ComfyUI Extension: ComfyUI_PHRenderFormerWrapper

Authored by paulh4x

Created

Updated

16 stars

A Wrapper and a set of Custom Nodes for using RenderFormer as a 3d Environment in ComfyUI.

Custom Nodes (0)

    README

    🧊 ComfyUI-PHRenderFormerWrapper 🧊

    Author: paulh4x

    WIP Screenshot 1 WIP Screenshot 2 WIP Screenshot 3 Resolution Comparison

    This repository contains a set of custom nodes for ComfyUI that provide a wrapper for Microsoft's RenderFormer model. It allows you to render complex 3D scenes (see Limits) with physically-based materials and global illumination directly within the ComfyUI interface.

    [!WARNING] WORK IN PROGRESS & HELLO WORLD: This project is my first "hello world" in contributing code of any kind and is currently under active development. It should be considered experimental. The model is in a very early stage and may be further developed later. I do not take responsibility if this breaks anything and do not plan to provide official support for it. PLEASE BE AWARE THAT EVERYTHING HERE IS ENTIRELY VIBECODED.

    🏆 Sponsorship

    • Please consider sponsoring me if you find the results of my work useful. A good way to keep code development open and free is through sponsorship.

    • BE A GITHUB SPONSOR ❤️ . DIRECTLY SUPPORT ME VIA PAYPAL . SUPPORT ME ON KO-FI!

    🚀 Features

    • 🎨 End-to-End Rendering: Load 3D models, define materials, set up cameras, and render—all within ComfyUI.
    • ⚙️ Modular Node-Based Workflow: Each step of the rendering pipeline is a separate node, allowing for flexible and complex setups.
    • 🎥 Animation & Video: Create camera and light animations by interpolating between keyframes. The nodes output image batches compatible with ComfyUI's native video-saving nodes.
    • 🔧 Advanced Mesh Processing: Includes nodes for loading, combining, remeshing, and applying simple color randomization to your 3D assets.
    • 💡 Lighting and Material Control: Easily add and combine multiple light sources and control PBR material properties like diffuse, specular, roughness, and emission.
    • ↔️ Full Transformation Control: Apply translation, rotation, and scaling to any object or light in the scene.

    🛠️ Installation

    Prerequisites

    • Git: Required for cloning the repository and installing certain dependencies.
    • Python: A Python version compatible with ComfyUI and PyTorch 2.0+.
    • PyTorch: A compatible version of PyTorch must be installed for your hardware (NVIDIA CUDA or Apple Metal). This wrapper is tested with PyTorch 2.7.1 and CUDA 12.6.
    • ComfyUI: A working installation of ComfyUI.

    Installation Steps

    1. Navigate to your ComfyUI custom_nodes directory: 
      cd ComfyUI/custom_nodes/
    2. Clone this repository: 
      git clone https://github.com/paulh4x/ComfyUI_PHRenderFormerWrapper.git
    3. Navigate into the newly cloned directory: 
      cd ComfyUI_PHRenderFormerWrapper
    4. Clone the official Microsoft RenderFormer repository into this directory. It must be named renderformer: 
      git clone https://github.com/microsoft/renderformer.git renderformer
    5. Install the required Python packages for this wrapper: 
      pip install -r requirements.txt
    6. RenderFormer requires an additional plugin for handling HDR image formats. Run the following command to download it: 
      python -c "import imageio; imageio.plugins.freeimage.download()"
    7. Restart ComfyUI.

    [!NOTE] For NVIDIA GPU users, the underlying RenderFormer library can optionally use Flash Attention for better performance. If you have it installed in your PyTorch environment, it should be used automatically.

