Save 4 hours! We auto-setup your workflow! Free!

Drop your workflow.json — we handle every dependency, custom node, and model. Just open the link and run.

Auto-Setup Workflow Json (Free) Now!
ComfyUI > Nodes > ComfyUI-SplatKit-World

ComfyUI Extension: ComfyUI-SplatKit-World

Repo Name

ComfyUI-SplatKit-World

Author
mickmumpitz (Account age: 476 days)
Nodes
View all nodes(1)
Latest Updated
2026-11-05
Github Stars
0.04K

How to Install ComfyUI-SplatKit-World

Install this extension via the ComfyUI Manager by searching for ComfyUI-SplatKit-World
  • 1. Click the Manager button in the main menu
  • 2. Select Custom Nodes Manager button
  • 3. Enter ComfyUI-SplatKit-World in the search bar
After installation, click the Restart button to restart ComfyUI. Then, manually refresh your browser to clear the cache and access the updated list of nodes.

Visit ComfyUI Online for ready-to-use ComfyUI environment

  • Free trial available
  • 16GB VRAM to 80GB VRAM GPU machines
  • 400+ preloaded models/nodes
  • Freedom to upload custom models/nodes
  • 200+ ready-to-run workflows
  • 100% private workspace with up to 200GB storage
  • Dedicated Support

Run ComfyUI Online

ComfyUI-SplatKit-World Description

ComfyUI-SplatKit-World guides image, text, or panorama inputs through camera planning, video generation, 3D reconstruction, and Gaussian splat training to create explorable environments.

ComfyUI-SplatKit-World Introduction

Turn one image—or one sentence—into a 3D environment you can explore in your browser.

ComfyUI-SplatKit-World is a guided app inside ComfyUI, built on ComfyUI-SplatKit. It brings panorama generation, camera-path planning, AI fly-through generation, 3D reconstruction, and Gaussian splat training into one step-by-step workspace.

You can start with:

  • A photo: Use an existing image as the starting point for a surrounding environment.
  • A text prompt: Describe a place you want to create.
  • A 360° panorama: Bring an existing panoramic image in a 2:1 format. The app guides you toward an explorable Gaussian splat: a 3D representation made from many small, soft, colored elements rather than a conventional surface mesh. Together, these elements recreate the appearance of a scene from different viewpoints.

For AI artists, this provides a way to move beyond a single flat composition. You can explore an environment concept spatially, plan camera movement through it, and download the resulting .ply splat for use with compatible tools.

What makes it useful?

  • A guided interface: Follow six stages instead of assembling the entire process manually.
  • Creative control over movement: Shape and preview camera paths before approving the expensive fly-through stage.
  • Help with unseen areas: AI generates plausible content where the original image provides no information.
  • Integrated training and viewing: Continue from the generated dataset to a trained splat and a browser viewer.
  • Persistent projects: Your worlds remain available after a ComfyUI restart.
  • Local processing with open weights: No cloud APIs or partner nodes are required.

Set your expectations: This is generative reconstruction, not an exact scan. A single image cannot reveal what is behind objects or outside its frame. Those areas must be estimated or invented, so the result should be treated as an AI-created environment rather than a verified copy of a real place.

How ComfyUI-SplatKit-World Works

The overall process is:

text Photo, prompt, or panorama ↓ 360° panorama ↓ MoGe depth estimate ↓ Camera paths you approve ↓ WAN-generated fly-throughs ↓ SphereSfM reconstruction ↓ Spirula Studio splat training ↓ Browser exploration and .ply download

1. Build a picture that surrounds you

An ordinary photo is like looking through a window: it shows only one direction. A 360° panorama is more like a picture wrapped around the inside of a sphere.

For photo and text inputs, Krea 2 and a specialized 360° LoRA help create this surrounding image. A LoRA is a smaller model add-on that guides a larger model toward a particular task—in this case, panoramic imagery.

The app also accepts an existing 2:1 panorama. This format is twice as wide as it is tall and represents a full spherical view laid out as a rectangle.

2. Estimate which surfaces are near and far

MoGe estimates depth from the panorama. Think of it as giving the flat picture a rough sense of distance: a nearby chair should sit in front of a distant wall.

This depth information becomes the starting geometry used to preview movement through the scene. It is an estimate from imagery, not a direct physical measurement.

3. Choose how the camera moves

You shape camera paths using floor and side views, then inspect a synchronized movement preview.

As the camera moves away from the original viewpoint, it reveals gaps. For example, moving sideways past a pillar exposes a section of wall that the original panorama could not see.

The Preview camera view shows these missing areas in black. At this stage, black means content that still needs to be generated, not necessarily an error.

4. Generate the missing views

Wan 2.1, guided by the rendered movement and panoramic LoRAs, generates fly-through imagery that fills the newly exposed areas.

