Many E-commerce brands were losing sales due to costly photoshoots, slow editing cycles, and inconsistent imagery.

To address this, I helped design Photogenix, an AI-powered product imaging platform that helps fashion and retail brands generate production-ready visuals at scale. Initially spun out from a micro-feature in Catalogix (just with background change + marketplace resizing), and quickly evolved into a standalone platform with a broader set of capabilities.

I was involved from 0→1, identifying pain points, defining requirements, and shaping the solution from the ground up.

The platform made catalog creation faster, cheaper, and more consistent, helping brands deliver high-quality visuals with a fraction of the time and cost.

Photoshoots are slow, expensive, and brittle. They block catalog velocity, force compromises on quality, and prevent brands (especially smaller ones) from staying visually relevant to fast-moving trends.

Through many in-depth user interviews and spending time with the cataloging team, we uncovered several key insights:

Key Insight: Users needed affordable, time efficient solution that also gives them creative control without technical complexity. They wanted to focus on brand decisions, not prompt engineering.

How might we enable non-technical merchants to create production-ready, contextual product imagery that matches their brand standards without complex AI prompting or expensive photoshoots?

Based on the derived design principles, we designed Photogenix around a simple 3-step flow:

This approach minimized friction for merchants who were frustrated with complex, text-based prompting.

A merchant uploads an image, and Photogenix instantly shows tailored options based on the use case on-model generation, mannequin-to-model transformation, product background change, or flat-lay integration. The user can make quick selections for model, background, footwear, or any specific visual element they want to modify or keep.

Through early feedback, we discovered that not every merchant needed the full workflow. Some only wanted to replace a background, while others just needed a new model. Instead of fragmenting the experience with multiple tools, we introduced a “Keep Original” option within the same flow. This gave users the flexibility to change the background, the model, or both, without context-switching or juggling separate modules.

For added creative control, merchants can also upload or describe a model, background, or other desired attributes within each section, ensuring complete customization while keeping the interface intuitive.

Since Photogenix supports multiple content types, from static product photos to generative videos, we applied a unified, card-based design system across all workflows. The preset architecture keeps the visual logic consistent, while modular controls adapt dynamically to each use case. This system ensures a predictable, scalable experience whether the user is editing a product, mannequin, or video frame.

As adoption grew, merchants wanted to process entire catalogs at once instead of running image generation SKU by SKU.

Designed for large-scale catalog operations.

• Users upload a zip file with multiple folders, each representing a product or category.
• Each folder can have custom settings (e.g., office look for shirts, street look for sneakers, minimal studio look for accessories).
• The system processes everything asynchronously in the background. Users are free to continue exploring the platform, and they get notified once assets are ready.

While Photogenix produced high-quality, production-ready images, generative AI models still had edge-case failures:

• Artifacts in complex areas (garments with fine textures, accessories not rendering correctly)
• Anatomical distortions (hands/fingers, a common SDXL limitation)
• Background blending issues (shadows, edges not harmonized)
• Inconsistent lighting or mismatched expressions

From looking at Posthog sessions and feedback sessions, ~30% of generated outputs showed at least one of these issues (n=500 images, July 2025), not enough to break the product, but significant enough to require an efficient, user-friendly correction path.

Instead of forcing users to regenerate repeatedly, we introduced a retouch toolkit inside the same "Renders" screen, so fixes could be made quickly without leaving the workflow.

Users felt in control rather than at the mercy of AI. Fixes took an average of 45 seconds compared to 8-15 minutes for regeneration or Photoshop edits. Despite AI’s limitations, professional output quality remained consistent.

We refined Photogenix V1 through insights from user interviews and PostHog sessions, fixing points of friction to raise clarity and speed. As feature requests grew, it became clear merchants would continue asking for new controls, so we redesigned the interface to support future additions without disturbing the main workflow.

At launch, users picked a subject type (model, mannequin, or product) from a single modal. Once we introduced options like flat-lay to model, video, and ad creation, this setup became confusing. We replaced the modal with feature cards that clearly show each use case. The cards also fix a recurring issue: many merchants uploaded weak source images without noticing the impact. Each card now shows good vs. bad examples to guide them before they start.

The old sidebar made the generation flow feel crowded and hard to extend. We moved everything into one contextual panel, which cleaned up the layout and made room for additions such as “Additional Prompt,” “Retouch,” and “Lighting.”

The original mask editor looked intuitive, but session recordings revealed repeated failed attempts, users couldn’t tell exactly what was being masked. We redesigned it to show the editable mask in real-time, directly on the image. This gave users instant visual feedback and far greater precision during edits.

Initially, we displayed generation results in a modal popup showing the original vs. generated image, credits used, and editing options like Retouch and Relight. As features expanded, this layout became cramped and hard to navigate.

We transitioned to a full-screen render view, adding a right-hand sidebar for all editing tools and detailed information. This not only improved readability but also allowed us to compare image variations side by side to ensure visual consistency.

Beyond these larger redesigns, we shipped continuous refinements based on usage data, improving tool discoverability, clarifying actions, and removing unnecessary steps to keep the experience simple and predictable as new capabilities were introduced.

Though my primary role was product design, I also contributed directly to the AI workflow exploration. I independently learned ComfyUI to better understand how workflows were constructed.

• Model-to-Model & Mannequin-to-Model workflows: Built early prototypes integrating Depth and Canny ControlNets to preserve pose and structure while maintaining stylistic flexibility. These prototypes provided a working foundation for our generative fashion pipeline.
• Continuous improvements: Iterated across multiple model updates, fine-tuning prompt conditioning, sampler settings, and ControlNet weights to improve realism, lighting consistency, and fabric detail.
• Bridging design & AI: This effort helped me translate design intent into technical workflows, ensuring that user-facing features (like mannequin-to-model conversion) were feasible and aligned with real model capabilities

Photogenix meaningfully improved catalog workflows for merchants across fashion and retail:

Designing Photogenix taught me that impactful AI products are less about "magic" and more about control, trust, and recovery. The best outcomes came from combining user empathy, technical curiosity, and awareness of business constraints. Users didn't need perfect outputs, they needed confidence the system worked with them, plus simple ways to fix mistakes.

This was a true 0→1 journey, from identifying pain points to shaping the end-to-end product experience. The work showed me the value of combining user-centered design with technical curiosity, and reinforced that impactful product design is not just about polished screens, but about making workflows faster, cheaper, and more reliable at scale.