    💡 Usage

    This node pack follows a standard 3D rendering pipeline. The basic workflow, as seen in the ph_renderformer_basic_01.json example, can be broken down into the following steps:

    graph TD
    subgraph "Step 1: Load Model"
        A["RenderFormer Model Loader"]
    end

    subgraph "Step 2: Define Scene Elements"
        B["RenderFormer Mesh Loader"] --> C["RenderFormer Random Colors"]
        D["RenderFormer Camera"]
        E["RenderFormer Lighting"]
    end

    subgraph "Step 3: Build Scene"
        C --> F["RenderFormer Scene Builder"]
        D --> F
        E --> F
    end

    subgraph "Step 4: Render (Sample)"
        A --> G["RenderFormer Sampler"]
        F --> G
    end

    subgraph "Step 5: Save Output"
        G --> H["Save Image"]
    end

    style A fill:#FDC501,stroke:#111417,stroke-width:2px,color:#111417
    style B fill:#FDC501,stroke:#111417,stroke-width:2px,color:#111417
    style C fill:#FDC501,stroke:#111417,stroke-width:2px,color:#111417
    style D fill:#FDC501,stroke:#111417,stroke-width:2px,color:#111417
    style E fill:#FDC501,stroke:#111417,stroke-width:2px,color:#111417
    style F fill:#FDC501,stroke:#111417,stroke-width:2px,color:#111417
    style G fill:#FDC501,stroke:#111417,stroke-width:2px,color:#111417
    style H fill:#a0a0a0,stroke:#111417,stroke-width:2px,color:#111417
    1. Load Model: Start by loading the RenderFormer Model Loader. This node downloads the specified model from Hugging Face and prepares it for the pipeline.
    2. Define Scene Elements:
      • Mesh: Use the RenderFormer Mesh Loader to load your 3D model (e.g., an .obj file). Here, you can also set its transformation (translation, rotation, scale) and define its core material properties. In the example, the output is passed to RenderFormer Random Colors for a creative effect.
      • Camera: Add a RenderFormer Camera to define the viewpoint, target, and field of view.
      • Lighting: Use the RenderFormer Lighting node to create one or more light sources. You can combine multiple lights with the RenderFormer Lighting Combine node.
    3. Build Scene: Connect the MESH, CAMERA, and LIGHTING outputs into the RenderFormer Scene Builder. This node collects all the components and assembles the final scene data that the model will render. You can also choose to add a default background plane and walls here.
    4. Render (Sample): The RenderFormer Sampler takes the MODEL and the SCENE as input. It performs the actual rendering operation and outputs the final image.
    5. Save Output: Connect the IMAGE output from the sampler to a Save Image node to save the result.

    For video, the workflow is similar. You would use a RenderFormer Camera Target to create a CAMERA_SEQUENCE, which then goes into the RenderFormer Scene Builder. The RenderFormer Sampler will automatically detect the sequence and output an image batch (IMAGES) ready for a video-saving node.

    🚀 Advanced Usage: Building Complex Scenes & Animations

    The ph_renderformer_advanced_01.json workflow is a powerful example that demonstrates how to build a complex scene with multiple, independent objects and generate both a static image and an animated video from a single, unified pipeline.

    graph TD
    subgraph "I. Model"
        ModelLoader["RenderFormer Model Loader"]
    end

    subgraph "II. Scene Inputs"
        subgraph "A. Meshes"
            Mesh1["LoadMesh (Processed)"] --> Remesh["Remesh / Randomize"] --> Combine["RenderFormerMeshCombine"]
            Mesh2["LoadMesh (Direct)"] --> Combine
        end
        
        subgraph "B. Camera"
            CamStart["RenderFormerCamera (Start)"] --> CamEnd["PHRenderFormerCameraTarget (End)"]
        end

        subgraph "C. Lighting"
            LightStart["RenderFormerLighting (Start)"] --> LightEnd["RenderFormerLightingTarget (End)"]
        end
    end

    subgraph "III. Scene Assembly"
        Combine --> SceneBuilder["RenderFormerSceneBuilder"]
        CamStart --> SceneBuilder
        CamEnd -- CAM_SEQUENCE --> SceneBuilder
        LightStart --> SceneBuilder
        LightEnd -- LIGHT_SEQUENCE --> SceneBuilder
    end

