify textual content in each Chinese language and English whereas preserving font, dimension, and elegance.
This expands on Qwen-Picture’s fame for robust textual content rendering, significantly in difficult situations like intricate Chinese language characters.
In apply, this permits for correct enhancing of posters, indicators, T-shirts, or calligraphy artworks the place small textual content particulars matter, as seen in one other instance from Replicate beneath.
One demonstration concerned correcting errors in a bit of generated Chinese language calligraphy via a step-by-step chained enhancing course of.
Customers might spotlight incorrect areas, instruct the system to repair them, after which additional refine particulars till the proper characters had been rendered. This iterative strategy exhibits how the mannequin may be utilized to high-stakes enhancing duties the place precision is crucial.
Functions and use circumstances
The Qwen staff has highlighted a variety of potential purposes:
- Inventive design and IP growth, equivalent to producing mascot-based emoji packs.
- Promoting and content material creation, the place logos, signage, and text-heavy visuals may be personalized.
- Digital avatars and artwork, with type switch supporting distinctive character representations.
- Images and private use, together with background changes, clothes adjustments, and object removing.
- Cultural preservation, demonstrated via correcting classical calligraphy works.
By bridging fine-grained enhancing with broader inventive transformations, Qwen-Picture-Edit caters to professionals who want management whereas remaining approachable for informal experimentation.
Benchmarking and efficiency
In response to the Qwen staff, evaluations throughout public benchmarks point out that Qwen-Picture-Edit delivers state-of-the-art efficiency in picture enhancing.
This follows from the broader Qwen-Picture technical evaluations, the place the bottom mannequin achieved main leads to each common picture era and textual content rendering duties.
Whereas particular enhancing benchmark figures weren’t detailed within the launch, Qwen-Picture itself ranked extremely in unbiased evaluations equivalent to AI Area, the place human raters in contrast outputs throughout fashions from completely different suppliers.
API pricing and availability
By means of Alibaba Cloud Mannequin Studio, builders can entry Qwen-Picture-Edit as an API. Pricing is about at $0.045 per picture, with a free quota of 100 photos legitimate for 180 days after activation.
The service is initially accessible within the Singapore area, with a fee restrict of 5 requests per second and as much as two concurrent duties per account.
To make use of the API, builders should acquire a Mannequin Studio API key and may name the mannequin by way of HTTP or via the DashScope SDK in Python or Java.
Photographs may be submitted as URLs or in Base64 format, with supported resolutions starting from 512 to 4,096 pixels and file sizes as much as 10 MB. Output photos are hosted on Alibaba Cloud Object Storage with hyperlinks legitimate for twenty-four hours, requiring customers to obtain and save outcomes promptly.
What’s subsequent for Qwen?
Qwen positions Picture-Edit as a step toward decreasing boundaries for visible content material creation. By making exact, style-consistent enhancing extra accessible, the mannequin might help purposes from design studios to informal customers refining private tasks.
The system additionally indicators a broader development in AI growth: shifting past single-purpose era towards instruments that combine enhancing, correction, and refinement.
With each semantic flexibility and appearance-level precision, Qwen-Picture-Edit displays this shift, mixing the generative strengths of enormous fashions with the reliability required for skilled enhancing.
