comfyui-lrw-nodes

ComfyUI custom nodes for latent-space geometry, geodesic interpolation, and WAN2.2 video guidance powered by latent-riemannian-world.

This project provides practical nodes for experimenting with latent-space paths in ComfyUI, including:

• geodesic and SLERP-style latent interpolation
• curvature and latent-distance analysis
• parallel transport style/vector operations
• WAN2.2 First-Last Frame video workflows with LRW-guided latent blending

WAN2.2 note:
The WAN2.2 integration is designed as a practical, VRAM-safe workflow.
Direct WAN2.2 First-Last Frame generation remains the main continuity path, while LRW geodesic keyframes are softly blended into the video latent before sampling.
This allows the geodesic structure to influence intermediate motion without sacrificing first-to-last continuity.

---

Demo

Direct WAN2.2 FLF vs LRW-Guided Latent Blend

In both comparison videos:

• Left: Direct WAN2.2 First-Last Frame baseline
• Right: WAN2.2 with LRW geodesic-guided latent blend

The right-side result uses LRW geodesic keyframes as a latent guide before KSampler.

4-step comparison

https://github.com/user-attachments/assets/07416cb8-5d9d-4c26-b306-48ea34323533

KSampler steps: 4
Left: Direct WAN2.2 FLF
Right: LRW-guided latent blend

6-step comparison

https://github.com/user-attachments/assets/43b13522-367e-4974-8920-d5dbab3dc597

KSampler steps: 6
Left: Direct WAN2.2 FLF
Right: LRW-guided latent blend

Recommended demo workflow:

examples/workflows/wan22_practical_direct_flf_plus_lrw_guided_blend.json

---

What the WAN2.2 LRW workflow does

The practical WAN2.2 workflow uses a two-branch structure.

Main continuity branch:

first image + last image
→ WanFirstLastFrameToVideo
→ base video latent

LRW guide branch:

first latent + last latent
→ LRW_WanTemporalMetric
→ LRW_WanCurvatureGuide
→ LRW_WanGeodesicKeyframes
→ geodesic guide latent

Final sampling branch:

base video latent + geodesic guide latent
→ LRW_WanLatentGuideBlend
→ KSampler
→ VAEDecode
→ VHS_VideoCombine

This keeps the target last frame active while allowing the LRW geodesic branch to influence the intermediate latent motion.

---

Node List

WAN2.2 Video Nodes (lrw/wan) ★ Core Feature

Designed for WAN2.2 First-Last Frame video generation.

| Node | Description | |---|---| | LRW WAN Temporal Metric | Creates a lightweight LRW-compatible metric descriptor for WAN latent-space analysis. Avoids WAN VAE decoder Jacobian paths for VRAM safety. | | LRW WAN Curvature Guide | Estimates latent distance / curvature-like complexity between first and last latent states. Useful for choosing steps, frame count, and keyframe strategy. | | LRW WAN Geodesic Keyframes | Generates LRW geodesic-style keyframe latents between first and last latent states. Supports WAN-safe 5D latent output. | | LRW Latent Keyframe Picker | Selects one geodesic keyframe from a stacked keyframe latent batch. Can output WAN-compatible 5D latent format. | | LRW WAN Latent Guide Blend | Blends a selected LRW geodesic guide into the actual WAN video latent before KSampler. This is the main practical node for making LRW affect video generation. |

Core Bridge Nodes (lrw/core)

| Node | Description | |---|---| | VAE Decoder Bridge | Wraps a decoder/VAE-style function for LRW metric experiments. | | Latent Blend / Frame Select | Selects or blends latent frames from a latent path. | | Latent Vector from Diff | Creates a latent direction vector from two latents. | | Apply Transported Vector | Applies a transported style/vector direction to a target latent. |

Metric Nodes (lrw/metric)

| Node | Description | |---|---| | Pullback Metric | Computes pullback metric experiments for compatible decoders. | | Latent Curvature Map | Visualizes local metric/volume behavior where supported. |

Geodesic Nodes (lrw/geodesic)

| Node | Description | |---|---| | Geodesic Interpolate | Interpolates between latents using LRW geodesic tools where supported. | | Geodesic Distance | Estimates geodesic distance between latent states. | | SLERP Interpolate | Baseline spherical interpolation for comparison. |

Transport Nodes (lrw/transport)

| Node | Description | |---|---| | Parallel Transport | Transports a latent/vector direction along a path using LRW transport tools. |

World Model Nodes (lrw/world)

| Node | Description | |---|---| | Latent Trajectory | Generates latent-space trajectories for experimental world-model workflows. |

---

Example Workflows

1. WAN2.2 Direct FLF + LRW-Guided Latent Blend

examples/workflows/wan22_practical_direct_flf_plus_lrw_guided_blend.json

Recommended practical WAN2.2 workflow.