    subgraph "IV. Rendering & Output"
        Sampler["PHRenderFormerVideoSampler"]
        ModelLoader --> Sampler
        SceneBuilder -- SCENE_SEQUENCE --> Sampler
        Sampler -- IMAGES --> SaveImg["Save Image"]
        Sampler -- IMAGES --> CreateVid["Create Video"] --> SaveVid["Save Video"]
    end

    style ModelLoader fill:#FDC501,stroke:#111417,stroke-width:2px,color:#111417
    style Mesh1 fill:#FDC501,stroke:#111417,stroke-width:2px,color:#111417
    style Mesh2 fill:#FDC501,stroke:#111417,stroke-width:2px,color:#111417
    style Remesh fill:#FDC501,stroke:#111417,stroke-width:2px,color:#111417
    style Combine fill:#FDC501,stroke:#111417,stroke-width:2px,color:#111417
    style CamStart fill:#FDC501,stroke:#111417,stroke-width:2px,color:#111417
    style CamEnd fill:#FDC501,stroke:#111417,stroke-width:2px,color:#111417
    style LightStart fill:#FDC501,stroke:#111417,stroke-width:2px,color:#111417
    style LightEnd fill:#FDC501,stroke:#111417,stroke-width:2px,color:#111417
    style SceneBuilder fill:#FDC501,stroke:#111417,stroke-width:2px,color:#111417
    style Sampler fill:#FDC501,stroke:#111417,stroke-width:2px,color:#111417
    style SaveImg fill:#a0a0a0,stroke:#111417,stroke-width:2px,color:#111417
    style CreateVid fill:#a0a0a0,stroke:#111417,stroke-width:2px,color:#111417
    style SaveVid fill:#a0a0a0,stroke:#111417,stroke-width:2px,color:#111417

    This workflow is designed for maximum flexibility and showcases several key features of the node pack:

    1. Model Loading: The process begins on the left with the RenderFormer Model Loader, which prepares the core AI model for rendering.
    2. Complex Asset Pipeline:
      • Multiple Mesh Inputs: The scene is composed of multiple LoadMesh nodes. This allows you to import various objects, characters, and background elements independently.
      • Processing Chains: Some meshes go through additional processing steps. For example, a main object might be passed through Remesh to optimize its geometry or RandomizeColors to apply procedural materials. Other meshes might be loaded directly without modification.
      • Combining Meshes: All individual mesh pipelines converge into a RenderFormerMeshCombine node. This powerful utility merges all separate objects into a single, organized list that the renderer can understand.
    3. Static and Animated Components:
      • Camera: A RenderFormerCamera defines the starting position and angle. This is then fed into a PHRenderFormerCameraTarget to create a smooth animation sequence from a start to an end state.
      • Lighting: Similarly, a RenderFormerLighting node sets the initial light properties, which are then animated over time by a RenderFormerLightingTarget node.
    4. Unified Scene Building: The RenderFormerSceneBuilder acts as the central assembly point. It is designed to be highly flexible and accepts all the core scene components:
      • It takes the combined MESH list.
      • It takes the static CAMERA and LIGHTING to define the scene's state at frame zero.
      • It takes the CAM_SEQUENCE and LIGHT_SEQUENCE outputs from the target nodes to understand how the scene should evolve over time.
      • It intelligently processes these inputs to generate a complete SCENE_SEQUENCE, which is a frame-by-frame description of the entire animation.
    5. Rendering and Unified Output:
      • The PHRenderFormerVideoSampler takes the MODEL and the SCENE_SEQUENCE and renders every frame of the animation.
      • Crucially, it outputs a single IMAGES batch. This batch can be used for multiple purposes simultaneously: you can save individual frames from the animation using a Save Image node, and also pass the entire batch to a Create Video node to produce the final movie file.

    📦 Nodes

    This wrapper provides a comprehensive set of nodes to build 3D scenes.

    Core Pipeline

    • RenderFormer Model Loader: Loads a specified RenderFormer model from Hugging Face or a local path.
    • RenderFormer Scene Builder: Assembles a scene from meshes, lighting, and camera inputs. It can handle both single camera inputs for static images and camera sequences for creating animations.
    • RenderFormer Sampler: Executes the RenderFormer pipeline on a scene or a sequence of scenes. It intelligently handles both single images and image batches for video, producing the final rendered output.

