ComfyUI-LATO2 Introduction
ComfyUI-LATO2 turns dense 3D models into new low-poly meshes directly inside ComfyUI. It uses LATO.2 to rebuild an existing shape with approximately 200–5,000 vertices, giving you control over how much geometry the result contains.
A mesh is the network of points, edges, and faces that forms a 3D object. AI-generated models and scans often contain far more geometry than you need. ComfyUI-LATO2 helps you turn those dense surfaces into simpler meshes with edges arranged around the object’s form.
You can use it with:
- Meshes generated by TRELLIS.2 or Hunyuan3D.
- Scanned models.
- Existing mesh files in supported formats. Unlike a conventional decimator, which simplifies the triangles already present, LATO.2 creates a new arrangement of vertices and connections. This makes it useful when you want a more deliberate low-poly appearance rather than simply fewer triangles.
Additional remeshing examples showing a car, fire hydrant, character, and market stall
The dense source models appear in the top rows, with the remeshed results below. These examples use 1,200 vertices, except the treasure chest and market stall, which use 2,500.
The extension also handles several practical cleanup tasks: removing hidden inner walls, rebuilding predicted faces into a closed surface, and restoring the original model’s position and size.
Important: The result is geometry only. Textures, UVs, and materials are not transferred. Thin surfaces and narrow gaps can still produce artifacts, so inspect the result before using it in a production asset.
How ComfyUI-LATO2 Works
Think of remeshing as rebuilding a sculpture with a new framework. Rather than preserving every tiny triangle, the extension decides where a smaller set of points should go and how those points should connect.
The process has four main stages.
1. Remove hidden inner walls
Some generated meshes—particularly TRELLIS.2 meshes—contain a second surface inside the visible outer wall. If left in place, LATO.2 may spend vertices rebuilding both layers.
With remove_inner_walls enabled, the extension:
- Represents the object on a 3D grid of small cells, called voxels.
- Fills empty space starting from outside the object.
- Removes faces that this outside space cannot reach. Imagine pouring water around a sealed sculpture: the water touches the exterior but cannot reach enclosed surfaces inside it. This provides a useful way to distinguish the visible shell from hidden geometry.
The cleanup also closes hairline gaps before the fill and performs a visibility check to retain faces visible from outside.
2. Place vertices with LATO.2
The shape is temporarily moved into a standardized size and position for processing.
LATO.2 renders a gray view of the object and uses DINOv2, a visual feature model, to extract information from that view. Its vertex flow, or V-Flow, then places approximately the number of vertices you requested.
For example, a barrel remeshed at 600 vertices has a much smaller geometry budget than the same barrel remeshed at 3,000 vertices.
3. Connect the vertices and rebuild the surface
LATO.2’s topology flow, or T-Flow, predicts which vertices should connect. “Topology” means the arrangement of those connections and faces.
The raw predictions can contain overlapping faces, holes, and faces pointing in inconsistent directions. With rebuild_surface enabled, ComfyUI-LATO2 reconstructs the faces from the predicted connections. It prioritizes likely triangles, limits each edge to at most two faces, fills remaining holes, and orients faces outward.
This rebuilding improves the surface, but it does not guarantee that every result is free from folds or self-intersections.
4. Restore the original position and size
The finished mesh is returned to the source model’s position and scale and written as a GLB, a binary glTF 3D file.
You also receive a native ComfyUI MESH output, allowing you to continue processing the geometry in your workflow.
ComfyUI-LATO2 Features
Controllable low-poly density
The vertex_count setting accepts values from 200 to 5,000. The finished mesh contains approximately this many vertices—not necessarily an exact count.
A vertex is a point in the mesh. Vertex count is not the same as triangle or face count.
| Vertex count | Suggested use | |---|---| | 600 | Explore a strongly simplified version, as shown in the comparison workflow. | | 1,000–1,500 | A useful starting range for the clean low-poly appearance shown in the prop, vehicle, and character examples. | | 2,500 and above | Retain more small features, such as coins, bolts, or chains. |
More vertices give the model additional room to describe detail, but do not guarantee that every narrow gap or delicate part will survive.
