Bypassing the Mesh II: Mapping the 3DGS Budget Spectrum

Part of the Bypassing the Mesh series. The hub article covers the full landscape. This spoke maps the budget spectrum — from enterprise 4DGS pipelines to the browser-based indie revolution.

There is a version of this story where 3D Gaussian Splatting (3DGS) is enterprise infrastructure — a technology that requires a 192-camera rig, a Framestore pipeline, and a Superman-sized budget to get anywhere near. That version exists. It is not the whole story.

The more interesting version is what happens when the same mathematical technique lands in a browser tab, runs on a phone camera, and produces something a single filmmaker can actually use.

The access barrier to volumetric production has not been lowered. It has been removed. What remains are three friction points the tooling has not yet abstracted away — and understanding them is the difference between a pipeline and a weekend of frustration.

The Ceiling

Framestore's work on Superman set the technical ceiling: a 4DGS pipeline capturing volumetric performances across approximately 40 shots using a 192-camera array and Infinite Realities' capture infrastructure. The result was living photography — actors preserved at a level of fidelity that would have required full CGI digital-double simulation three years ago — integrated directly into a studio VFX pipeline.

That is what the top of the stack looks like. It is useful to know it exists. It is not where the democratisation story lives.

What Phone-Tier Actually Produces

Polycam and Luma AI both run 3DGS processing from phone capture. Point the camera, walk the space, upload. The processing happens in the cloud and returns a splat asset within minutes.

The honest output is highly dependent on capture discipline. Move too fast, shoot in low light, or skip overlap passes, and the algorithm struggles with missing angular data. Where the camera cannot see, it creates occlusion shadows — blurry, stretched smears or blank holes in the geometry. But shoot carefully — consistent speed, good coverage, adequate light — and the result is a usable volumetric environment from hardware you already own.

What it is good for: location reference, previs stand-ins, background environments for short-form work, music video backdrops. What it is not: a replacement for a controlled capture session with a drone and a professional rig.

The Drone Step-Up

A consumer drone — DJI Mini series — extends coverage to aerials and wide spatial structure that handheld capture cannot reach. The step up from phone to drone is not a cliff. It is a gradient: better overlap, more consistent altitude, wider environment coverage.

The professional tier — a DJI Matrice carrying a Zenmuse P1 45MP full-frame sensor — is where capture discipline becomes capture engineering. Systematic flight plans, 80% image overlap, LiDAR integration. The output is a highly dense splat that can sustain closer camera work and more demanding deployment.

The gap between consumer and professional drone capture is real. It is not the gap between impossible and possible — it is the gap between a background asset and a foreground-capable environment.

The Janitor Work

The transition from a raw cloud render to a usable scene is where the indie filmmaker faces the true tactical bottleneck. Raw splats are incredibly messy. They inherit everything in the frame, producing floaters — stray floating point-cloud artefacts — and noisy, distorted backgrounds from across the street.

Cleaning this data requires a deliberate choice in the pipeline toolset.

The Web-Native Route — SuperSplat

For pure volumetric pipelines, browser-based tools like SuperSplat allow you to manipulate raw PLY files directly using WebGL or WebGPU. Using lasso, brush, and box selection tools, you can slice away background noise and delete floaters without ever converting the splat into standard 3D geometry.

The DCC Route — Blender

Traditional Digital Content Creation software like Blender becomes relevant if you convert your splat into a standard textured polygonal mesh via Polycam or Luma exports. Once converted, you can use Sculpt Mode to smooth jagged boundaries or Texture Paint to fix blurry patches. The trade-off is significant: converting a radiance field into a standard mesh strips away its view-dependent specular reflections — the way light realistically bounces off a wet surface or glints through a window based on camera angle.

For most indie workflows, the hybrid compromise is becoming standard: use SuperSplat to clean the raw volumetric background, then bring the scene into Blender solely to build low-poly invisible geometry — floor colliders, physics boundaries, or shadow catchers for digital characters.

