ComfyUI-loger

A ComfyUI custom node pack for running LoGeR (Long-sequence Generalizable Reconstruction) inference directly inside ComfyUI workflows. LoGeR estimates per-frame depth maps, confidence maps, and full 3-D point clouds from a sequence of images (treated as a video).

---

Nodes

Node	Description
LoGeR Model (Down)Loader	Loads LoGeR or LoGeR_star weights. Downloads automatically from HuggingFace (Junyi42/LoGeR) if not present locally.
LoGeR Inference	Runs the reconstruction model on an image batch and returns a LOGER_OUTPUT bundle.
LoGeR Depth Output	Extracts per-frame depth maps and confidence maps as IMAGE tensors.
LoGeR to Pointcloud	Exports the reconstruction as a pointcloud.ply and optionally COLMAP cameras.txt / images.txt.
LoGeR to Houdini Script	Generates a Houdini Python script that imports camera poses as keyframes, aligned to the exported PLY.

---

Requirements

• ComfyUI with the V3 node API (comfy_api.latest)
• Python packages: see requirements.txt
  • plyfile
  • huggingface_hub (for automatic weight download)
  • Standard ML stack: torch, numpy, Pillow, pyyaml

Install dependencies:

pip install -r requirements.txt

---

Installation

1. Clone this repository into your ComfyUI/custom_nodes/ directory:

cd ComfyUI/custom_nodes
git clone https://github.com/shanef3d/ComfyUI-loger

2. Restart ComfyUI. The LoGeR nodes will appear in the LoGeR category.

---

Model Weights

Weights are downloaded automatically on first use from Junyi42/LoGeR via huggingface_hub. They are stored under ComfyUI/models/LoGeR/ by default. You can override the path in the LoGeR Model (Down)Loader node.

Two variants are available:

• LoGeR — base model
• LoGeR_star — improved variant (default)

---

Typical Workflow

Load Images → LoGeR Model Loader → LoGeR Inference → LoGeR Depth Output
                                                    → LoGeR to Pointcloud
                                                    → LoGeR to Houdini Script

LoGeR Inference parameters

Parameter	Default	Description
window_size	64	Sliding-window size (-1 = full sequence)
overlap_size	3	Frame overlap between windows
filter_edges	true	Zero out depth-edge confidence
target_height	280	Inference height (snapped to multiples of 14)
target_width	504	Inference width (snapped to multiples of 14)
use_sim3	false	Use Sim3 alignment instead of SE3

LoGeR to Pointcloud parameters

Parameter	Default	Description
filename_prefix	loger	Output sub-folder inside ComfyUI/output/
save_colmap_txt	false	Also write COLMAP cameras.txt / images.txt
fov_degrees	60.0	Assumed horizontal FOV for intrinsics
conf_threshold	0.1	Minimum confidence to include a point
downsample	1	Spatial grid downsampling factor
keep_random_points	1.0	Random fraction of points to keep

---

Houdini Integration

The LoGeR to Houdini Script node writes a self-contained Python script. Open Houdini's Python Script editor and run it to create a LoGeR_Camera object with per-frame keyframes. Import pointcloud.ply without any axis transform — the script handles the OpenCV → Houdini convention conversion internally.

---

Credits

• LoGeR model: Junyi42/LoGeR
• ComfyUI custom node integration: this repository

---

License

MIT — see LICENSE.