Skip to main content
Marigold V2 is a family of monocular estimation models from Huawei Bayer Lab that repurpose a diffusion transformer (Qwen-Image Edit 2509) for geometry and intrinsic image estimation. From a single RGB image, the models predict a depth map, surface normals, or albedo, following up on the original Marigold (CVPR 2024) which repurposed Stable Diffusion for the same task. The V2 models ship as LoRAs on top of a shared base diffusion model, which keeps each download small and the workflow structure identical across the three tasks. Compared with classical depth networks, the diffusion-based approach produces sharp, boundary-aware predictions and generalizes to open-domain images without retraining. Typical uses include depth-guided control for image and video generation, 3D scene reconstruction pipelines, relighting, and material editing driven by the estimated albedo. Marigold V2 is supported natively in ComfyUI. Update to the latest version of ComfyUI before starting.
Make sure your ComfyUI is updated.Workflows in this guide can be found in the Workflow Templates. If you canโ€™t find them in the template, your ComfyUI may be outdated.If nodes are missing when loading a workflow, possible reasons:
  1. You are not using the latest ComfyUI version (Nightly version)
  2. Some nodes failed to import at startup

โ€‹
Model Installation

Each workflow uses four files: the shared base diffusion model, plus a task-specific LoRA, VAE, and conditioning embedding. The three tasks use different LoRA, VAE, and conditioning files, so each workflow section below links the exact files it needs. The base diffusion model is shared across all three workflows:

Diffusion Model

qwen_image_edit_2509_int8_convrot.safetensors โ†’ ComfyUI/models/diffusion_models/
If you plan to run all three tasks, download the per-task LoRAs, VAEs, and conditioning embeddings from the workflow sections and store them as shown in each section.

โ€‹
Depth Estimation

Predict a depth map from a single image. The decoded prediction is normalized so that near surfaces appear bright. Marigold V2 depth estimation workflow preview

Run on Comfy Cloud

Open in Comfy Cloud

Download Workflow

Download JSON or search โ€œDepth Estimation: Marigold V2โ€ in Template Library
Input materials Upload this file to the matching LoadImage node:

ev_suv_gold_desert_pavilion.png

LoadImage node ยท ev_suv_gold_desert_pavilion.png

โ€‹
1. Steps to Run

  1. Load your input image in the LoadImage node
  2. Queue the workflow: the base model runs with the depth LoRA, and the Marigold V2 Post-Process node normalizes the decoded prediction
  3. Save or preview the resulting depth map

โ€‹
2. Model Downloads

Depth LoRA

marigold_v2_depth_log_stage2.safetensors โ†’ ComfyUI/models/loras/

Depth VAE

marigold_v2_depth_log_stage2_vae.safetensors โ†’ ComfyUI/models/vae/

Depth conditioning embedding

marigold_v2_depth_conditioning.safetensors โ†’ ComfyUI/models/embeddings/

โ€‹
Surface Normal Estimation

Predict per-pixel surface normals from a single image. The output maps each pixel to the direction the surface faces, useful for relighting and geometry analysis. Marigold V2 surface normal estimation workflow preview

Run on Comfy Cloud

Open in Comfy Cloud

Download Workflow

Download JSON or search โ€œSurface Normal Estimation: Marigold V2โ€ in Template Library
Input materials Upload this file to the matching LoadImage node:

greek_statue_orange_cloth.png

LoadImage node ยท greek_statue_orange_cloth.png

โ€‹
1. Steps to Run

  1. Load your input image in the LoadImage node
  2. Queue the workflow with the normals LoRA loaded
  3. Save or preview the resulting normal map

โ€‹
2. Model Downloads

Normals LoRA

marigold_v2_normals.safetensors โ†’ ComfyUI/models/loras/

Normals VAE

marigold_v2_normals_vae.safetensors โ†’ ComfyUI/models/vae/

Normals conditioning embedding

marigold_v2_normals_conditioning.safetensors โ†’ ComfyUI/models/embeddings/

โ€‹
Albedo Estimation

Predict the albedo (intrinsic color) of a single image, separating material color from lighting and shading. The output is converted to sRGB by the Marigold V2 Post-Process node. Marigold V2 albedo estimation workflow preview

Run on Comfy Cloud

Open in Comfy Cloud

Download Workflow

Download JSON or search โ€œAlbedo Estimation: Marigold V2โ€ in Template Library
Input materials Upload this file to the matching LoadImage node:

flower_with_texture.png

LoadImage node ยท flower_with_texture.png

โ€‹
1. Steps to Run

  1. Load your input image in the LoadImage node
  2. Queue the workflow with the albedo LoRA loaded
  3. Save or preview the resulting albedo image

โ€‹
2. Model Downloads

Albedo LoRA

marigold_v2_albedo.safetensors โ†’ ComfyUI/models/loras/

Albedo VAE

marigold_v2_albedo_vae.safetensors โ†’ ComfyUI/models/vae/

Albedo conditioning embedding

marigold_v2_albedo_conditioning.safetensors โ†’ ComfyUI/models/embeddings/

โ€‹
Marigold V2 Post-Process Node

Every workflow ends in the Marigold V2 Post-Process node, which turns the raw decoded prediction into a viewable image. The prediction option must match the task:
  • depth: normalizes the prediction so near surfaces are bright
  • normals: maps raw values to unit surface normals
  • albedo: converts the prediction to sRGB