The Browser Stack and the Optimisation Wall

A complete 3DGS workflow exists today that requires no local installation, no hardware beyond a phone and a laptop, and costs nothing to start:

• Capture — Polycam or Luma AI on a phone
• Process — cloud-based training pipelines
• Edit — SuperSplat in a browser tab: floater removal, boundary cleanup, basic colour work
• Deploy — Niantic Studio for browser-based scene building and distribution, or embed directly via Luma AI's iframe player

That is a virtual production pipeline. A minimal one — no relighting, no LED wall, no camera tracking — but a pipeline. An indie filmmaker can capture a location, clean it, and embed it in a distributable experience without touching a terminal.

However, the real gatekeeper is optimisation. A raw, pristine splat can easily weigh 50MB to 100MB or beyond. Deploy that raw file to the web and it will choke mobile VRAM, cause standalone VR headsets to stutter, or crash a browser tab entirely. The indie pipeline relies on the compression steps built into tools like SuperSplat — converting archival PLY files into compressed formats or streaming-friendly WebGPU scenes — to keep the experience distributable over standard 4G/5G connections. Optimisation is not an afterthought. It is the last production step.

Where the Gap Still Is

The tools are free. The ceiling is not.

Relighting costs money. Volinga's decomposition pipeline remains gated away from the consumer tier. A raw splat carries its lighting locked into its static radiance fields — the exact view-dependent specular reflections present at capture. The sunlight on the wall at 2pm, the shadow cast by a tree, the specular highlight on a wet road: all of it is baked in. Without a relighting pass, you are committed to those ambient conditions. For controlled studio work, that is a hard constraint. For documentary, music video, or immersive experience work where the authentic look of the location is the point, it may not matter.

LED volume access remains gated — not by technology, but by infrastructure cost. Running a splat on a browser player is accessible. Running it on a 20-metre LED wall with real-time camera tracking and a full stage crew is not. That gap has not closed.

Live events remain a professional domain. The Linkin Park Allianz Arena scan is a remarkable volumetric artefact — but the pipeline to produce and stream those fields in real time, at scale, for a touring show, is still enterprise territory.

The Creative Opportunity

None of that closes the door on independent work. It defines where the door is.

Short Film — Virtual Production Without the Volume

A 3DGS background environment captured on location replaces a greenscreen or a static photograph. The splice between a physical foreground actor and a volumetric background with depth is now within reach of a one-person production — no LED wall, no studio deal required.

Corridor Crew — THIS is the Biggest Thing Since CGI

See also: Convert Any Video to a 3D Model — drone capture to splat to After Effects compositing, no studio budget.

Music Video — Performer Against the World

A performer composited against a clean splat of a real environment — a forest, an arena, an abandoned building — without a greenscreen session or a location permit for a set that no longer needs to be present. Bilawal Sidhu's deep dive covers both ends of the budget spectrum: A$AP Rocky's volumetric shoots alongside the Superman 4DGS pipeline.

Bilawal Sidhu — Holographic Video is Finally Here. 4D Gaussian Splats Explained!

See also: How to Create an Interactive 3D Gaussian Splat From 1 Image — single-image to dynamic 3DGS asset, no enterprise pipeline.

Immersive Experience — Scan Once, Share Everywhere

A browser-embedded splat is a distributable artefact. A location, a building, a performance space — scanned once, optimised for web delivery, shared everywhere, and experienced spatially by anyone with a standard browser or headset.

Create Online Model Using XR Tour with 3DGS in D5 — packaging high-fidelity spatial scans for barrier-free browser delivery

For the VR end: VR-GS (UCLA / HKU) demonstrates real-time physics-aware interactive Gaussian Splatting inside headsets — touch, drag, and deform splat assets without crashing mobile VRAM boundaries.

The technology exists. Budget determines your position on the spectrum, not your access to it. Since the vast majority of the stack is now free, your entry point is the only thing dictated by finances. The constraint is no longer financial, it is entirely a matter of craft and execution.

This article is a snapshot. The stack it describes will have already shifted by the time you read this, bringing new tools, lower thresholds, and deeper integration. That is simply the nature of technology when it moves this fast.