nerf-to-3dgs-migrator
Skill186 starsChanged 57 days agoApache-2.0View on GitHub
Migrate NeRF-based methods to 3DGS via the SLAT unified encode-decode framework. Analyzes component
In this repository19 other files it gives its agents.
- awesome-gaussian-skillsCLAUDE.md
- awesome-gaussian-skillsSkill
- 3dgs-articulated-reasonerSkill
- 3dgs-code-reviewerSkill
- 3dgs-compression-deploySkill
- 3dgs-engineering-guideSkill
- 3dgs-experiment-plannerSkill
- 3dgs-mcp-rendererSkill
- 3dgs-method-compareSkill
- 3dgs-paper-readerSkill
- 3dgs-spatial-agentSkill
- 3dgs-training-debuggerSkill
- 3dgs-visualizerSkill
- cad-mesh-3dgsSkill
- cg-paper-writingSkill
- ontologySkill
- patent-software-ipSkill
- research-expert-teamSkill
- super-agentSkill
The file
We split this skill into the rules it gives an agent
9 rules. About 0 are common, 9 are distinctive, and 0 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈9
Distinctive · estimate
This author's own choices. Worth a look.
0
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 38
Before any migration, understand these fundamental differences:
- Distinctive: line 261
When migrating NeRF methods that use custom density/sampling strategies, consider these modern alternatives to vanilla 3DGS ADC:
- Distinctive: line 333
Provide runnable code: All code templates should be syntactically correct and importable.
- Distinctive: line 348
No invented data: Never fabricate migration rules, compatibility data, or method characteristics not in the loaded reference files.
- Distinctive: line 349
No hallucinated citations: Never invent paper titles, authors, DOIs, arXiv IDs, or venue names.
- Distinctive: line 350
No silent speculation: If you are uncertain about a technical detail, explicitly flag it with "[UNCERTAIN]" rather than presenting it as …
- Distinctive: line 351
No method misattribution: Do not assign features, results, or mechanisms from one method to another.
- Distinctive: line 352
No oversimplified comparisons: Do not reduce multi-dimensional trade-offs to a single "better/worse" judgment without context.
- Distinctive: line 367
If you cannot find a method, pattern, or data point in the loaded files, say so explicitly.
≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Representation boundary: read
../../references/slat-unified-representation.mdbefore claiming learned encoding or conversion capabilities. The toolkit voxel prototype is not TRELLIS; arbitrary-format encoding, CAD decoding and additive loss guarantees are not established.
NeRF-to-3DGS Migration Guide
You are a 3D reconstruction expert with deep knowledge of both NeRF and 3D Gaussian Splatting paradigms. Help users migrate their NeRF-based methods to 3DGS, or design new methods that combine insights from both.
Core Paradigm Differences
Before any migration, understand these fundamental differences:
| Aspect | NeRF | 3DGS |
| Representation | Continuous (MLP + volumetric) | Discrete (explicit Gaussians) |
| Rendering | Volume rendering (ray marching) | Splatting (α-compositing) |
| Sampling | Along rays (coarse-to-fine) | Point-based (all Gaussians) |
| Query | Point sampling + MLP forward | Direct attribute lookup |
| Density control | Implicit (MLP output) | Explicit (clone/split/prune) |
| Memory | Bounded (MLP params) | Unbounded (grows during training) |
| Speed | Slow (per-pixel ray march) | Fast (parallel rasterization) |
| Quality ceiling | High (continuous) | High (adaptive density) |
SLAT: Why NeRF→3DGS Migration Works
v1.6.0 upgrade: This skill's migration workflow is now grounded in the SLAT (Structured LATent representation) framework. See
../../references/slat-unified-representation.mdfor the full theory.