This approach comes from Matrix-3D: move a camera through estimated geometry, mark the unknown regions, and use a video model to paint plausible content consistently across frames.

SplatKit's HiRes Composite stage then reprojects detail from the original high-resolution panorama onto the generated frames. In simple terms, it tries to keep the original image's detail where that information is available, rather than relying on the video model to repaint everything.

5. Reconstruct a training dataset

SphereSfM compares the generated views and finds matching visual features. It uses those matches to estimate camera positions and a sparse set of 3D points.

This is called structure from motion: recovering a scene's structure by comparing images taken from different positions.

The result is a COLMAP-format dataset, a standard arrangement of images, camera information, and reconstruction data that splat trainers can use.

6. Train and explore the splat

Spirula Studio trains a Gaussian splat from the dataset.

Training repeatedly adjusts the splat's elements so that rendered views resemble the training images. The finished scene is displayed in the browser using Spark, and you can download its .ply file.

ComfyUI-SplatKit-World Features

Six-stage guided workspace

Open SplatKit World from ComfyUI's top bar or the SplatKit menu. You can also open:

text <your ComfyUI address>/splatkit/world/

For a local ComfyUI session using the usual address, this is:

text http://127.0.0.1:8188/splatkit/world/

The left side of the app guides you through the following stages:

| Stage | What you do | What the app does | |---|---|---| | 01 Input | Drop a photo, describe a place, or provide a 2:1 panorama. | Establishes the starting material for your world. | | 02 Panorama | Inspect the surrounding image with View 360°. | Generates the panorama where needed and estimates its depth with MoGe. | | 03 Camera paths | Edit, preview, and approve camera movement. | Shows the planned movement and the imagery that will guide WAN. | | 04 Fly-throughs | Continue with the approved paths. | Generates missing views, applies HiRes Composite, and runs dual-resolution SphereSfM to produce COLMAP data. | | 05 Splat training | Choose a Preview or Full training run. | Runs Spirula Studio to create the splat. | | 06 Your world | Explore the scene and download the result. | Displays the trained world in the browser. |

Interactive camera-path editing

The path editor gives you control over where the virtual camera travels.

  • Floor view helps you inspect movement across the scene.
  • Side view helps you inspect changes in height.
  • Synchronized previews let you compare the path with the camera's movement.
  • Preview camera shows the control imagery that WAN will receive.
  • Approval before generation gives you a chance to correct the route before the expensive fly-through stage begins. For example, a low path through a courtyard emphasizes a ground-level experience, while a rising path reveals more of the surrounding environment. Moving behind nearby objects exposes more unseen surfaces and therefore asks the AI to invent more content.

Practical approach: Inspect the route for movement through walls or other surfaces before approving it. The estimated geometry is useful guidance, but it is not a guarantee of a physically accurate scene.

Panorama inspection

Use View 360° to inspect the panorama before committing to later stages.

This is an important creative checkpoint: the panorama supplies much of the world's appearance. Look around the full image, not just the original photo's direction, and check whether the generated surroundings match your intended setting.

High-resolution detail preservation

The integrated dataset workflow includes HiRes Composite.

Rather than using only the detail produced by WAN, this stage brings information from the original panorama back into the fly-through frames wherever the geometry supports it. WAN supplies content where the original view has no answer.

This is especially relevant when your panorama contains fine textures that you want to preserve. For a deeper explanation, see the upstream HiRes Composite guide.

Preview and Full splat training

The documented training presets are:

| Preset | Training steps | Best suited to | |---|---:|---| | Preview | 7,000 | Evaluating the scene before committing to a longer run. | | Full | 30,000 | Giving a promising dataset more time to train. |

The training stage uses a random background and supports up to 3 million splats.

A longer run gives training more opportunities to refine the result, but it cannot guarantee that incorrect geometry or inconsistent generated imagery will be repaired. Use Preview to assess the overall scene before choosing Full.

Browser exploration and export

The finished world supports these controls:

| Control | Action | |---|---| | W / A / S / D | Move around | | R / F | Move up or down | | Drag | Look around | | Shift | Move faster |

You can download the trained world as a .ply file.

The exported result is a Gaussian splat. Do not assume it is a conventional textured mesh or a ready-to-use game level.

Saved worlds and restart recovery

Jobs are stored in:

text ComfyUI/output/splatkit_world/

  • Use the Worlds menu to switch between projects.
  • Use Resume to continue from the first unfinished stage.
  • Jobs survive ComfyUI restarts. This stage-based recovery helps you continue a project without automatically repeating every completed part. It does not imply that an interrupted generation or training operation resumes at its exact internal step.

Optional access to the underlying workflows

If you prefer ComfyUI's node canvas, the equivalent UI workflows are available in example_workflows/.

