Midhun Harikumar
ML Lead and Engineering Manager at Morphic
I lead ML efforts at Morphic as both ML Lead and Engineering Manager. My team builds zero-shot character customization, camera controls, and video outpainting/inpainting for video diffusion models.
Previously at Adobe, I was a founding member of Adobe Firefly and deployed Adobe's first production diffusion model to Photoshop as tech lead. I worked on controlled image generation, led zero-shot customization efforts, and managed the enterprise foundational model team for model customizations.
Outside of work, I have a keen interest in DIY electronics, Physical AI, and the history of science. I also follow Formula 1 and the World Endurance Championship racing series.
LinkedIn | Google Scholar | X
Publications
Reshoot Anything: A Self-Supervised Model for In-the-Wild Video Reshooting
CVPR 4D World Models Workshop 2026
Precise camera control of video diffusion models using self-supervised dataset generation strategy and training without synthetic pairs.
TexSliders: Diffusion-Based Texture Editing in CLIP Space
ACM SIGGRAPH 2024
A novel framework for semantic texture manipulation by leveraging the latent space of pretrained diffusion models and CLIP embeddings. Enables fine-grained control over texture attributes while preserving structural integrity.
PREDITOR: Text-Guided Image Editing with Diffusion Prior
A novel approach for text-guided image editing that utilizes diffusion priors to achieve semantically meaningful modifications while maintaining image coherence and photorealism.
Enhanced Controllability in Diffusion Models through Feature Disentanglement
ICML 2024
An innovative architecture for diffusion models that separates spatial content and style representations, leading to improved manipulation capabilities and more precise control over generated outputs.
Patents
19 granted US patents — Adobe Inc., Amazon Technologies
2026
| Patent | Title |
|---|---|
| US 12,596,766 | Automatically generating an image dataset based on object instance similarity |
| US 12,586,271 | Color conditioned diffusion prior |
| US 12,586,259 | Image generation using a text and image conditioned machine learning model |
| US 12,579,608 | Generating tile-able patterns from text |
| US 12,555,288 | Controllable diffusion model |
| US 12,536,722 | Utilizing a diffusion neural network for mask aware image and typography editing |
| US 12,530,822 | Utilizing a diffusion prior neural network for text guided digital image editing |
2025
| Patent | Title |
|---|---|
| US 12,511,877 | Object-agnostic image representation |
| US 12,493,937 | Prior guided latent diffusion |
| US 12,406,334 | Preset style transfer |
| US 12,322,007 | Systems and methods for color palette optimization |
| US 12,299,939 | Generating novel images using sketch image representations |
| US 12,277,630 | Unsupervised style and color cues for transformer-based image generation |
| US 12,260,480 | Machine learning-based layout generation (also CN, AU, DE) |
2024
| Patent | Title |
|---|---|
| US 12,008,698 | Image segmentation using text embedding |
| US 11,934,448 | Keyword localization digital image search |
2023 and earlier
| Patent | Year | Title |
|---|---|---|
| US 11,574,392 | 2023 | Automatically merging people and objects from multiple digital images |
| US 11,138,257 | 2021 | Object search in digital images |
| US 10,789,569 | 2020 | System to determine item footprint (Amazon) |