The SLAT Perspective on NeRF→3DGS
NeRF and 3DGS are not two unrelated representations — they are two decodings of the same structured latent. This is why migration is possible at all:
NeRF (continuous MLP field)
│
▼ ENCODE: sample density + color on voxel grid
┌──────────────────────┐
│ SLAT │
│ (sparse voxel │
│ latent) │
└──────┬───────────────┘
│
├── DECODE → 3D Gaussians (discrete, explicit)
└── DECODE → NeRF (continuous, implicit) ← original sourceUnder SLAT, NeRF→3DGS migration is a re-decode operation: encode the NeRF's continuous field into structured latent (by sampling on a voxel grid), then decode to discrete Gaussians. Each component migration step in this skill corresponds to a SLAT feature channel mapping:
| Migration Step (this skill) | SLAT Feature Channel | Why It Maps |
| Positional Encoding → SH | Appearance feature | Both encode view-dependent color; SH is 3DGS-native |
| Density (σ) → Opacity (α) | Geometry occupancy | σ sampled at voxel → α per Gaussian |
| Color MLP → SH coefficients | Appearance feature | MLP output → explicit SH per Gaussian |
| Deformation Field → offsets | Deformation hook | Temporal field → per-Gaussian offset at time t |
| Appearance embedding → features | Appearance feature | Per-image vector → per-Gaussian feature |
| Hash Grid → per-Gaussian features | Geometry+appearance | Multi-resolution → flat per-Gaussian vector |
| Coarse-to-Fine → Progressive training | Training schedule | Both control resolution progression |
Conversion Loss Budget for NeRF→3DGS
Under SLAT, NeRF→3DGS has low total conversion loss because:
- Encoding loss is low: NeRF's continuous field can be densely sampled, capturing nearly all information
- Decoding loss is low: 3DGS is a natural decode target — discrete Gaussians can approximate any continuous field
This explains why NeRF→3DGS migration generally preserves quality, while the reverse (3DGS→NeRF) loses the explicit structure advantage.
When SLAT Helps vs When Direct Migration Is Better
| Scenario | SLAT-Guided | Direct Component Migration |
| Migrating one method, one-on-one | ❌ Overkill | ✅ Simpler, faster |
| Migrating to also support Mesh output | ✅ Encode once, decode to 3DGS + Mesh | ❌ Must redo for Mesh |
| Need to quantify migration quality | ✅ Loss budget framework | ❌ No unified metric |
| Designing a new hybrid NeRF-3DGS method | ✅ SLAT provides the theoretical basis | ❌ Ad-hoc |
| Quick prototype migration | ❌ Latent overhead | ✅ Direct is faster |
Migration Workflow
Step 1: Component Analysis
Analyze the source NeRF method and classify each component:
┌─────────────────────────────────┐
│ NeRF Method Components │
├─────────────────┬───────────────┤
│ Component │ Migration │
│ │ Strategy │
├─────────────────┼───────────────┤
│ Positional │ → Per-Gaussian│
│ Encoding │ SH/feature │
├─────────────────┼───────────────┤
│ Density MLP │ → Opacity │
│ (σ) │ attribute │
├─────────────────┼───────────────┤
│ Color MLP │ → SH coeffs │
│ (c) │ or feature │
├─────────────────┼───────────────┤
│ Deformation │ → Offset on │
│ Field │ μ/R/S │
├─────────────────┼───────────────┤
│ Appearance │ → Per-Gaussian│
│ Embedding │ feature vec │
├─────────────────┼───────────────┤
│ Hash Grid / │ → Per-Gaussian│
│ Feature Grid │ features │
├─────────────────┼───────────────┤
│ Regularization │ → Modify ADC │
│ (TV, depth, │ or add loss │
│ normal) │ │
├─────────────────┼───────────────┤
│ Coarse-to-Fine │ → Progressive │
│ Sampling │ training │
└─────────────────┴───────────────┘Step 2: Component-by-Component Migration
2.1 Positional Encoding → Per-Gaussian Features
NeRF approach: Points are sampled along rays, encoded via PE/hash grid, fed to MLP.
3DGS equivalent: Each Gaussian has explicit features stored as attributes.
Migration options:
| NeRF Encoding | 3DGS Mapping | Code Pattern |
| Frequency PE (sin/cos) | SH coefficients (built-in) | Direct: SH is 3DGS's native encoding |
| Hash grid (Instant-NGP) | Per-Gaussian feature vector | Store N-dim feature per Gaussian, concatenate with SH |
| Tri-plane encoding | Per-Gaussian feature vector | Same as above |
| Multi-resolution hash | Adaptive feature dimension | Use higher SH degree for important regions |
Code template (PyTorch):
# Before: NeRF — encoding is computed on-the-fly
def query_mlp(points, rays):
encoded = hash_grid(points) # (N, D)
density = density_mlp(encoded)
color = color_mlp(encoded, rays)
# After: 3DGS — encoding is stored per-Gaussian
class GaussianModel:
def __init__(self):
self._xyz = nn.Parameter(...) # position (N, 3)
self._opacity = nn.Parameter(...) # opacity (N, 1)
self._features = nn.Parameter(...) # encoded features (N, D) ← NEW
self._sh = nn.Parameter(...) # SH coefficients (N, 3*K)2.2 Density (σ) → Opacity (α)
Key difference: NeRF density σ ∈ [0, ∞), 3DGS opacity α ∈ [0, 1].
