This submit was co-written with Bradley Grantham and Hugo Dugdale from Popsa.
Popsa is a expertise firm that helps customers rediscover and relive the significant recollections hidden of their picture libraries. Obtainable throughout greater than 50 nations and 12 languages, we use design automation and AI to rework on a regular basis images into private, shareable experiences, together with fantastically printed Picture Books.
In 2016, we launched PrintAI, a pioneering algorithm to take full management of making a different and attention-grabbing design from a consumer’s images. Our clients might use the algorithm to create Picture Books that appeared professionally designed, in lower than 5 minutes.
A core philosophy of our enterprise is that expertise ought to do the heavy lifting for our customers, so automation has all the time been an intrinsic a part of our product. Within the present Generative AI age, we are able to develop much more methods to raise our clients’ expertise, with out making our software program extra sophisticated to make use of.
On this submit, we share how we utilized Amazon Bedrock and the Amazon Nova household of fashions to reimagine our Title Suggestion characteristic. By combining metadata, pc imaginative and prescient, and retrieval-augmented generative AI, we now mechanically generate inventive, brand-aligned titles and subtitles throughout 12 languages. Utilizing the unified API of Amazon Bedrock, Anthropic’s Claude 3 Haiku, and Amazon Nova Lite and Professional, we improved high quality, diminished value, and lower response occasions. This resulted in larger buyer satisfaction, measurable uplifts in engagement and buy charges, and over 5.5 million personalised titles generated in 2025.
Producing title options with Amazon Bedrock
When a buyer receives their Picture Guide, the very first thing they see is the entrance cowl, with a outstanding title and subtitle. A high-quality title and subtitle elevate a Picture Guide’s design, nevertheless most clients aren’t skilled copywriters and lots of of them settle for easy titles like “France 2024”, “Images from Spain” and even, “Images”.
To assist customers elevate their images, we developed and launched a characteristic known as Title Suggestion, which has been accessible to our customers since 2021.
When customers choose images for a Picture Guide design, our cell app reads metadata—comparable to timestamps and geocoordinates—from the photographs and runs on-device convolutional neural networks to extract related options. For instance, whether or not the picture incorporates a seaside, a barbecue, or a pet.
To make use of this information, we created an algorithm known as Title Suggestion Graph. This algorithm used the metadata and information of the chosen images to construct a listing of doable titles, following a algorithm and templates to reach at a set of appropriate options. For instance:
If all images within the design have been taken on the identical day
then recommend “On this Day” as a title with a subtitle of the precise date
In June 2024, we recognized a chance to enhance Title Suggestion by making use of generative AI, with the purpose of inspiring our customers with extra inventive titles. We started by clearly defining the issue and establishing analysis metrics.Our resolution needed to meet strict necessities:
- Character restrict
- Each the title and subtitle should not exceed 36 characters on account of format restrictions affecting how the textual content can be displayed on a entrance cowl.
- Title class
- Every title–subtitle pair should even have an related class that determines the icon displayed alongside the pair to customers. Imagined or incorrect classes would stop an icon from being rendered.
- JSON format
- Lastly, all outputs have to be legitimate JSON with keys `title`, `subtitle` and `class`. This helped with constant parsing, validation, and rendering within the app.
These guidelines helped with analysis as a result of they may very well be outlined in code, so we constructed a dataset of over 100 instance Picture Books and outlined our metrics in an analysis pipeline:
- % of title/subtitle options throughout the character restrict
- % of legitimate title classes
- % of responses within the appropriate JSON format
Along with these strict guidelines, we wanted our resolution to fulfill some broader tips:
- Theme consistency
- Classes ought to match the content material (for instance, snowboarding icons wouldn’t be applicable if the design topic was a seaside vacation)
- Model model
- Ideas ought to mirror Popsa’s tone and model id
- Title-subtitle cohesion
- Pairs ought to complement one another; they shouldn’t be repetitive or disjointed.
- Multilingual high quality
- Ideas wanted to be prime quality in all 12 languages we assist.
We determined to make use of an LLM-as-a-judge to judge efficiency towards these tips. This helped us to quickly take a look at totally different fashions, prompts, and strategies to establish essentially the most dependable strategy. After narrowing to 2 or three choices, we performed in depth inner testing.
Our high outcomes got here from Retrieval-based few-shot prompting. We created a database of instance Picture Books and acceptable title options. For a brand new Picture Guide, we retrieved just a few comparable Picture Guide designs and a random collection of their recommended titles.
Utilizing Amazon Bedrock and Anthropic’s Claude 3 Haiku, we seeded the dialog with these examples as – messages earlier than appending the consumer’s new design doc as the ultimate message. This allowed the massive language mannequin (LLM) to emulate prior responses whereas naturally following the foundations that we outlined.
Our full structure for this resolution may be seen within the following diagram:
When our Title Suggestion Service receives a request, it first decrypts and processes the consumer’s design to extract the timestamps. Then, it performs a reverse geocoding operation on any latitudes and longitudes included within the design, after which classifies the topic of the design based mostly on object landmarks.
This generates an outline like “A snowboarding photobook with 21 images taken within the Alps between twenty first January 2025 and twenty third January 2025”. We then cross this description to our retrieval-based few-shot prompting part to supply a closing set of user-facing options.
Comparisons to our earlier graph-based methodology present higher outcomes:
To quantify enhancements, we relied on a suggestions loop, the place clients rated options as optimistic, impartial, or adverse. We additionally performed multivariate testing with a whole lot of hundreds of customers. Suggestions strongly favored generative AI titles, and key metrics like Design Created and Buy additionally improved. After a number of months, we rolled the characteristic out to 100% of our customers.
