More consultations booked with AI before-and-after previews

The situation

The undecided consultation is the most expensive moment in any transformation service business. A client sits across from a practitioner, genuinely interested, but unwilling to commit because they cannot picture the result on their own face, their own hair, or their own space.

Stock before-and-after photos help, but they feature someone else. Makeover generates the preview using the client's own photo. The result lands differently: it is personal, it is theirs, and it closes the gap between "I'm interested" and "let's book."

Dental practices, aesthetic clinics, hair salons, and interior designers in 40+ transformation categories use Makeover to show what the outcome looks like before the client says yes. The client reported that consultation-to-booking conversion increased 35% for practices using it.

That set the quality bar for the build: the preview had to be realistic enough to trust at consultation distance, which meant category-specific models trained on real before-and-after pairs rather than one general image model.

What we had to solve

• Getting previews realistic enough that clients trust them without overselling

  An AI preview that looks obviously artificial does more damage than no preview at all: it signals that the practice is using technology it does not fully control. The output had to look realistic at consultation viewing distance on a tablet, which meant training category-specific models on real before-and-after pairs rather than using a general image generation approach. Dental whitening, hair color, aesthetic contouring, and interior finishes each needed their own model. Getting the quality bar right for each category was the core research challenge.

• Generating the preview fast enough for use during an active consultation

  A preview that takes five minutes to generate is not a consultation tool. It is a disruption. The entire flow had to complete in under 60 seconds: photo capture, upload, inference, and display back to the practitioner and client. That required a job queue architecture that could route requests to the right category model, handle inference retries without adding visible latency, and deliver the result directly to the consultation interface without requiring a page refresh. Replicate's GPU infrastructure handled the inference; the engineering challenge was the end-to-end latency budget.

1. Client sees a photorealistic result in under 60 seconds, from a model trained on real before-and-after pairs

   The client uploads a photo or it is captured on a tablet in the consultation room. Makeover runs the image through a category-specific AI model (dental, hair, aesthetics, interiors) and returns a photorealistic result in under 60 seconds. The model is trained on real before-and-after pairs for each category, not on general image generation data.

   

2. Clients express a preference rather than committing blindly to one fixed outcome

   Practitioners show multiple versions in the same session: lighter vs. heavier treatment, one approach vs. another. The client sees options rather than a single fixed outcome, which supports a more natural consultation conversation and allows the client to express a preference rather than committing blindly.

   

3. Returning clients compare new previews against previous sessions, no file hunting before appointments

   Previews are saved to the client's record and linked to their profile. Returning clients can compare a new preview against previous sessions. Practitioners access the image history before a follow-up appointment without hunting for a saved file.

   

4. Clients who leave without booking get a reminder with their own preview, not a generic brochure

   After the consultation, the practitioner sends the client their personalized preview via email or SMS. Clients who leave without booking receive a reminder with their preview attached, a follow-up that is personal because the image is theirs, not a generic brochure.

   

How it runs

Next.js

The practitioner dashboard and client-facing share pages needed fast initial loads with optimistic UI updates so the consultation flow did not stall while waiting for the inference result to arrive.

Replicate

Category-specific image generation models run on Replicate's GPU infrastructure. Replicate's API allowed us to deploy separate fine-tuned models for each transformation category without managing GPU servers.

AWS

Preview images are stored on S3 with signed URLs that expire after 30 days, keeping client photos secure and ensuring share links sent to clients cannot be accessed indefinitely.

Node.js

The job queue manages inference requests, routes each job to the correct category model, handles retries for failed inference calls, and delivers the completed preview back to the practitioner dashboard without requiring a page reload.