These expose the panorama stage and SplatKit's dataset graph, giving you another way to inspect the process. You do not need to use them to work through the guided app.

ComfyUI-SplatKit-World Models

ComfyUI-SplatKit-World uses several specialized models together rather than offering one model that performs the entire task.

| Model or model group | Role in the pipeline | What it means for your work | |---|---|---| | Krea 2 Turbo | Panorama image generation. | Creates the surrounding image used as the visual foundation of the world. | | Qwen3-VL 4B | Text and image encoding for the Krea 2 stage. | Helps provide the text and reference-image information used during generation. It is not a separate 3D generator. | | Two 360° panorama LoRAs | Specialized guidance for panoramic generation; the requirements specify using one of the two. | Adapt panorama generation to the extension's supported workflows. Exact checkpoint details are listed in the project documentation. | | MoGe | Depth and geometry estimation. | Supplies the estimated spatial structure used for camera previews and rendered control views. | | Wan 2.1 I2V 14B with LoRAs | Fly-through generation and completion of unseen regions. | Produces the additional viewpoints needed for reconstruction. | | Real-ESRGAN x2 | Image upscaling. | Supports the workflow's image-enlargement stage. |

The panoramic video adaptation is available as the Wan 2.1 Pano360 LoRA, converted from the Matrix-3D LoRA into ComfyUI's format.

Important compatibility boundaries

  • The extension documentation does not identify the exact MoGe version used. The broader MoGe project's models should not be assumed to be interchangeable within this app.
  • Smaller WAN or Matrix-3D models mentioned in upstream projects are not documented as selectable alternatives here.
  • The wider model lists supported by AI Toolkit do not describe models supported by ComfyUI-SplatKit-World.
  • SphereSfM, Spirula Studio, and Spark are reconstruction, training, and rendering tools—not alternative generation checkpoints. For the exact checkpoint filenames and expected model set, consult the project's model information.

ComfyUI-SplatKit-World Requirements and Practical Limits

This is a demanding local workflow. Hardware requirements apply to the whole pipeline, not just to the browser viewer or the splat trainer.

| Item | Documented guidance | |---|---| | ComfyUI | Tested with version 0.37.0 and frontend 1.52.7. | | Operating system | Windows or Linux; macOS is untested for this extension. | | GPU | NVIDIA GPU; plan for at least 24 GB VRAM. | | System memory | Plan for at least 64 GB RAM and a large page file. | | Author's test system | 96 GB VRAM / 62 GB RAM. | | Model storage | Approximately 46 GB. | | Output storage | Approximately 13 GB per world in ComfyUI's output folder. |

The workflow depends on these five custom node packs:

  • ComfyUI-SplatKit
  • ComfyUI-Mickmumpitz-Nodes
  • ComfyUI-Krea2-Ostris-Edit
  • ComfyUI_essentials
  • ComfyUI_UltimateSDUpscale It also uses a separately supplied Spirula Studio executable.

Spirula Studio itself supports additional GPU vendors and platforms, and Spark can render splats on many WebGL2 devices. Those capabilities do not change the documented NVIDIA and operating-system requirements of the complete ComfyUI-SplatKit-World pipeline.

Troubleshooting ComfyUI-SplatKit-World

The app shows a red box listing missing components

The app checks its dependencies and reports missing items.

  1. Read the names in the red box.
  2. Identify whether each item is a node pack, model, or Spirula Studio component.
  3. Compare the report with the project requirements and model table.
  4. Have the missing components made available on the ComfyUI machine.
  5. Restart ComfyUI and check the app again. The World app does not install packages or download the Spirula executable for you. Upstream SplatKit separately documents first-use downloads for some of its components, including MoGe and SphereSfM.

The SplatKit World button is missing

  1. Check the SplatKit menu as well as the top bar.
  2. Try opening /splatkit/world/ directly at your ComfyUI address.
  3. If the app still does not open, restart ComfyUI.
  4. Check for startup or dependency errors and consult the project troubleshooting guide. Use your actual ComfyUI address if it differs from the local example.

The camera preview contains black areas

Black regions in Preview camera indicate areas that the original panorama does not supply.

  1. Play through the preview.
  2. Check whether those regions appear where movement reveals previously hidden surfaces.
  3. Inspect the floor and side views for a path crossing scene geometry.
  4. Adjust the route if it exposes more unknown content than you intended.
  5. Preview again before approving. Some black areas are expected: filling those gaps is part of WAN's job.

Generated areas do not match the scene

Upstream SplatKit documentation emphasizes that the prompt must describe the actual panorama.

  1. Inspect the full panorama with View 360°.
  2. Compare its contents with your scene description.
  3. Remove contradictory objects, materials, lighting, or setting details.
  4. Describe the environment that is actually visible.
  5. Check whether the camera route reveals large amounts of unseen space. For example, if the panorama shows a stone courtyard, describing a futuristic metal corridor gives WAN conflicting guidance.

