Prime Video
Interactive / Animated
Key Art Automation
Developed a custom GenAI tool to automate the extraction and animation of character key art, turning standard 2D assets into interactive HTML parallax campaigns in minutes.
-
Gemini (Web & CLI)
-
Meta SAM (Foreground Extraction)
-
Stable Diffusion 1.5 Inpainting (R&D)
-
LaMa (Background Object Removal)
-
React & OpenCV
.jpeg)
1. AV & Motion Generation
2. Static Asset Production
3. Custom AI Tools & Workflows
4. UI & Prototyping
1. AV & Motion Generation
2D UI Animation
Tools: Lovable, Topaz Video AI, ElevenLabs
Producing high-impact digital content and interactive pitch decks traditionally required navigating lengthy manual animation timelines. This workflow bypasses those bottlenecks to scale premium, native digital video internally, combining strong UX/UI sensibilities with rapid motion delivery.
Workflow
Traditional
Agency Storyboarding ➔ Manual After Effects Keyframing ➔ Studio Voiceover Recording
-
Timeline: 1.5 Months | £3,000
AI Integrated
Lovable Prototype Capture ➔ Topaz 4K Upscale ➔ ElevenLabs Audio Sync
-
Timeline: 1 Week | £300
Tips
While this pipeline drastically reduces cost and time, it requires strict operational sequencing to ensure a premium final product.
-
The Trade-Off (Speed vs. Control): This workflow lacks the granular, frame-by-frame keyframe control of traditional Adobe After Effects. Because you cannot easily tweak micro-animations post-capture, the interactive motion within your Lovable prototype must be executed and recorded perfectly in real-time.
-
Resolution Preservation: When capturing live UI motion from Lovable, push the raw footage through Topaz to upscale to 4K immediately. Do this before any editorial cutting to prevent typography artifacting.
-
Audio Sequencing: Always generate and sync the ElevenLabs AI voiceover after locking the final visual edit. Generating audio too early leads to pacing mismatches.
Tool Comparison
-
Lovable vs. Manual UI Keyframing (Figma/After Effects): Traditional UI motion requires handing off static Figma screens to animators who manually keyframe every hover state, click, and transition. Lovable generates fully interactive, code-backed prototypes instantly, allowing you to capture authentic, real-time UI motion directly in the browser.
-
Topaz Video AI vs. Premiere Pro Upscaling: Standard NLE (Non-Linear Editor) upscaling simply stretches pixels, which heavily blurs critical typography and UI elements. Topaz uses temporal neural networks to physically reconstruct missing pixels, resulting in artifact-free 4K text.
-
ElevenLabs vs. Standard TTS: Standard text-to-speech sounds robotic and actively cheapens a brand's perceived value. ElevenLabs offers the granular emotional pacing and natural breath control that is strictly essential for premium enterprise marketing.
3D Cinematic Animation
Tools: Midjourney, Runway Gen-3, Luma Dream Machine, Seedance 2.0, ElevenLabs
High-end entertainment promotions and online trailers require photorealistic visualization, dynamic camera moves, and cinematic scale. Traditionally, this meant building heavy digital sets in 3D software (like Cinema 4D). This Gen-AI workflow translates static key art into rich, animated AV exploration instantly, using diffusion models to hallucinate realistic motion, lighting, and textures.
Workflow
Traditional
Agency Storyboarding ➔ Cinema 4D Modeling ➔ OctaneRender Lighting ➔ After Effects Compositing ➔Studio Voiceover Recording
-
Timeline: 2 Months | £8,000
AI Integrated
Midjourney Concept Art ➔ AI Video Generation (Runway / Luma / Seedance) ➔ ElevenLabs Studio Assembly & Audio Sync
-
Timeline: 1-2 Weeks | £500
Tips
-
Starting with Static Key Art: Do not rely purely on text-to-video for precise brand visuals. Generate hyper-realistic, branded anchor frames in Midjourney first, then pass those static images into a model like Luma or Runway to establish the 3D depth and cinematic motion.
-
Model Selection: Different models excel at different movements. Use Runway Gen-3 for photorealistic human subjects, Luma for sweeping 3D camera pushes, and Seedance 2.0 for precise, director-level control over complex lighting changes.