It keeps WAN2.2 First-Last Frame as the main continuity path, then injects LRW geodesic guidance into the video latent before KSampler.

LoadImage(first) ─┐
                  ├─ WanFirstLastFrameToVideo ─┐
LoadImage(last)  ─┘                             │
                                                ├─ LRW_WanLatentGuideBlend → KSampler → VAEDecode → VHS_VideoCombine
VAEEncode(first) ─┐                             │
                  ├─ LRW_WanGeodesicKeyframes ─┘
VAEEncode(last)  ─┘

2. WAN2.2 Direct FLF Baseline

examples/workflows/wan22_direct_flf_baseline.json

Baseline workflow for comparison.

3. Geodesic Interpolation for Images

examples/workflows/geodesic_interpolation_workflow.json

Image-to-image latent geodesic interpolation for compatible image models.

4. Style Transfer via Parallel Transport

examples/workflows/style_transfer_workflow.json

Transport style direction vectors along latent paths.

---

Recommended repository layout

comfyui-lrw-nodes/
├─ examples/
│  ├─ workflows/
│  │  ├─ wan22_direct_flf_baseline.json
│  │  └─ wan22_practical_direct_flf_plus_lrw_guided_blend.json
│  │
│  └─ media/
│     ├─ output_compare_side_by_side.mp4
│     └─ output_compare_side_by_side01.mp4
│
├─ nodes/
│  ├─ wan_nodes.py
│  ├─ latent_keyframe_picker.py
│  └─ lrw_wan_latent_guide_blend.py
│
├─ README.md
└─ requirements.txt

---

Installation

Via ComfyUI Manager

Search for:

comfyui-lrw-nodes

Manual installation

cd ComfyUI/custom_nodes
git clone https://github.com/lajjadred/comfyui-lrw-nodes
cd comfyui-lrw-nodes
pip install -r requirements.txt

Restart ComfyUI after installation.

---

Required custom nodes for WAN2.2 workflows

The WAN2.2 demo workflows may require:

• ComfyUI-GGUF — UnetLoaderGGUF
• WAN2.2 First-Last Frame support — WanFirstLastFrameToVideo
• ComfyUI-VideoHelperSuite — VHS_VideoCombine
• Optional: ComfyUI-KJNodes for image resizing and utility nodes

---

VRAM Guide

| VRAM | Recommended WAN2.2 Setting | |---|---| | 12GB | GGUF Q3/Q4, blend_strength=0.05–0.10, lower frame count | | 16GB | GGUF Q3/Q4, blend_strength=0.10–0.20, 81 frames practical | | 24GB+ | Larger models or higher steps, stronger experimentation possible |

For the practical LRW-guided WAN workflow:

LRW_WanLatentGuideBlend
blend_strength: 0.10–0.20
time_schedule: middle_focus
keyframe_index: middle keyframe

Suggested starting point:

blend_strength = 0.15
time_schedule = middle_focus
normalize_guide = true

---

Why use LRW guidance with WAN2.2?

Direct WAN2.2 First-Last Frame generation is usually the best source of first-to-last continuity.
However, with large changes in pose, gaze, framing, or camera distance, the intermediate motion can become abrupt or unstable.

LRW-guided latent blending adds a soft latent-space guide:

| Method | Role | Expected effect | |---|---|---| | Direct FLF | Main continuity path | Strong first-to-last arrival | | LRW Geodesic Keyframes | Latent-space guide | Provides intermediate direction | | LRW Latent Guide Blend | Practical injection point | Lets the geodesic guide influence KSampler input | | Strong midpoint replacement | Not recommended as default | Can weaken arrival at the real last frame |

The recommended workflow does not replace the last frame with a midpoint.
Instead, it keeps the actual last frame in WAN FLF and blends LRW guidance into the video latent before sampling.

---

Mathematical note

For general compatible decoders, LRW can be used to explore pullback metrics, geodesic interpolation, transport, and latent-space trajectories.

For WAN2.2, the integration is intentionally practical and VRAM-safe:

• WAN VAE decoder Jacobian / dense pullback metric paths can be too expensive or incompatible with some vmap/in-place operations.
• The WAN workflow therefore uses LRW geodesic keyframes as a latent guide.
• The guide is softly blended into the WAN video latent rather than replacing WAN's first-last conditioning.

This keeps the workflow usable on consumer GPUs while still allowing LRW geometry to influence video generation.

---

Requirements

• ComfyUI
• latent-riemannian-world >= 0.3.0
• torch >= 2.4
• Python compatible with your ComfyUI installation

---

License

BSL-1.1 — (c) 2025 lajjadred