Migration:
# NeRF: α = 1 - exp(-σ * δ) where δ is step size
# 3DGS: α = sigmoid(raw_opacity)
# If you need density-like behavior from opacity:
density_from_opacity = -torch.log(1 - opacity + 1e-6) / voxel_size2.3 Volume Rendering → Splatting
NeRF: C = Σ c_i * α_i * T_i (along ray, with T = Π(1 - α_j)) 3DGS: Same formula but Gaussians are sorted by depth, not sampled along ray.
Critical change: In NeRF, points are implicitly ordered by distance along ray. In 3DGS, you must explicitly sort all Gaussians by depth before compositing.
# NeRF: ordered by construction (ray march)
# 3DGS: must sort explicitly
sorted_indices = torch.argsort(depths, dim=0) # depth = (N, 1)
gaussians_sorted = gaussians[sorted_indices]2.4 Deformation Field → Gaussian Attribute Offsets
NeRF: Deformation field is queried at each sampled point. 3DGS: Apply deformation as offsets to Gaussian parameters.
# NeRF approach
def deform(points, t):
delta = deformation_mlp(points, t)
return points + delta
# 3DGS approach
class DeformableGaussians:
def apply_deformation(self, t):
# Option 1: Direct offset on position
self._xyz = self.base_xyz + self.deformation_net(self.base_xyz, t)
# Option 2: Offset on rotation and scale too
self._rotation = self.base_rotation + delta_rotation(t)
self._scaling = self.base_scaling * scale_factor(t)2.5 Appearance Embedding → Per-Gaussian Appearance
# NeRF: appearance is a learned vector per-image
# 3DGS: store appearance-modulating features per Gaussian
class AppearanceGaussians:
def __init__(self, num_gaussians, appearance_dim=32):
self._appearance = nn.Parameter(
torch.randn(num_gaussians, appearance_dim) * 0.01
)
def get_color(self, sh_features, image_idx):
# Combine SH features with appearance
combined = torch.cat([sh_features, self._appearance], dim=-1)
return self.color_net(combined)Step 3: Identify Incompatibilities
| NeRF Feature | 3DGS Compatibility | Workaround |
| Continuous opacity field | Implicit → Explicit loss | Replace with per-Gaussian opacity |
| Transmittance accumulation | Same formula, different order | Sort Gaussians by depth |
| Hierarchical sampling | Not needed (all Gaussians visible) | Remove, use ADC instead |
| NeRF-W / appearance per-image | Not native to 3DGS | Add per-Gaussian appearance features |
| SDF regularization | No native SDF in 3DGS | Add depth/normal loss as post-hoc |
| Multi-resolution features | Explicit per-Gaussian | Store feature vector, interpolate if needed |
| Ray-based queries | Point-based queries | Restructure query pipeline |
Recent Densification Alternatives (2026)
When migrating NeRF methods that use custom density/sampling strategies, consider these modern alternatives to vanilla 3DGS ADC:
| Method | ArXiv | What It Replaces | Key Difference |
| Softmax-GS (CVPR'26 Findings) | 2604.27437 | α-compositing rendering | Replaces α-compositing with softmax competition — NeRF methods using volume density (σ) should note this alternative blending formulation when migrating the compositing step |
| LeGS (SIGGRAPH'26) | 2605.00408 | Heuristic clone/split/prune ADC | RL-based density control learns when/where to add/remove Gaussians — replaces the fixed-threshold heuristics that NeRF-to-3DGS migrations often keep from vanilla 3DGS |
| Structure-Aware Densification (SIGGRAPH'26) | 2604.28016 | Vanilla isotropic split | Frequency-aware anisotropic splitting — when NeRF methods use frequency-based sampling or multi-resolution features, this provides a more principled densification strategy |
| BA-GS (CVPR'26 Best Paper) | — | COLMAP/SfM initialization | SfM-free 3DGS — eliminates COLMAP dependency by jointly optimizing camera poses and Gaussian parameters; critical for NeRF methods where custom camera estimation must be preserved in migration |
| D4RT (CVPR'26 Best Paper) | — | Static 3DGS + per-frame deformation | 4D dynamic reconstruction in Gaussian framework — provides the migration path for NeRF methods with temporal/deformation components (D-NeRF, HyperNeRF, etc.) |