-
Temporal Consistency: Rather than attempting a 30-second continuous camera pan, break your storyboard into 5 to 10-second micro-shots and stitch them together seamlessly during your final editorial assembly.
Tool Comparison
-
AI Video Models vs. 3D Render Engines: Traditional 3D requires physical material mapping and hours of rendering per frame. Diffusion models (like Runway and Seedance) replicate real-world physics, lighting, and motion dynamics with extreme accuracy by predicting pixel behaviors instantly.
-
AI Sound Design vs. Foley Artists: Custom soundscapes for trailers usually require hunting through stock libraries. Tools like ElevenLabs SFX allow you to generate custom, high-fidelity sound textures directly from text prompts to match the exact pacing of your visual edit.
-
Runway Gen-3: Best for photorealistic humans, emotional micro-expressions, and subtle cinematic pacing.
-
Luma Dream Machine: Best for aggressive 3D camera sweeps, spatial tracking, and deep Z-axis pushes.
-
Seedance 2.0: Best for multi-shot continuity and granular, director-level control over complex lighting shifts.
-
ElevenLabs SFX: Best for replacing stock Foley libraries with custom, high-fidelity cinematic impacts and atmospheric soundscapes.
2. Static Asset Generation
Tools: Midjourney v6, Nano Banana 2 (Gemini 3 Flash Image)
Over the past few years of experimenting with AI prompting, I’ve found that generating a single, beautiful piece of art is simple. The real challenge, and the core focus of asset generation workflows, is programmatically generating high-fidelity, brand-consistent artwork at scale.
By shifting from one-off prompting to a systematic pipeline, the ultimate business impact becomes clear: it unlocks infinite scalability for visual assets while maintaining strict brand consistency across distinct categories.
Workflow
Traditional
Concept Sketching ➔ Individual Manual Illustration (1-by-1) ➔ Manual QA for Visual Drift
-
Timeline: 2 days for 10 assets
AI Integrated
Single Figma Anchor ➔ Midjourney --sref Batching ➔ Automated SVG Vectorization
-
Timeline: 3 hours for 10 assets
Tips
While a haphazard, trial-and-error method can be fun for casual use, it is less than ideal for enterprise workflows. Technical, guided prompting is the key to high-quality, consistent asset generation. Instead of guessing, I follow a strict structural baseline to eliminate variability and ensure high-fidelity results.
Structure:
-
Medium: What is the fundamental format? (e.g., Flat vector graphic, 35mm film photography, 3D render).
-
Subject: Tell the AI exactly what you want to be the focus (e.g., A minimalist coffee cup, no text, no human hands).
-
Environment: Where does this object live? (e.g., On a plain white background, underwater, in a sunlit studio).
-
Composition: Use actual photography, camera, or framing terms to control the angle. (e.g., Top-down flat lay, isometric angle, extreme close-up).
-
Style: The overarching aesthetic finish. (e.g., Corporate minimalism, cyberpunk, Bauhaus).
Tool Comparison
-
Midjourney v6 (Best for "Art"): It has a natural bias toward cinematic, fantastical, and highly stylized compositions, making it the premier engine for concept art and brand-level visual storytelling.
-
Nano Banana 2 (Best for "Design"): It has significantly higher accuracy for rendering actual text within images and is much more obedient to literal, un-stylized prompts. This makes it the superior choice for practical UI design, exact product shots, and assets where clarity outweighs artistic flair.
.png)
3. Custom AI Tools & Workflows
Tools: Python, Gemini CLI, Meta SAM, LaMa, OpenAI Custom GPTs
While off-the-shelf AI tools are powerful, the highest enterprise ROI comes from architecting secure, proprietary scripts to eliminate heavy manual labor. The core focus of this workflow is moving beyond standard software to build bespoke pipelines tailored exactly to a company's unique operational bottlenecks.
By replacing repetitive manual labor with customized automated systems, the ultimate business impact is immense: it transforms hours of specialized work into minutes of processing time, scaling output dramatically without scaling headcount.