    Scene Components

    • RenderFormer Mesh Loader: Loads a 3D mesh file (e.g., .obj, .glb). This node is also used to define an object's material properties and transformations.
    • RenderFormer Camera: Defines the camera's position, look-at target, and field of view (FOV).
    • RenderFormer Lighting: Creates a configurable emissive light source.

    Animation

    • RenderFormer Camera Target: Creates a camera animation sequence by defining a start and end camera state (position, look-at, FOV).
    • RenderFormer Lighting Target: Creates a light animation sequence by defining a start and end light state (position, rotation, scale, and emission).
    • RenderFormer Mesh Target: Creates a mesh animation sequence by defining a start and end state for a mesh's transformation (position, rotation, scale) and material properties (diffuse color, specular color, roughness).

    Utilities

    • RenderFormer Mesh Combine: Combines multiple MESH outputs into a single object list.
    • RenderFormer Lighting Combine: A powerful node that enables multi-light animation. It accepts any combination of static lights (LIGHTING) and animated lights (LIGHTING_SEQUENCE), and outputs a single, unified LIGHTING_SEQUENCE ready for the Scene Builder. It also outputs the start_frame_lighting separately for a direct, convenient connection.
    • RenderFormer Remesh: Simplifies the geometry of a mesh to a target face count using pymeshlab.
    • RenderFormer Random Colors: Applies random colors to a mesh. It offers three modes: per-object (one color for the whole object), per-shading (uses the RenderFormer backend to color based on shading groups), and per-triangle (assigns a unique color to every face).

    Advanced & Experimental

    • RenderFormer From JSON: Loads a scene from a JSON definition, allowing for more complex and customized setups based on the original RenderFormer format.
    • RenderFormer Example Scene: A test node to quickly load one of the official RenderFormer example scenes.

    🧱 Model Limits

    • Meshes: RenderFormer can handle meshes of up to 8192 polygons.
    • Lights: Up to 8 lightsources can be combined for lighting, emission color is for now limited to rgb 255, 255, 255 (white).
    • Resolution: Best tested resolutions for now are between 512 x 512 to 1024 x 1024 pixels. The model can produce resolutions up to 2048 x 2048 pixel, quality of outputs decrease with higher resolutions (see comparison img).
    • Animations: Due to slightly varying precision in each frame rendered, camera animations for now contain some flickering, especially with high reflective materials.

    📝 Progress & To-Do

    This project is under active development. Here is a summary of the progress so far and the features planned for the future.

    ✅ Done

    • [x] Core Rendering Pipeline: All essential nodes for building and rendering a static 3D scene are implemented.
    • [x] In-Memory Processing: The entire scene preparation pipeline, from loading meshes to generating the final HDF5 data, is now handled in-memory. This significantly improves performance by avoiding slow disk I/O.
    • [x] Video Rendering: A complete, dedicated workflow for video rendering is in place.
    • [x] Light Animation: Full animation support for lights, including position, rotation, scale, and emission.
    • [x] Advanced Utilities: Nodes for mesh combination, remeshing, and JSON-based scene loading are functional.
    • [x] Custom UI Elements: The nodes feature custom colors for better visual organization, and progress bars are implemented for long-running operations.
    • [x] Bug Fixes: Addressed various bugs related to file handling, data types, and temporary file management.
    • [x] Proper RGB Values for diffuse Color: Colorpicker support for white-values for more variety of colors including.
    • [x] Randomize Color Fix: The RandomizeColors node now correctly applies colors in all modes. The original per-triangle mode was renamed to per-shading for clarity, and a new per-triangle mode was added for true per-face coloring.
    • [x] Mesh Animation: Implemented animation capabilities for MESH properties (translation, rotation, scale) and materials (diffuse, specular, roughness).
    • [x] Mesh Animation: Add multi mesh animation for complex movement
    • [x] Public Release: Prepare for a more stable, public release with better documentation and examples.

