NeurCo
Neural Rendering Compiler

Neural Rendering Compiler: Generative Video Models as Programmable Renderers

NeurCo probes frozen video models to build model-specific capability manuals, then compiles user intent into execution strategies that keep reliable requirements and adapt the rest.

  • Author One*
  • Author Two*
  • Author Three
  • Author Four
  • Author Five
Institution Placeholder  ·  *Equal contribution
380 VComBench compositional tasks across T2V / I2V / R2V
68.5→69.7% Seedance 2.5 Req. with NeurCo (T)
46.8→52.3% Wan requirement satisfaction with NeurCo
66.5→69.3% H3 requirement satisfaction with NeurCo
Key features

From implicit model ability to an executable control interface

Instead of assuming a fixed instruction set, NeurCo measures what a generative renderer can reliably do, then compiles requests against that measured interface.

Profiling

Offline capability manuals

Probe semantic controls under T2V, I2V, and R2V to estimate which requirements a frozen backend can satisfy.

IR

Model-agnostic programs

Represent intent as objects, constraints, events, and cameras, independent of any single video model.

Compile

Capability-guided lowering

Keep reliable requirements; rewrite, reroute, or invoke external operators when native execution is weak.

Benchmark

VComBench

380 multi-requirement tasks with requirement-level scoring for compositional controllable generation.

Abstract
Generative video models are becoming powerful renderers, yet lack the predictable control interfaces of conventional systems such as HTML/CSS engines or Blender. Their capabilities emerge implicitly from training, making it difficult to determine what they can reliably execute. We present Neural Rendering Compiler (NeurCo), which automatically probes frozen video models to construct model-specific capability manuals describing which semantic requirements they can reliably satisfy. NeurCo represents user intent in a model-agnostic intermediate representation of objects, constraints, events, and cameras, then compiles it into model-specific execution strategies. Reliable requirements remain unchanged, while unreliable ones are selectively rewritten or compiled into alternative execution strategies. We introduce VComBench, a benchmark of 380 compositional, multi-requirement tasks spanning text-to-video, image-to-video, and reference-to-video generation. NeurCo improves requirement satisfaction from 46.8% to 52.3% on Wan and from 66.5% to 69.3% on MiniMax H3, and achieves further gains when combined with prompt enhancers, including on Seedance 2.5. Our analysis reveals model-specific differences in control capabilities and trade-offs between controllability and photorealism, highlighting capability-guided compilation as a promising approach to programmable generative video rendering.
Method

Profile once. Compile per request.

NeurCo separates a shared semantic vocabulary from model-specific manuals, then lowers each request into backend-aware execution strategies.

01

Probe

Isolate controls at graded difficulties under applicable input modes.

02

Manual

Build a capability manual of reliable settings and preferred realizations.

03

Compile

Parse IR, preserve strong requirements, adapt weak ones with operators.

04

Execute

Run text, first-frame, reference, proxy, or shot-concatenation strategies.

Results

VComBench main table

Requirement satisfaction (Req.) and all-requirements success (All) on H3, Wan, and Seedance 2.5. NeurCo (T) is text-only compilation; NeurCo (T+O) adds external operators. VBench-2.0 testmini and Sci-VBench testmini are released as data, not shown as tables.

Method H3 Wan Seedance 2.5
Req.All Req.All Req.All
Native inputs
Native 66.528.2 46.88.4 68.526.9
NeurCo (T) 67.727.9 47.98.4 69.727.5
NeurCo (T+O) 69.325.3 52.311.8 68.225.8
With external PE
External PE 72.533.2 50.612.7 71.331.1
External PE + NeurCo (O) 73.033.7 52.514.5 72.132.2
External PE + NeurCo (T) 72.432.7 46.39.0 71.732.1
External PE + NeurCo (T+O) 72.034.5 50.012.3 71.831.5

On Seedance 2.5, text-only compilation is strongest on native inputs (68.5→69.7 Req.), while operator-augmented NeurCo helps more after external PE. Qualitative videos currently cover H3, Wan, and Seedance 2.5.

Citation

BibTeX

@article{neurco2026,
  title   = {Neural Rendering Compiler: Generative Video Models as Programmable Renderers},
  author  = {Author One and Author Two and Author Three and Author Four and Author Five},
  journal = {arXiv preprint},
  year    = {2026}
}