The same object remeshed at 600, 1,500, and 3,000 vertices
Alternative layouts through seeds
LATO.2 samples its vertex placement and connections. Changing seed gives you a different arrangement for the same source shape.
For an important asset, try two or three seeds and compare:
- The overall silhouette.
- The placement of edges around recognizable features.
- Creases or folds on thin parts.
- Whether narrow openings remain visible. Keep the vertex count unchanged while comparing seeds so you can judge the layout rather than the amount of geometry.
Hidden-surface cleanup
remove_inner_walls is enabled by default. It removes unreachable interior faces before remeshing, helping avoid unnecessary geometry inside the result.
This is especially relevant for TRELLIS.2 models with hidden inner shells. Disabling it skips that cleanup stage.
Closed-surface reconstruction
rebuild_surface is enabled by default. It replaces LATO.2’s raw face construction with the extension’s rebuilt surface.
Leave it enabled for normal use. Turning it off exposes the raw faces, which can contain holes, overlaps, and inconsistent face directions.
A closed surface should not be confused with a guaranteed print-ready surface: thin regions can still cross or fold.
Native ComfyUI connections
The node accepts both native MESH data and supported 3D files. It returns outputs suitable for previewing, saving, and further mesh processing.
This lets you place remeshing directly after a mesh-generation node instead of manually exporting and reloading the model.
Windows and Linux support
The extension supports Windows and Linux without requiring you to compile CUDA extensions.
Where certain upstream packages cannot be imported, it provides portable replacements for voxel processing, scatter operations, attention, and rendering. Its alternative renderer also avoids the black-image problem associated with some display-less rendering environments.
An NVIDIA GPU with CUDA and 8 GB of VRAM is required. Portable replacements do not mean that the remeshing pipeline supports CPU-only operation.
Using ComfyUI-LATO2
A simple first workflow
Once the extension is available in ComfyUI:
- Open Browse Templates > Extensions > ComfyUI-LATO2, or drag a workflow from
example_workflows/onto the canvas. - Add your mesh with upload mesh, or connect a mesh-producing node.
- Start with a
vertex_countaround 1,000–1,500. - Leave
remove_inner_wallsandrebuild_surfaceenabled. - Connect
model_3dto Preview 3D. - Run the workflow and inspect the result from several angles.
- If needed, try another seed or increase the vertex count. Example ComfyUI workflow using LATO.2 Low-Poly Remesh, Preview 3D, and Save 3D Model
Mesh inputs and file selection
The LATO.2 Low-Poly Remesh node provides three ways to select a source.
| Input | What it does |
|---|---|
| mesh | Accepts a connected native MESH or 3D file. Takes priority over both file-selection fields below. |
| input_path | Accepts an absolute mesh path, or a path relative to ComfyUI’s input or output folder. Takes priority over model_file. |
| model_file | Selects a mesh from ComfyUI’s input folder, including input/3d. Use upload mesh to add a file. |
Supported file extensions are .glb, .gltf, .obj, .ply, and .stl.
Selection priority: connected mesh → input_path → model_file.
If the node processes an unexpected object, check whether a higher-priority input is overriding your selection.
Main and advanced controls
| Setting | Explanation |
|---|---|
| vertex_count | Target number of vertices, from 200 to 5,000. Controls the overall geometry budget. |
| seed | Changes the sampled arrangement of vertices and edges. |
| filename_prefix | Sets the name prefix for the GLB saved in output/lato2. |
| vflow_steps | Advanced control for the number of vertex-flow processing steps. |
| tflow_steps | Advanced control for the number of topology-flow processing steps. |
| cfg | Advanced control for how strongly the vertex flow follows the input’s rendered view. |
| remove_inner_walls | Removes enclosed, unreachable faces before remeshing. Enabled by default. |
| rebuild_surface | Reconstructs the predicted faces into a closed surface. Enabled by default. |
For your first comparisons, focus on vertex count and seed. Leave the flow-step and guidance controls at their provided values until you have a baseline result.