Generation fails with an out-of-memory error

  1. Check your GPU memory against the 24 GB or greater planning guidance.
  2. Close other GPU-heavy applications.
  3. Check available system RAM and the machine's page-file configuration.
  4. Confirm that there is enough free storage for the world's outputs.
  5. Reopen the project from Worlds and use Resume after addressing the resource problem. Do not apply low-memory claims from standalone WAN or Matrix-3D examples to this complete workflow. No specific low-VRAM mode is documented for the World app.

ComfyUI restarted before the world finished

  1. Open SplatKit World.
  2. Select the project from Worlds.
  3. Choose Resume. The app continues from the first unfinished stage. Preserve the project's files in ComfyUI/output/splatkit_world/.

The trained scene looks soft or structurally wrong

  1. Inspect the panorama for problems already present in the source.
  2. Check that the prompt accurately describes it.
  3. Review camera paths for movement through estimated surfaces.
  4. Use Preview training to evaluate the scene before a Full run.
  5. Remember that additional training cannot reliably correct contradictory views or invented geometry. The upstream detail-preservation guide explains how panorama resolution and splat capacity affect sharpness.

Frequently asked questions

Do you need a 360° camera? No. You can start from a photo or text prompt, or bring an existing 2:1 panorama.

Is the output only a 360° image? No. The panorama is an intermediate stage. The completed pipeline produces a trained 3D Gaussian splat that supports movement between viewpoints.

Does SplatKit already train the world? The upstream panorama dataset workflow produces training data. ComfyUI-SplatKit-World adds a guided process that includes Spirula Studio training and browser viewing.

Can you resume an unfinished world? Yes. Select it in Worlds and choose Resume to continue from the first unfinished stage.

Does it need a cloud API key? No cloud APIs or partner nodes are required.

Can you use it on macOS or an AMD GPU? The complete extension is documented for NVIDIA GPUs on Windows or Linux. macOS is untested, even though some individual tools support broader hardware.

Can you export a conventional mesh? The documented World app download is a .ply Gaussian splat. Mesh export is not documented as a feature of this app.

ComfyUI-SplatKit-World Security and Licensing

Local execution and network access

The author documents that the World app does not run package installers, use a shell, execute eval or exec, or change your environment. The only external program it starts is the Spirula Studio executable supplied on the ComfyUI machine.

The Spirula executable's location can only be set from that machine.

If you launch ComfyUI with --listen, the World app becomes reachable from your network just like ComfyUI itself. Local processing should not be confused with restricted network access.

See SECURITY.md for the project's security details.

Licenses to check before commercial use

The extension's code is licensed under MIT, but its dependencies and models retain their own licenses.

Important distinctions include:

  • Krea 2 and the two 360° LoRAs: Covered by the Krea 2 Community License. The project notes commercial use only below USD 1 million in annual revenue, with content filtering required when running it for others.
  • Spirula Studio: GPL-3.0, used as a separate program.
  • ComfyUI and ComfyUI_UltimateSDUpscale: GPL-3.0.
  • Bundled three.js and Spark: MIT. The extension's MIT license does not replace the terms of the models it runs. Review THIRD_PARTY_NOTICES.md and the applicable licenses before using the workflow commercially or providing it as a service.

Learn More about ComfyUI-SplatKit-World

Understand the artist workflow

Explore the underlying ideas visually

  • Matrix-3D project page: View examples of text- and image-based explorable scene generation. These demonstrate the underlying approach, not guaranteed results or controls in this extension.
  • MoGe project page (https://wangrc.site/MoGePage): See how a single image can be interpreted as 3D geometry.
  • MoGe interactive demo: Explore the depth-estimation concept separately from the local World workflow.

Learn about training and viewing splats

  • Spirula Studio: Read about the trainer used for the final reconstruction stage.
  • Spirula Studio web viewer: Explore a separate browser-based splat-viewing resource.
  • Spark documentation and FAQ (https://sparkjs.dev/): Learn about the rendering technology behind the World viewer.
  • Spark examples (https://sparkjs.dev/examples/): See demonstrations of Gaussian splat rendering. Spark's wider capabilities are not all exposed by the World app. For more advanced inspection of the panorama editing stage, see ComfyUI-Krea2-Ostris-Edit. Its documentation explains how Krea 2 receives text and reference-image information.

ComfyUI-SplatKit-World Related Nodes

RunComfy
Copyright 2025 RunComfy. All Rights Reserved.

RunComfy is the premier ComfyUI platform, offering ComfyUI online environment and services, along with ComfyUI workflows featuring stunning visuals. RunComfy also provides AI Models, enabling artists to harness the latest AI tools to create incredible art.

ComfyUI-SplatKit-World detailed guide | ComfyUI