Depth LoRA

Convert depth maps into high-quality images

Introduction

This is a depth LoRA that converts depth maps into high-quality images. It was trained in multiple stages on a variety of datasets to balance overall structural accuracy with fine-grained detail reconstruction.

HF download:

https://huggingface.co/lrzjason/Krea2-Lora-Collection/tree/main/loras

EditUtils

https://github.com/lrzjason/ComfyUI-EditUtils/tree/main

Training Details

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

Stage 1 — 512 Resolution Pretraining

Trained on a 2,000-pair subset sampled from a public depth dataset, with all captions regenerated by a VLM (VLM re-caption) to improve text-image alignment. This stage used a constant learning rate of 1e-4 to establish a solid depth-structure prior.

Stage 2 — 1536 High-Resolution Refinement

Continued training on a 1536-resolution depth dataset, focusing on detail reconstruction (edges, thin structures, and fine geometry). This stage switched to a cosine learning rate schedule for more stable convergence.

Stage 1 512 res, 2,000 pairs (public depth dataset + VLM re-caption), constant LR 1e-4Stage 2 1536 res depth dataset, cosine LR scheduleTotal Epochs50+

Usage

⚠️ This LoRA must be used together with the EditUtils plugin:
👉 https://github.com/lrzjason/ComfyUI-EditUtils

1. Install the ComfyUI-EditUtils plugin.

2. Load this LoRA in your workflow.

3. Feed in the source image to obtain the corresponding depth map.

Notes

• Works better on 10240- 2048 with long captions; with higher-resolution inputs yielding richer depth details.

• Will not work as intended without the EditUtils plugin.

RH Workflow

Experience link: https://www.runninghub.ai/post/2082077636234313729/?inviteCode=rh-v1279

📬 Contact

• 𝕏 Twitter: @Lrzjason

• 📧 Email: [email protected]

• 💬 QQ Group: 866612947

• 💬 WeChat ID: fkdeai

• 🎨 CivitAI: xiaozhijason