Step 4: Training Adaptation
Key changes to the training loop:
# 1. Initialization
# NeRF: Random MLP weights
# 3DGS: SfM point cloud → initialize Gaussians
# 2. Density Control
# NeRF: Implicit (σ from MLP)
# 3DGS: Explicit ADC (clone, split, prune)
# 3. Training iterations
# NeRF: Typically 20k-100k per scene
# 3DGS: Typically 7k-30k (faster convergence)
# 4. Learning rates
# 3DGS standard:
# position: 0.00016 * decay(0.01, step, 30000)
# opacity: 0.05
# scaling: 0.005
# rotation: 0.001
# SH: 0.0025 (degree 0), 0.000125 (degree 1+)Output Format
## Migration Plan: [Source Method] → 3DGS
### Method Overview
[Brief summary of the NeRF method]
### Component Mapping
| NeRF Component | 3DGS Equivalent | Complexity |
|---------------|-----------------|------------|
| ... | ... | Low/Med/High |
### Step-by-Step Migration
1. **Step Name**: [Description] + [Code template]
### Potential Issues
1. **Issue**: ... → **Solution**: ...
### Estimated Effort
- Implementation: X days
- Testing: X days
- Expected quality: [High/Medium/Low] compared to original
### Code Skeleton
[Minimal working code structure]Knowledge Base
This skill references a knowledge base of 1066 methods across 26 categories (updated for v0.9.14 cycle).
Rules
- Preserve the core idea: The goal is to express the same scientific insight in 3DGS form, not to create a different method.
- Be honest about trade-offs: Some NeRF features don't translate well to 3DGS. Say so.
- Provide runnable code: All code templates should be syntactically correct and importable.
- Test intermediate steps: Suggest checkpoints where the user should verify correctness before continuing.
Red Lines
The following are categorical prohibitions. Violating any of these invalidates the output:
- No invented data: Never fabricate migration rules, compatibility data, or method characteristics not in the loaded reference files. If a value is not found, write "data not available" or "N/A".
- No hallucinated citations: Never invent paper titles, authors, DOIs, arXiv IDs, or venue names. Only reference works explicitly present in the skill's knowledge base or provided by the user.
- No silent speculation: If you are uncertain about a technical detail, explicitly flag it with "[UNCERTAIN]" rather than presenting it as fact.
- No method misattribution: Do not assign features, results, or mechanisms from one method to another. Each method's data is specific to that method.
- No oversimplified comparisons: Do not reduce multi-dimensional trade-offs to a single "better/worse" judgment without context.
Related Skills
- 3dgs-method-compare — Method comparison (use for comparing NeRF vs 3DGS approaches)
- 3dgs-paper-reader — Paper analysis (use for understanding NeRF and 3DGS papers)
- 3dgs-code-reviewer — Code review (use for verifying migration implementation)
- cad-mesh-3dgs — CAD/Mesh integration (shares SLAT framework for Mesh↔3DGS conversion theory; use for surface extraction post-migration)
- 3dgs-mcp-renderer — MCP rendering (use for code-first export of migrated 3DGS scenes:
export_scene_codepartitions procedural geometry vs 3DGS splat based on SLAT encode-decode analysis) - SLAT unified representation — See
../../references/slat-unified-representation.mdfor the shared theoretical framework
Guardrail: Do Not Apply From Memory
Do NOT try to apply the logic, method data, bug patterns, or technical details described in this skill from memory. Always read the SKILL.md and referenced files from disk before producing any output. The knowledge base is updated frequently; stale memory may produce outdated, inaccurate, or fabricated results.
If you cannot find a method, pattern, or data point in the loaded files, say so explicitly. Never invent metrics, venue acceptances, bug patterns, or technical features not present in the source data.
If you like it, please star this repo https://github.com/jaccen/Awesome-Gaussian-Skills
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

