Precision Image Creation: Harnessing the Power of Dual ControlNets
1. Workflow Overview
This workflow is a FLUX-based precision image generation pipeline, specialized for indoor scenes (e.g., staircase entrances, carpets, furniture). It uses dual ControlNets (Canny Edge + Depth) for composition control, enhances details with Lora models, and outputs HD images via UltimateSD upscaling.
2. Core Models
Model Name | Function | Source/Installation |
|---|---|---|
Base Algorithm_F.1 | Main model (FP8 optimized) | Manual download to |
FLUX.1-ControlNet (Canny/Depth) | Dual ControlNet for构图控制 | Download |
4x-UltraSharp | Image super-resolution | Install via |
3. Key Nodes
Node Name | Function | Installation |
|---|---|---|
BaiduTranslateNode | Auto-translates prompts (ZH→EN) | Manual custom node install |
FluxGuidance | FLUX architecture conditioning | Built-in |
UltimateSDUpscale | Tile-based upscale + detail repair | Install via |
AIO_Preprocessor | All-in-one preprocessor (Canny/Depth) | Requires |
4. Workflow Groups
Group 1: Model Loading
Loads base model, VAE, dual CLIP (clip_l + t5xxl)
Input: None | Output: Model/CLIP/VAE objects
Group 2: Conditional Control
Linear Control: Canny Edge + ControlNet
Depth Control: Depth Map + ControlNet
Input: Reference image | Output: Conditioned Latent
Group 3: Image Generation
KSampler → VAEDecode → UltimateSD Upscale
Input: Latent/ControlNet conditions | Output: HD image (1024x1024)
5. Inputs & Outputs
Inputs:
Reference image: 768x1024 PNG (e.g.,
2 (2).png)Prompts: Supports ZH/EN (auto-translated)
Lora weights:
Entry Mat Style(strength 1.0) +Indoor Realistic Render(strength 0.3)
Output:
4x upscaled image (with metadata)
6. Notes
⚠️ VRAM: Minimum 12GB (peaks during UltimateSD tiling)
⚠️ Dependencies: Requires ComfyUI-Impact-Pack and rgthree extensions
⚠️ Model Paths:
ControlNet models in
models/controlnetLora models in
models/loras