Depth LoRA (Depth LoRA)

Introduction

This is a Depth LoRA that can convert depth maps into high-quality images. The model is trained in stages on multiple datasets, balancing overall structural accuracy and detail restoration capability.

Training Details

The training adopts a two-stage, resolution-progressive strategy:

**First Stage — 512 Resolution Pretraining**
Trained on a subset of 2000 image pairs sampled from public depth datasets, with all annotations regenerated through VLM re-captioning to enhance text-image alignment quality. This stage uses a constant learning rate of `1e 4` to establish a solid depth structure prior.

**Second Stage — 1536 High-Resolution Refinement**
Training continues on a 1536 resolution depth dataset, focusing on improving detail reconstruction (edges, fine structures, and intricate geometry). This stage switches to a cosine learning rate schedule for more stable convergence.

| Item | Parameter |
| | |
| LoRA Rank | xxx |
| First Stage | 512 resolution, 2000 image pairs (public depth dataset, VLM re-captioned), constant learning rate `1e 4` |
| Second Stage | 1536 resolution depth dataset, cosine learning rate schedule |
| Total Training Epochs | 50 epochs |

Usage

⚠️ **This LoRA must be used with the EditUtils plugin:**

👉 https://github.com/lrzjason/ComfyUI EditUtils

1. Install the ComfyUI EditUtils plugin.
2. Load this LoRA into the workflow.
3. Input the original image to obtain the corresponding depth map.

Notes

Supports long sides ranging from 1024 to 2048, adapting to a wide range of input resolutions; the higher the input resolution, the richer the depth details.
If not used in conjunction with the EditUtils plugin, it will not function properly.