Divyanshu Goyal
Applied Scientist at Adobe
AI researcher and engineering leader with 8+ years at Adobe, specializing in agentic AI and multimodal LLMs. I led the technical strategy that took Adobe GenStudio for Performance Marketing from 0 to 1, driving multi-million-dollar adoption across the Fortune 500. Today I am tech lead for a 7-person team building VLMs, judge models, and retouchers. My work includes 6 patents and publications at CVPR, ICML, and AAAI workshops.
About Me
Education
M.S. in Computer Science
Georgia Institute of Technology
Machine Learning
B.E. in Computer Science
Birla Institute of Technology and Science, Pilani
M.Sc. in Mathematics
Birla Institute of Technology and Science, Pilani
Expertise
Tools & Languages
Publications & Patents
ATune: Unsupervised Photo Retouching Through Image-Editor Controls
Divyanshu Goyal*, Gabriel Ruiz*, Amir Meisami (*equal contribution)
DistortBench: Benchmarking Vision Language Models on Image Distortion Identification
Divyanshu Goyal, Akhil Eppa, Vanya Bannihatti Kumar
JudgeBreak: Is Taste Hidden in the Layers? Stress-Testing Reward Models and LLM Judges with Representation-Based Subjective Evaluation
Vanya Bannihatti Kumar, Anish Narang, Mayank Anand, Akhil Eppa, Divyanshu Goyal
Curiosity-Driven LLM-as-a-judge for Personalized Creative Judgment
Vanya Bannihatti Kumar, Divyanshu Goyal, Akhil Eppa, Neel Bhandari
Cultural Adaptation of Images Using Generative Artificial Intelligence
Akhil Eppa, Mayank Anand, Divyanshu Goyal
Guiding Language Translation with Translation Documents Using Machine Learning
Divyanshu Goyal, Akhil Eppa, Mayank Anand
Video-to-Video Culturalization Using Generative AI
Akhil Eppa, Divyanshu Goyal
Learning to Be Surprised: A Curiosity-Driven Framework for Creative Judgment
Divyanshu Goyal, Vanya Bannihatti Kumar, Akhil Eppa, Neel Bhandari
Systems and Methods to Provide Parameter-Efficient Fine-Tuned Models
Yuanyou Wang, Naveen Vangala, Mayank Anand, Kunal Kumar Jain, Jose Mathew, Eapen Jose, Divyanshu Goyal, Asmita Chihnara, Arif Abdullah, Anand Dantu
Brand-Aligned Marketing Content Generation Using Structured Brand Data and Generative Models
Divyanshu Goyal, Mayank Anand, Jose Mathew
Experience
Applied Scientist
/Adobe, San Jose, CAJune 2021 – Present
FixIt: Multimodal Models for Marketing & Visual Quality (Tech Lead, Team of 7)
- Architected a sub-second JEV-style judge model integrated directly into the Firefly AI Agent evaluation harness, improving the agent's outputs through real-time feedback and reducing user revisions by 40%.
- Directed the development of three specialized model streams: domain-adapted VLMs, marketing-quality judges, and an unsupervised photo retoucher trained entirely without human-edited pairs.
ATune: Unsupervised Photo Retouching (Co-Lead with Adobe Research)
- Co-led project architecture and evaluation design with Adobe Research as co-first author, owning the synthetic data pipeline, training recipe, and human evaluation framework, and mentored a PhD intern on the project.
- Trained a compact 54M-parameter model that performs non-destructive photo adjustments via off-the-shelf editors (RawTherapee/darktable), learned from 10M unedited photographs with zero expert targets.
- In a 157-rater human study (3,524 judgments), beat GPT-6 Astra (54.3% of decisive votes, 95% CI 50.3–58.3) and matched Claude Opus 5 (50.5%); preferred over GPT-6 Astra by ≥8 of 9 published quality judges across FiveK, PPR10K, and SPAQ.
- Ran data-scaling studies from 1M to 10M images: each larger corpus wins or ties, and no judge on any test set prefers a smaller one.