Outputs and saved files
| Output | How you use it |
|---|---|
| mesh | Geometry-only native MESH, suitable for Save 3D Model, Get Mesh Info, and other mesh nodes. |
| model_3d | GLB represented as a 3D file, suitable for Preview 3D and Save 3D. |
| glb_path | The GLB file path relative to ComfyUI’s output folder. |
The extension also saves the GLB automatically in output/lato2/.
Included example workflows
The examples use the included hydrant.glb. Add it through upload mesh when needed.
| Workflow | What you can learn |
|---|---|
| lato2_remesh_a_mesh.json | Remesh a file, preview the result, and connect it to Save 3D Model. |
| lato2_vertex_counts.json | Compare 600, 1,500, and 3,000 vertices side by side. |
| lato2_from_load_3d.json | Feed the 3D file output of Load 3D into the remesher. |
| lato2_after_trellis2.json | Generate a mesh from an image with ComfyUI’s native TRELLIS.2 nodes, then remesh it. Requires the TRELLIS.2 models. |
ComfyUI-LATO2 Models
The documented extension uses LATO.2 with DINOv2 visual conditioning. These are complementary components, not alternative artistic presets.
| Component | Role | |---|---| | LATO.2 | Generates the new vertex positions and predicts their connections through V-Flow and T-Flow. | | DINOv2 | Extracts visual information from the gray rendered view used to guide LATO.2. | The documented downloads are approximately 3.4 GB for LATO.2 weights and 1.1 GB for DINOv2.
The upstream LATO.2 project includes separate inference tools for its individual stages and demonstrates capabilities such as part-wise generation and topology-adaptive editing. These should not be confused with controls exposed by this ComfyUI remeshing node.
Likewise, the wider DINOv2 project offers several model variants and research applications, but the provided extension documentation does not describe a user-facing selector for switching among them.
For artistic variation in this node, use vertex count and seed rather than treating these components as interchangeable models.
ComfyUI-LATO2 Limitations and Quality Checks
Before texturing or exporting an important asset, inspect the silhouette, thin features, openings, and surface shading.
- Thin parts can fold or intersect. Slats, blades, and cloth edges are difficult cases because vertices may be placed on both sides of a very thin surface.
- Narrow slots can close. If no vertices are placed inside a gap, the new surface may bridge it.
- Closed does not mean self-intersection-free. Outward-facing surfaces can still contain local folds.
- 3D printing and Boolean operations need extra care. These tasks may require stricter surface consistency than the result provides.
- Textures, UVs, and materials are not transferred. The extension creates new geometry.
- Only one mesh is processed at a time. If you supply a
MESHbatch, the node uses its first mesh. - Results can vary across hardware. The same source and seed may produce a slightly different layout on another machine or GPU. After remeshing, you can texture the new mesh—for example with StableGen in Blender—or bake the source model’s texture onto it as a separate step.
Troubleshooting ComfyUI-LATO2
“LATO.2 is not set up”
- Find LATO.2 Setup Check in the
3d/LATO.2category. - Run it and read the missing-component report.
- Consult the extension repository for setup guidance.
- Restart ComfyUI after the missing components have been addressed.
- Run the check again; it should report ready.
“CUDA out of memory”
LATO.2 peaks at approximately 7 GB of VRAM, so an 8 GB GPU may have little room for other loaded models.
- Use ComfyUI’s Free model and node cache button.
- Unload other models before remeshing.
- Run the remeshing workflow again. The extension starts a fresh process for each run and does not retain its model VRAM between runs.
