Skip to content
Sumit Chatterjee

Index — Selected Research

Selected Research

Six investigations into closing the gap between generative diffusion models and the working tools of a VFX compositor — HDR linear generation, color management, dataset infrastructure, and the surrounding workflows.

  1. 01

    Research · 2024 — present

    HDR Image Generation from Diffusion Models

    Bringing scene-referred linear 16-bit HDR generation to Qwen-Image-2512 — first via LogC4 encoding, then through native linear-space VAE and MMDiT fine-tuning.

    • HDR
    • DIFFUSION
    • QWEN-IMAGE
    • MMDIT
    • LogC4
    • EXR
    • VAE
  2. 02

    Open source · 2024 — present

    Foundry Nuke Nodes for ComfyUI

    A package of ComfyUI nodes that replicate common Foundry Nuke nodes — IO, merge, transform, grade, blur, OCIO, LUT — so a compositor can build AI-augmented workflows in tools that behave like the ones they already trust.

    • COMFYUI
    • NUKE
    • OCIO
    • ACES
    • OPENIMAGEIO
    • VFX
  3. 03

    Infrastructure · 2024 — present

    Autonomous Dataset Agent for HDR Training Data

    An autonomous agent that builds training datasets from 360° HDR panoramas — extracting perspective views, evaluating each one with vision, and exporting only what passes.

    • LANGGRAPH
    • VLM
    • MISTRAL
    • vLLM
    • AGENT
    • HDR
  4. 04

    Private infrastructure · 2024 — present

    Bit-Depth Expansion Network

    A neural network that lifts 8-bit log-encoded source material into 16-bit float without the banding and quantization artifacts that defeat naive precision-padding.

    • U-NET
    • DEBANDING
    • PYTORCH
    • VFX
    • OCIO
  5. 05

    Open source · 2024 — present

    ComfyUI MCP Server

    An MCP server that lets AI coding agents — Claude Code, Cursor, Kilo Code, Cline — request images from a local ComfyUI instance. Closes the gap between code generation and asset generation in agent workflows.

    • MCP
    • COMFYUI
    • AGENTS
    • CLAUDE-CODE
    • CURSOR
  6. 06

    Methodology · 2024 — present

    Custom LoRA Training Pipelines

    Notes on training diffusion LoRAs as a loss-design problem rather than an architecture problem — across Flux, Wan 2.1, and Qwen-Image, for HDR generation, HDR reconstruction, and multi-character coherence.

    • LORA
    • DIFFUSION
    • FLUX
    • WAN
    • QWEN-IMAGE
    • FLOW-MATCHING