Pixshop — consistent portraits from a single selfie
I built Pixshop (https://www.pixshop.art) after running into the same issue with most AI photo tools:
they either need a bunch of photos + a training step, or they drift your face across outputs.
The goal:
use a single selfie, skip training, and still keep identity consistent while changing the setting.
How it works:
- Upload one selfie
- Pick a “Look” (headshot, dating photo, travel, etc.)
- Each look runs on a fixed “recipe” (structured prompt + tuned config)
- Output is a small batch (4–6 images), with preserved identity
The interesting part is the recipe layer.
Instead of open-ended prompting, each look encodes:
- camera distance + framing
- lighting direction and intensity
- background constraints
- facial anchoring to reduce drift
In practice, this mattered more than model choice for consistency.
There’s no per-user training or fine-tuning step — generation runs directly on image models, so results come back quickly.
Stack: Next.js, Vercel Blob, Neon + Drizzle, QStash for async jobs, Clerk + Stripe.
Free tier: 3 credits, no card required.
Happy to answer anything about:
- how we keep identity stable across different looks
- what failed before landing on the recipe approach
- async generation pipeline
It is a major frustration for users when they have to upload twenty different photos of themselves just to get one decent AI headshot that ends up looking like a distant cousin rather than a real portrait.
The real breakthrough here is the "recipe" layer because by standardizing the lighting and camera angles you remove the variables that usually cause an AI model to hallucinate facial features which provides a much more predictable result than an open-ended prompt.
Since you are skipping the per-user training step to keep things fast are you finding that certain lighting "recipes" work better than others for maintaining skin tone and facial structure consistency across different ethnic backgrounds?
I think latest image models are getting stronger for character consistency, techniques like extra Lora weights were useful but slowed down the end-to-end experience. For high-end character consistency requirements, we still need special prompting for maintaining detailed skin tone and facial structure consistency across different ethnic backgrounds.
Using special prompting to maintain skin textures and ethnic features is a smart move because relying solely on the base model often leads to generic or smoothed-out results.
Avoiding extra Lora weights keeps the generation speed fast which is the biggest factor for users who want to see their "Look" instantly without waiting for a training queue.
I focus on this kind of technical positioning in high-tier PR and media placement where we highlight a product's unique logic to build brand authority on major news platforms.
Are you planning to let users create their own custom "recipes" or will you keep the selection curated to ensure the output quality stays consistent?
Both:
- User can leverage baked recipes to generate the results quickly
- User can also create their own custom recipes
Giving users both options is the right call — curated recipes lower the barrier for casual users while custom recipes unlock the power users who will eventually become your most vocal advocates.
The interesting challenge with custom recipes will be preventing users from creating combinations that break identity consistency — have you thought about adding guardrails or a 'consistency score' preview before they generate?
This is a really clean take on a messy problem — skipping training and keeping identity stable is a big win 👍
The “recipe layer” is the interesting part. That’s basically turning randomness into controlled output, which most tools don’t do well.
A few thoughts / questions:
→ You’re right — structure > model choice here
That could actually be your main positioning:
“consistent results without training”
→ One thing to watch:
users will push it with edge cases
(different angles, expressions, lighting in the original selfie)
Curious how robust it is there?
→ Also feels like a big unlock could be:
→ before/after examples for each “look”
That builds instant trust
→ And maybe:
→ let users slightly tweak recipes (lighting, vibe)
without breaking consistency
Curious — are users coming back to generate more looks, or mostly trying it once?
Also, I’m running a small project (Tokyo Lore) where we highlight tools like this with a focused group of early users.
Since you’ve solved a real UX gap in AI images, this could be a strong fit — happy to share more 👍
Thanks for the input!
It is true that user may push with edge cases, I try to follow the 80/20 rule -- solving the 80 first, then optimizing the rest 20.
I am just getting started, let's see how it works for end users, will share more for the "coming back/trying it once" question.
BTW, is there a homepage for 'Tokyo Lore'?
Makes sense — 80/20 is the right way to start, especially for something like this 👍
Curious to see how it holds up once people really start pushing it.
On Tokyo Lore — yeah, here’s the homepage:
→ tokyolore.com
It’s basically a small, focused round where we bring in builders + early users and see how products actually perform (not just feedback, but what people use, question, come back to, etc.).
Given what you’re building around consistency (which is a real gap in AI images), this could be a good fit to validate retention + edge cases in a more structured way.
Happy to share more or get you in if it sounds useful 👍
The "recipe layer" approach is smart — encoding camera distance, lighting, and framing as fixed parameters rather than letting the model interpret free-form prompts explains a lot of why identity stays consistent.
Thanks! Let's see how it works :-)