Production Diffusion Pipelines at Dashtoon

As a Research Engineer at Dashtoon (Mar 2025 – Aug 2025), I built and shipped production-grade diffusion pipelines for animated character generation at 2048 resolution, spanning LoRA training, model integration, dataset curation, and video generation evaluation.

Flux LoRA Adapters

• Trained and shipped Flux LoRA adapters for 20+ characters across 8 animation styles
• Production inference optimized with FP16, batching, and caching, validated at 2048 resolution
• Improved character consistency, prompt alignment, and style fidelity across a diverse cast

Hidream-l1 (MoE) Integration

• Integrated Hidream-l1, a Mixture-of-Experts diffusion model, into the animated character pipeline during early development
• Adapted inference and post-processing to deliver smoother 2048 resolution outputs with improved prompt alignment
• Hardened pipeline reliability around high-resolution generation, model switching, and output consistency

Regional Bias Mitigation Dataset

• Built a targeted dataset of 10,000+ images from 100+ Bollywood movies via automated extraction pipeline
• Goal: reduce regional bias in image generation for South Asian character representation
• Pipeline covered automated clip extraction, frame selection, quality filtering, and metadata tagging

Google Veo 2 & Frameo AI

• Worked hands-on with Google Veo 2 during early access/POC to evaluate video generation quality and workflow fit for character animation
• Partnered with the GCP team to integrate the Veo 2 workflow into the new Frameo AI platform, aligning API usage, input/output formats, and reliability requirements for production

Stack

Python · PyTorch · Flux · LoRA · Hidream-l1 (MoE) · Google Veo 2 · GCP · FP16 · Diffusion Models