By shifting from the Graph Algorithm to Claude 3 Haiku for producing title options, we elevated optimistic consumer suggestions by 13% (from 58% to 71%).
Enhancing buyer satisfaction and decreasing value with Amazon Nova
For the reason that generative AI based mostly re-launch of Title Ideas in 2024, LLM expertise has improved considerably in efficiency, value, and velocity. The unified API of Amazon Bedrock has helped us to check and take a look at new fashions by flipping mannequin IDs and delivery experiments in hours as a substitute of weeks. We lately examined the Amazon Nova household (Micro, Lite, and Professional) which assist greater than 200 languages at low latency.
In early 2025, we ran a multivariate A/B take a look at evaluating Claude 3 Haiku and Nova fashions, monitoring guardrail metrics and gathering direct consumer preferences by means of our in-app suggestions characteristic.
Testing varied fashions for title technology confirmed that whereas Claude 3 Haiku (71% optimistic) carried out nicely, Nova Professional achieved the best consumer satisfaction at 73% optimistic suggestions with the bottom adverse suggestions at 12%.
Whereas Nova Micro-outperformed our legacy Graph methodology, it lagged in consumer satisfaction in comparison with the opposite LLMs and have been put aside. Among the many remaining fashions, we targeted not solely on high quality, but in addition on value, latency and throughput, as proven within the following desk. These comparisons made it clear that Nova Lite supplied near-identical high quality to Claude Haiku at decrease value and sooner response occasions.
Mannequin
Worth per 1,000 enter tokens
Worth per 1,000 output tokens
Response Time (Seconds To Output 500 Tokens)
Claude 3 Haiku
$0.00025
$0.00125
6.8
Amazon Nova Lite
$0.000069
$0.000276
2.4
Amazon Nova Professional
$0.00092
$0.00368
3.4
*pricing taken from the Amazon Bedrock pricing web page
*efficiency metrics taken from Synthetic Evaluation
Lowering Time to First Suggestion with the ConverseStream API
One of many key latency metrics that we monitor is Time to First Suggestion (TTFS), which measures how rapidly the primary legitimate suggestion seems after a consumer request. Even when extra choices are being generated within the background, reducing TTFS makes the characteristic really feel extra responsive, so options are seen earlier than the consumer strikes on.
To enhance our TTFS, we migrated from the InvokeModel API of Amazon Bedrock to the ConverseStream API, to stream tokens as they’re generated.As a result of our providers require legitimate title-subtitle-category triplets, we prolonged the FastAPI to parse streams in actual time, returning the primary suggestion instantly upon validation. Extra options proceed streaming within the background, however the shopper already has one thing able to show.
This shift dramatically diminished TTFS to below one second for the primary polished suggestion, as a substitute of ready for a complete batch of options to finish.
By migrating to the ConverseStream API, we diminished the typical time to first suggestion from 1.41 seconds to 0.92 seconds, delivering title options 35% sooner to customers.
What’s subsequent
In 2025, our Title Suggestion characteristic has generated over 5.5 million titles, offering insights into what resonates, what doesn’t, and the way individuals work together with our options. That suggestions loop will proceed to drive evolution of the characteristic.
Trying forward, we plan to make use of bigger fashions like Nova Professional for a portion of our consumer base, to seize creativity and nuance whereas nonetheless working cost-effectively at scale. The information that we collect from these experiments will assist us to fine-tune smaller fashions, serving to them inherit the strengths of their bigger counterparts with out compromising latency or affordability.
Future work contains instrument integrations that give the LLM richer context about every Picture Guide, from occasion particulars to seasonal cues, with the purpose of producing extra personalised, thematic, and brand-aligned titles.
These developments proceed our mission: enabling anybody, regardless of their talent degree, to rapidly flip their images into significant, inventive, and treasured keepsakes.
Concerning the authors
Bradley Grantham
Bradley is the Lead Knowledge Scientist at Popsa, the place his workforce builds the AI methods that assist tens of millions of individuals revisit and organise their private picture libraries. His work spans on-device pc imaginative and prescient, generative AI powered by Amazon Bedrock, and manufacturing ML methods constructed from analysis by means of deployment.
Hugo Dugdale
Hugo is a Knowledge Scientist at Popsa, the expertise firm serving to tens of millions of individuals flip their digital recollections into bodily picture merchandise. He works throughout pc imaginative and prescient, geospatial information, and generative AI – constructing and deploying the methods that energy how Popsa understands and organises private picture libraries at scale.
Ayman ElSayed
Ayman is a Startup Options Architect and Gen AI specialist at AWS, partnering with UK & Eire startups to scale their AI ambitions and obtain their enterprise goals. Beforehand CTO/Product at Hawaya (acquired by MatchGroup) and co-founder/ CTO at EdTech AI startup Mavericks, he brings hands-on expertise guiding, constructing and scaling AI merchandise globally throughout the UK startup ecosystem.
Ellen Franklin
Ellen is a Senior Account Supervisor at AWS with 7 years’ expertise advising high-growth UKI startups throughout B2C, FSI, and ISV sectors. An authorized AWS Options Architect Affiliate and AI Practitioner, she sits on the intersection of economic technique and technical innovation, partnering with founders and management groups to take away limitations to scale, navigate enterprise go-to-market, unlock progress, and maximise the worth of their AWS partnership.
