AI/ML Expertise

LoRA & Diffusion Model Fine-tuning

I train custom LoRAs and fine-tune diffusion models for VFX-specific styles, HDR output, and character consistency on my RTX 5090 32GB.

Models I Train

Image Generation

Flux.1Flux.1-devFlux.1-schnellQwen-ImageZ-Image-TurboSDXLSD3

Video Generation

Wan2.2-T2V-A14BWan2.2-I2V-A14BWan2.2-TI2V-5BCogVideoXMochi-1HunyuanVideoLTX-Video

Training Tools I Use

KohyaSS

My go-to for complex training configs. Full control over every parameter.

AI-Toolkit

Quick Flux LoRA training. Best for rapid iteration.

SimpleTuner

When I need Prodigy optimizer and advanced schedulers.

My Captioning Pipeline

Built a custom captioning tool using Gemma 3-12B via vLLM. Concurrent batch processing gives me 50+ images/min - 10x faster than BLIP-2.

Gemma 3-12BvLLM servingAsync batch processingCustom VFX prompts

What I Train For

VFX Style Matching

Training models to match specific film looks and color grades for concept art.

Character Consistency

LoRAs for consistent character appearance across storyboard generations.

HDR Output

Fine-tuning for LogC4/ACEScg color spaces for professional VFX integration.

Custom Assets

Training on proprietary environments, props, and styles for studio work.

Technology Stack

KohyaSSAI-ToolkitSimpleTunerPyTorchCUDAProdigyvLLMGemma 3-12BWandBDiffusers

Expertise by Sumit Chatterjee

Industrial Light & Magic, Sydney

Back to Portfolio