Itim Kongsakulvatnasook

Week 04 - Thesis Topic Exploration

26 Jan - 01 Feb 2026

Thesis Topic Exploration

Topic 1: Gaussian Splats for Stylized/NPR FX

Concept

  • Investigate if Gaussian splatting's discrete representation better suits non-photorealistic VFX (anime smoke, hand-drawn fire, stylized magic) than traditional volume rendering.

    • Why?

  • Problem: Volume rendering optimized for photorealism; NPR effects fight against natural smoothnesst

  • Hypothesis: Gaussian splats' discrete nature naturally matches hand-drawn, stylized aesthetics

    • Approach

  • Convert Houdini sims to Gaussian representations

  • Develop art-direction tools for Gaussian manipulation

  • Compare quality, workflow, performance vs. volumes

  • Focus: anime smoke, magical effects, stylized explosions

    • Topic 2: Real-Time Gaussian Splat Previews for Volume Look-dev

    Concept

  • Use real-time Gaussian splat rendering as fast, accurate preview system for volume look development, while final renders use traditional high-quality volume rendering.

    • Why?

  • Problem: Volume rendering slow (minutes/frame), slowing iteration

  • Solution: Gaussian previews render in real-time

    • Approach

  • Convert volumes to Gaussian representations

  • Build real-time viewport integration (Houdini plugin)

  • Artists iterate with instant Gaussian preview

  • Validate preview accuracy vs. final renders

  • Measure workflow time savings

    • Topic 3: AI Style Transfer Tool for VFX Production Pipeline

    Concept

  • Build production-ready style transfer system that solves VFX-specific with temporal coherence using 3D render data, deep pipeline integration, and real-time interactive workflows.

    • Why?

  • Problem: Look development slow/expensive; existing tools lack temporal coherence and pipeline integration

    • Functions

    VFX-Grade Temporal Coherence

  • 3D-aware warping using motion vectors + depth + cryptomatte

  • Per-object style consistency - characters, background etc.

  • Depth-stratified application

    • Pipeline Integration

  • Nuke plugin with multi-pass render awareness

  • Render pass preservation

  • Shot metadata integration

    • Real-Time Interactive System

  • Progressive refinement (0.3s - 2s - 10s quality levels)

  • Semantic region editing

  • Multi-reference style mixing

  • Temporal scrubbing preview

    • Triangle Splatting Exploration

    Create triangle splat from images

    Source

    Key Innovation

  • Trains triangle geometry directly from multi-view images

  • Each triangle has position, orientation, size, color, and opacity

    • Prerequisites

  • Python 3.11

  • CUDA 12.6

  • Git with submodule support

    • Input Requirements

  • Multi-view images (COLMAP format or NeRF synthetic)

  • Camera poses (from COLMAP or provided transforms.json)

  • Shot metadata integration

    • Mesh Export Pipeline

  • Convert to .OFF Format

  • Importing to Unreal Engine 5 or any tool supporting .OFF or convertible to .OBJ

  • Multi-reference style mixing

  • Temporal scrubbing preview

    • Triangle Splatting VS Photogrammetry

    Triangle Splatting

  • Handles reflections/specularity better than Photogrammetry

  • Fewer images needed (can work with gaps)

  • No texture seams (vertex colors)

  • Best for complex reflective surfaces (wet, shiny objects) and organic forms (foliage, natural scenes)

    • Photogrammetry

  • Geometrically accurate (real-world measurements)

  • Clean topology

  • Standard UV textures

  • Best for architecture/buildings, hard-surface objects (mechanical parts), assets needing manual editing/retopology

    • To-Do Lists

  • Use OpenUSD to store multiple variants of style of the same gaussian splat object

  • Testing original and stylized gaussian splat in Unreal

  • Refine Gaussian splat style AI transfer tool