A run seems slow
The author reports approximately 30–60 seconds per run on an RTX 3090 or RTX 4080, with a substantial part of that time spent loading weights.
Removing inner walls from a very dense input—around a million faces—can add one to three minutes of CPU processing.
- Check ComfyUI’s console for messages prefixed with
[LATO.2]. - Allow additional time for dense-input cleanup.
- Remember that weights are loaded again for each run. These timings are reference measurements, not guaranteed completion times.
The result loses small details or closes a gap
- Increase
vertex_count; 2,500 or more is the suggested range for small details. - Try two or three different seeds.
- Compare the affected area from several angles. Narrow slots remain a known limitation. More vertices can help, but do not guarantee preservation.
Thin parts have creases or folded surfaces
- Confirm that
rebuild_surfaceis enabled. - Try another seed and compare the affected region.
- Inspect the geometry before using it for printing or Boolean operations. Surface rebuilding addresses raw face-construction problems, but cannot guarantee the removal of thin-part folds.
The output has unexpected holes or overlapping faces
- Check whether
rebuild_surfacewas disabled. - Enable it and run the node again.
- Compare the rebuilt result with the raw output. Use the surface-rebuilding comparison in How ComfyUI-LATO2 Works as a visual reference.
The wrong source mesh is being processed
- Check for a connected
meshinput. - If there is no connected source, check
input_path. - Use
model_fileonly when neither higher-priority source is active.
Frequently asked questions
Does this generate an object directly from a text prompt or image? This node remeshes an existing 3D object. The included TRELLIS.2 workflow generates the initial mesh from an image before passing it to LATO.2.
Does it preserve my original textures? No. The output contains geometry only. Texturing or texture baking must happen separately.
Will I get exactly the requested vertex count? The result contains approximately the requested number of vertices.
Is the output a quad mesh? The documented rebuilding process constructs triangular faces. It is not described as a quad-retopology tool.
Can I process a batch of meshes at once?
A supplied MESH batch uses only its first mesh.
Is a closed result automatically ready for 3D printing? No. Folds and self-intersections can still require inspection and repair.
Why does the same seed look slightly different on another GPU? LATO.2 samples its result, and machine or GPU differences can affect the final layout.
Learn More about ComfyUI-LATO2
Practical resources
- ComfyUI-LATO2 repository: The extension’s main documentation and source.
- Included example workflows: The most direct starting point for file remeshing, vertex-count comparisons, and TRELLIS.2 integration.
- LATO.2 project page: Visual examples and an overview of the underlying method.
- LATO.2 interactive demo: Explore the upstream model separately from this extension.
Texturing and broader asset workflows
The author first built this node as part of StableGen, a Blender add-on for generating, texturing, and rigging 3D assets. Its broader features are separate from ComfyUI-LATO2.
- StableGen quick start: An introduction to its Blender asset workflow.
- Generation modes: Learn about generating and projecting texture views.
- TRELLIS.2 mesh generation: Background on a source of meshes used in the examples.
- StableGen Discussions: Community discussions for the related Blender workflow.
Technical background
- LATO.2 paper: The research behind vertex flow and topology flow.
- LATO.2 model repository: Upstream pretrained weights.
- DINOv2 repository: Background on the visual feature model used for conditioning.
ComfyUI-LATO2 Credits and License
ComfyUI-LATO2 builds on LATO.2, DINOv2, and cleanup work from StableGen.
- Extension code: GPL-3.0; see LICENSE.
- LATO.2: MIT, as identified in the extension documentation. Its code and weights are downloaded separately rather than bundled with the extension.
- DINOv2: Apache-2.0 for the code and model weights used by the documented dependency.
- Portable voxel implementation: Follows TRELLIS.2’s MIT-licensed
o-voxel; see THIRD_PARTY_NOTICES.md. The extension and its model dependencies retain their respective licenses. For research use, consult the upstream LATO.2 paper and citation information.