Workflow
Traditional
Parallax effect: Manual Pen-Tool Masking ➔ Manual Layer Separation➔Background recreation: Photoshop Content-Aware Fill➔Parallax Motion Animation in After Effects
-
Timeline: 4 Hours per asset
AI Integrated
Parallax effect: Meta SAM Auto-Masking ➔ LaMa Background Inpainting ➔ Automated Animation
-
Timeline: 2 Minutes per asset
Tips
Building proprietary AI tools carries massive upside, but requires a strategic approach to security and development resources.
-
The Trade-Off (Capability vs. Barrier to Entry): Custom workflows have a high technical barrier to entry and require rigorous data sanitization to ensure enterprise IP security is maintained when interacting with AI models.
-
The Coding Co-Pilot: Always use an LLM as a coding co-pilot to rapidly validate your logic and build proof-of-concepts before utilizing expensive, dedicated engineering hours.
-
Data Sanitization: Never feed raw company data into a model. Always scrub and sanitize all proprietary IP before building custom knowledge bases or RAG (Retrieval-Augmented Generation) systems.
Tool Comparison
-
Meta SAM + LaMa vs. Photoshop Content-Aware Fill: Photoshop requires manual lassoing and often hallucinates background patterns on complex imagery like movie posters. Meta SAM automates the exact pixel boundary extraction, while LaMa utilizes a much larger parameter network to ensure clean, artifact-free background reconstruction.
-
OpenAI Custom GPTs vs. Standard Chatbots: Traditional chatbots rely on rigid, frustrating decision trees. Custom GPTs allow you to rapidly build a proof-of-concept over a scrubbed knowledge base, using advanced semantic understanding to validate complex prompt logic before a full backend deployment.
Tools: Gemini CLI, Meta SAM2, LaMa
As an exploration into 'vibe coding', I wanted to explore how I could create a tool that designers could use to speed up their workflows by automating traditionally labour-intensive work.
The below video is an example of a tool I created using Gemini CLI and various other tools. Use of this tool enables the designer to upload static key art, such as movie posters, select the key foreground and midground elements, and use AI models to generate a dynamic, parallax effect.
See the full project (and interactive element!) here:
Tools: Gemini (Canvas)
I created the below tool as an example of a real tool I made as part of the Galatasary FC project. I designed the app with a focus on scalability, by using a semantic colouring architecture.
I used Gemini's web tool, 'Canvas' to rapidly create a quick app where I could visualise screens of the app in different colour schemes, to match the branding of different organisations.
Using the below tool, you can see how components of the design system and a sample screen change dependent on the semantic colours which have been set up. You can create additional examples by providing the brand colours and auto-generating the rest of the colour scheme using standard colour theory.
4. UI & Prototyping
Tools: Lovable, Custom GPTs, Figma (Tokens Studio), JSON
The traditional product design cycle heavily relies on static wireframes, often resulting in painful developer handoffs and weeks of waiting just to test basic interactions. The core focus of this workflow is bypassing static screens entirely to instantly build interactive, on-brand front-end code.
By collapsing the design-to-development pipeline, the ultimate business impact is profound: it allows for the immediate, live testing of complex micro-interactions and logic, aggressively accelerating enterprise go-to-market strategies.
Workflow
Traditional
Static Figma Screens➔ Custom Interactions➔ Developer Handoff
-
Timeline: 3 Weeks for MVP project
AI Integrated
Figma JSON Export ➔ Custom GPT Translation ➔ Gen-AI design in Lovable
-
Timeline: 2 days for MVP project
Tips
To successfully generate production-ready UI, you must tightly control the AI's structural and visual outputs.
-
The Trade-Off (Speed vs. Hallucination): AI code generators are incredibly fast, but they will heavily hallucinate styling, padding, and CSS if strict brand guardrails are not mathematically enforced from the start.
-
The JSON Injection Protocol: Never let the AI guess your brand styles. Export your entire Figma design system as JSON tokens (using Token Studio) and feed them to a Custom GPT. This acts as an irrefutable structural rulebook, forcing the UI generator into absolute brand compliance.
Tool Comparison
-
Lovable vs. Static Figma Prototypes: Figma only mimics interaction via static, pre-determined screen transitions. Lovable generates actual, functional React/HTML code. This allows for live, hands-on testing of complex spatial logic (like bezier curve connections, state changes, and drag-and-drop physics) that are mathematically impossible to validate in a static design tool.
