ComfyUI.org

Unlock 3D Magic: A Step-by-Step Workflow for Converting 2D Line Art

CN
ComfyUI.org
2025-04-02 09:06:50

1. Workflow Overview

m8zpcmbj5x27u6151kn2f33adc508abca0898a3d9f42653c7dab56323c1167824767a5e92e439a6635d.png

This workflow converts 2D line art into 3D-styled images using ControlNet for sketch control and depth map generation, with tile-based upscaling for HD output. Key stages:

  • Line art preprocessing β†’ 3D style generation β†’ Depth map control β†’ Upscaling

2. Core Models

Model Name

Function

3DMix.fp16

Stable Diffusion 1.5 fine-tuned for 3D cartoon rendering.

ControlNet-lineart

Controls sketch structure (control_v11p_sd15_lineart_fp16.safetensors).

ControlNet-depth

Enhances 3D depth (control_v11f1p_sd15_depth_fp16.safetensors).

4x-UltraSharp

Image super-resolution model (4x-UltraSharp.pth).

3. Key Nodes

Node Name

Purpose

Installation

AIO_Preprocessor

Preprocesses line art/depth maps (supports AnimeLineArt/DepthAnythingV2).

Via ComfyUI Manager.

ControlNetApplyAdvanced

Dual-ControlNet integration (lineart + depth).

Built-in.

UltimateSDUpscale

Tile-based upscaling to avoid VRAM overflow.

Manual install from GitHub.

WD14Tagger

Auto-tags input images to assist prompt generation.

Install via pysssss plugin.

Dependencies:

  • LoRA: Download zhidiao.safetensors (3D style enhancer) to models/loras.

  • Note: Model links are provided in the workflow note (from liblib.art).

4. Workflow Structure

Group Name

Function

Input/Output

Model Loading

Loads 3D base model, ControlNets, and LoRA.

Input: Model paths / Output: Initialized models.

ControlNet

Processes line art and depth maps in parallel.

Input: Sketch image / Output: Conditioning.

Tagging

Generates tags via WD14Tagger and combines with manual prompts.

Input: Image / Output: Combined prompts.

Sampling

Generates 3D images using KSampler (default: 28 steps, Euler).

Input: Conditioning / Output: Latent.

Upscaling

Performs UltimateSDUpscale (2x) with tile-based processing.

Input: Low-res image / Output: HD image.

5. Inputs & Outputs

  • Input Parameters:

    • Image: Upload line art via LoadImage node (supports .webp).

    • Prompt: Default style is clay character (editable via CR Text node).

    • Resolution: Preprocessing fixed at 768x768; final output depends on upscale settings.

  • Output:

    • Saved as PNG (via SaveImage node) to ComfyUI/output.

6. Notes

  1. VRAM: β‰₯8GB GPU recommended (upscaling is VRAM-intensive).

  2. Preprocessors:

    • Use AnimeLineArtPreprocessor for sketches.

    • Use DepthAnythingV2Preprocessor for depth maps.

  3. Debugging: If ControlNet fails, verify model filenames match the JSON.

previous
Minimalist Masterpiece: AI-Powered Workflow for Gradient-Style Art
next
Unlock the Power of Text-to-Video Generation with Aliyun's Wan2.1 Model

Recommend