[Release] The Geometry Duo for ComfyUI: Clean 'White Models' & Normal Guidance for MiniMax H3, Wan 2.1 & Video Diffusion
retained text
ComfyUI-Depth-Anything-3: [https://github.com/1038lab/ComfyUI-Depth-Anything-3](https://github.com/1038lab/ComfyUI-Depth-Anything-3) ComfyUI-Lotus-2: [https://github.com/1038lab/ComfyUI-Lotus-2](https://github.com/1038lab/ComfyUI-Lotus-2) Hey everyone, If you have been working with modern video generation models like MiniMax H3, Wan 2.1, or CogVideoX, you already know the biggest obstacle to getting cinematic, consistent motion: geometry instability. Without solid spatial guidance, characters morph, limbs glitch, and backgrounds wobble between frames. The standard technique among high-end AI creators is feeding geometric 'white models' (clay proxies), metric depth, and 3D surface normal passes into the generation pipeline to anchor every movement. To make generating these white models effortless and flicker-free in ComfyUI, we just published two companion nodes under the AILab/Geometry category: 1. ComfyUI-Depth-Anything-3: Real-time batching workhorse for video white models, pure PyTorch GPU colormaps (zero CPU bottleneck), and true physical metric depth (in meters). 2. ComfyUI-Lotus-2: Dual-engine generative normals and disparity (SD2.1 Fast Engine + FLUX.1-dev DiT Detail Engine) with 2-wire plug-and-play, built-in 2MB empty prompt vector (no 10GB T5 CLIP needed), and Stage 2 Detail Sharpener for hair and fur. # Why Two Nodes? The Right Tool for the Right Pipeline |Creative Workflow|Recommended Node|Why?| |:-|:-|:-| |MiniMax H3 / Wan 2.1 Video Guidance|Depth Anything 3|Rock-solid temporal consistency (min\_max), zero CPU stalls on 60+ frame batches.| |3D Camera Tracking / Blender Projection|Depth Anything 3 (Metric)|Calculates real-world distance in physical meters.| |Micro-Detail Hair / Fur / Delicate Fabric|Lotus-2 (FLUX)|Generative diffusion prior + Detail Sharpener eliminates latent patch grid lines.| |Game-Ready 3D Surface Normal Maps|Lotus-2 (FLUX / SD2.1)|Generates authentic XYZ direction vector normal passes.| |Ultra-Fast Local Previews (<1.5s)|Lotus (SD2.1) or DA3|Single forward-pass execution without DiT overhead.| # 1. Depth Anything 3: High-Throughput Video Batching & Metric Depth * Pure PyTorch GPU Colorizer: Standard depth colorizers offload tensors to CPU NumPy or matplotlib. DA3 does all color mapping (inferno, magma, plasma, viridis, cividis, turbo, gray) directly in VRAM via pure PyTorch kernels. When batching long video sequences, you get 0% CPU stall. * Flicker-Free Temporal Consistency: The min\_max mode normalizes depth across the entire video batch uniformly, preventing the frame-to-frame flashing common in standard depth estimators. * Physical Distance in Meters: The Metric checkpoints output real-world physical depth, giving you accurate depth-of-field and 3D spatial alignment. Repository: [https://github.com/1038lab/ComfyUI-Depth-Anything-3](https://github.com/1038lab/ComfyUI-Depth-Anything-3) # 2. Lotus-2: Generative Precision with Zero Clutter (FLUX + SD2.1) Traditional implementations of Lotus-2 require dragging 6 to 7 nodes onto your canvas (Base Model, LoRA, Dual CLIP, Sampler, VAE). We redesigned the entire node from scratch: * 2-Wire Plug-and-Play: Connect only image and model. * Zero CLIP Clutter: FLUX requires an empty text conditioning prompt. Instead of forcing you to load a 10GB T5 CLIP loader, Lotus-2 has a precomputed 2MB empty prompt vector (flux\_empty\_embed.pt) built-in. * Auto VAE Resolution: Automatically detects and caches your local ae.safetensors in VRAM. * Detail Sharpener: Running steps = 1 takes \~13 seconds on GGUF for clean disparity and normals. Setting steps = 2 to 4 engages the Stage 2 Detail Sharpener to smooth patch boundaries and recover individual strands of wavy hair or animal fur. Repository: [https://github.com/1038lab/ComfyUI-Lotus-2](https://github.com/1038lab/ComfyUI-Lotus-2) # Installation & Workflows Both nodes are available right now in the ComfyUI Manager: * Search keyword: ailab (or Depth-Anything-3 / Lotus-2) Manual Installation: `ash cd ComfyUI/custom_nodes git clone https://github.com/1038lab/ComfyUI-Depth-Anything-3.git git clone https://github.com/1038lab/ComfyUI-Lotus-2.git` Both repositories come with tested drag-and-drop workflows in their example\_workflows directory: * DA3\_video\_depth.json: Tested batch video depth workflow. * Lotus-2.json: Side-by-side comparison workflow for both SD2.1 and FLUX engines. Check them out and let us know how they work in your MiniMax and video generation pipelines!
document fetched 2026-10-06 00:39 · canonical https://reddit.com/r/comfyui/comments/1wycdex/release_the_geometry_duo_for_comfyui_clean_white