    📋 To Do

    • [ ] Animation Flickering: Investigate and fix flickering in animations, especially for objects with highly reflective materials.
    • [ ] Mesh Format Support: Add support for loading .glb and .fbx files, including their materials and textures.
    • [ ] Camera Adoption: Integrate with the Load 3D core node to adopt its camera transformations.
    • [ ] Material Presets: Create a system for saving and loading material presets.
    • [ ] Overall Animation: Implement multiple keyframes and different interpolation options (bezier?) for complexer animations

    📜 Version History

    Version 0.5.0 - Multi Mesh Animation & Public Release

    • Feature: Mesh & Material Animation: Multi object support for animations now integrated.

    Version 0.4.0 - Mesh Animation

    • Feature: Mesh & Material Animation: A complete animation workflow for meshes has been implemented.
    • New Node: RenderFormerMeshTarget: This new node allows you to define an animation for a mesh by specifying its end-state transformation (position, rotation, scale) and material properties (diffuse color, specular color, roughness).
    • Enhancement: RenderFormerMeshCombine: This node has been upgraded to intelligently handle both static meshes and the new animated mesh sequences, allowing you to combine multiple animated and static objects in a scene.
    • Enhancement: RenderFormerSceneBuilder: The scene builder now correctly processes the new mesh animation sequence, enabling it to generate interpolated frames for mesh animations just as it does for cameras and lights.

    Version 0.3.35 - Bug Fixes & Stability

    • Fix: Resolved a critical WrongTypeError in the RenderFormerRandomizeColors node. The random_diffuse_type was being set to an incorrect string value ("per triangle"), which has now been corrected to the valid literal ("per-triangle").
    • Fix: The per-shading mode in the RenderFormerRandomizeColors node has been updated to use the correct literal "per-shading-group" for consistency with the scene configuration, ensuring it functions as expected.
    • Change: The seed input on the RenderFormerRandomizeColors node is now correctly limited to a 10-digit integer.

    Version 0.3.3 - Workflow Optimization

    • Workflow: The animation workflow has been significantly simplified and made more robust. The RenderFormerSceneBuilder now has a required lighting input that explicitly defines the start-frame lighting.
    • Workflow: The end_lighting input on the Scene Builder has been removed. All lighting animation is now handled by the optional lighting_sequence input, which provides the end-frame state.
    • Data Flow: The RenderFormerLightingTarget and RenderFormerLightingCombine nodes have been updated to output a standardized LIGHTING_SEQUENCE dictionary ({ "start_lights": [...], "end_lights": [...] }), ensuring consistent data flow.
    • Convenience: The RenderFormerLightingCombine node now provides a separate start_frame_lighting output for a direct connection to the Scene Builder's required lighting input, preserving a clean and logical graph.

    Version 0.3.0 - Animation Overhaul and Performance Boost

    • Feature: Full Light Animation: The RenderFormerLightingTarget node now supports animating a light's position, rotation, scale, and emissive strength, enabling complex dynamic lighting effects.
    • Fix: Animation Transformation: Reworked the animation pipeline to correctly interpolate object and light transformations, ensuring smooth movement, rotation, and scaling.
    • Performance: In-Memory Scene Generation: The entire scene-building process now runs in-memory, dramatically reducing processing time by avoiding disk I/O for intermediate files.
    • Workflow: The RenderFormer Sampler and Scene Builder nodes now handle both static scenes and animation sequences more robustly.

    Version 0.2.5 - Node Consolidation and Workflow Simplification

    • Workflow: Merged the RenderFormerVideoSamplerBatched into the main RenderFormerGenerator (now RenderFormer Sampler). The sampler now intelligently handles both single images and video sequences, with an optional IMAGES output.
    • Workflow: Merged the RenderFormerVideoSceneBuilder into the main RenderFormerSceneBuilder. The node now accepts both a single CAMERA and a CAMERA_SEQUENCE, with optional SCENE and SCENE_SEQUENCE outputs.
    • Fix: Resolved a critical AttributeError in the RenderFormerSceneBuilder caused by a missing helper method after the node merge.
    • Fix: Removed hardcoded material properties for background meshes in the RenderFormerLoadMesh node, ensuring that materials are always controlled by the node's inputs.
    • UI/UX: Added a colored startup message with an ASCII logo to provide clear feedback on whether the nodes loaded successfully.
    • UI/UX: Fixed a SyntaxWarning related to escape sequences in the startup message.