Adobe Marketing VLM (Gemma-3-12B, SFT & GRPO Alignment)
- Fine-tuned Gemma-3-12B via distributed SFT followed by multi-round GRPO on 8× H100s, leveraging LoRA at rank 256 (1.19B trainable parameters) across 7 custom reward functions.
- Identified reward exploitation in ordinal rating prediction: the policy collapsed toward the middle rating (84% of predictions vs. 56% in human annotations), exploiting the bounded distance-based reward. Introduced a cubic distance penalty and class-weighted advantages for underrepresented ratings, reducing middle-rating predictions to 64%.
- Built a fault-tolerant annotation pipeline that labeled ~10M images with a teacher VLM to create the SFT dataset.
- Built a Photoshop agent powered by the marketing VLM that edits photos toward marketing taste, turning the model's judgment of what makes an image effective into concrete Photoshop edits.
JEV-Style Marketing Judge Models
- Trained Qwen3.5-4B LoRA judges that output calibrated scores in a single forward pass instead of generating text, on a 1.1M soft-labeled dataset curated from 14 human-rated sources, raising KonIQ correlation to 0.663 SRCC (vs. 0.421 for the reference model).
- Curated a 67K-item held-out benchmark, leak-checked against all training images via SSCD copy detection, and gated every model release on a pre-registered non-inferiority test, so a new model shipped only if it held its quality on unseen data.
- Conceived and built a live demo suite that turns the judge from a scorer into a creative engine: it picks the best crop of an image (0.82 pairwise accuracy on GAIC vs. 0.71 for the score-based baseline), guides photo retouching, scores marketing and custom attributes, re-ranks Adobe Stock for marketing use, and flags low-quality assets. Presented it to leadership and external partners, opening new customer engagements.
- Diagnosed a vision-encoder collapse that left every checkpoint image-blind within 200 steps while the loss looked normal, and added an automated vision-health check that halts training on collapse, catching silent failures the loss misses.
Adobe GenStudio for Performance Marketing (Tech Lead, Team of 20+)
- Took GenStudio from prototype to a multi-million-dollar enterprise product (launched Oct 2024), leading its technical strategy and the production API layer that serves its LLM-powered features to Fortune 500 customers.
- Co-invented automated brand-guideline extraction and compliance scoring (US20250117829), the foundational engine for Adobe Brand Intelligence.
- Developed parameter-efficient fine-tuning pipelines on a Fortune 500 customer's brand data, directly closing a custom-model enterprise deal.
Evaluation Research
- DistortBench (First Author; CVPR/ICML 2026 Workshops): Established a 13.5K-question visual distortion benchmark across 27 types and 5 severity levels; the best of 18 VLMs (61.9%) trails a human majority vote (65.7%).
Machine Learning Engineer
/Adobe, Bangalore, IndiaJuly 2016 – August 2019
ML Infrastructure & Serving
- Joined Adobe's core ML serving platform team at inception and helped plan the end-to-end serving infrastructure for internal workloads, including field-aware factorization machines.
- Scaled production inference throughput to 3,500 QPS per node at 0.3–0.6 ms p99 latency (Adobe Tech Blog), implementing a garbage-collection-free asynchronous event pipeline on LMAX Disruptor.
- Built the model lifecycle platform for versioned deployment, monitoring, and staged rollout.
Projects
VibeNanoChat: GPT-2 Scale LLM Pretraining from Scratch
A from-scratch framework for pretraining, scaling, and evaluating GPT-2 scale language models, featuring depth-parameterized scaling laws, a custom distributed Muon optimizer, and a 35+ benchmark evaluation suite.
Marketing Quality Vision-Language Models
Fine-tuned Gemma-3-12B with SFT and multi-round GRPO, then trained Qwen3.5-4B judges for marketing attributes, checked on a leak-checked benchmark suite.
Adobe GenStudio for Performance Marketing
Built the founding prototype and pitched it to Adobe leadership, leading to the launch of a generative-AI platform for creating on-brand marketing content at scale.