    Version 0.2.21 Beta Test - UI/UX and Workflow Enhancements

    • Examples: added example workflows for ComfyUI.
    • UI/UX: All nodes now have custom colors for better visual organization in the graph. The header is yellow (#FDC501) and the body is anthracite (#111417).
    • UI/UX: Added progress bars to all long-running nodes (SceneBuilder, VideoSceneBuilder, Sampler, VideoSampler, Remesh) to provide real-time feedback.
    • Workflow: The Remesh node now processes all meshes in a list, not just the first one.
    • Workflow: The SceneBuilder nodes will no longer add the default background if a background mesh has already been loaded manually, preventing duplicates.
    • Input Precision: The emissive_strength input on the Lighting node now has a smaller step value for finer control.

    Version 0.1.9 - Video Rendering and Stability

    • Feature: Added a complete, dedicated workflow for video rendering (Camera Target, Video Scene Builder, Video Sampler).
    • Fix: Resolved a dacite.exceptions.MissingValueError in the video scene builder.
    • Fix: Addressed TypeError and FileNotFoundError bugs related to temporary file handling in the video pipeline.
    • Change: Separated single-image and video workflows for improved stability and clarity.

    Version 0.1.5 - Refinements and Fixes

    • Feature: The LoadMesh node now has a RENDERFORMER_MATERIAL output, allowing material properties to be shared.
    • Feature: Added default transformations for specific background meshes from the official examples.
    • Change: Increased input precision for rotation and scale transformations.

    Version 0.1.0 - Initial Development

    • Feature: Established the core rendering pipeline and all essential nodes for building and rendering a static 3D scene.
    • Feature: Implemented in-memory scene processing to patch trimesh and h5py, avoiding slow disk I/O.
    • Feature: Added advanced utility nodes like RemeshMesh, RandomizeColors, and FromJSON.
    • UI/UX: Integrated a file uploader and a custom canvas-based color picker for a better user experience.

    💖 Contact

    🌱 My Journey

    This project represents my first steps into coding and open-source contribution. It was born out of a desire to learn and create. What started as a simple experiment has been a journey of discovery, and I'm excited to see where it goes. I would love to develop this project further with the help of the community and welcome any contributions or feedback.

    🙏 Acknowledgements

    This project would not be possible without the foundational work of others.

    • RenderFormer This project is a wrapper for the incredible RenderFormer model. All credit for the underlying rendering technology goes to the original authors: Chong Zeng, Yue Dong, Pieter Peers, Hongzhi Wu, and Xin Tong.

    • comfyui-hunyuan3dwrapper Special thanks to kijai for his work on the comfyui-hunyan3dwrapper, which served as an invaluable reference and starting point for this project during its development in vibecoding.

    • ComfyUI_Fill-Example-Nodes A huge thank you to filliptm for creating the ComfyUI_Fill-Example-Nodes repository. The advanced color picker in this project was implemented by adapting the excellent example code provided in that repo. It was an essential learning resource for understanding how to build custom UI elements in ComfyUI.

    • Awesome people from the Internet Special thanks to m4d, rem431, Mel and hicho for pointing me directions, providing tips and feedback.

    ⚖️ License

    • <a href="https://github.com/paulh4x/ComfyUI_PHRenderFormerWrapper">ComfyUI_PHRenderFormerWrapper</a> © 2025 by <a href="https://www.linkedin.com/in/paul-hansen-410695b6/">Paul Hansen</a> is licensed under <a href="https://creativecommons.org/licenses/by-sa/4.0/">CC BY-SA 4.0</a>