ControlNet教程：用姿态、深度和边缘控制AI图像
文生图无法控制构图。ControlNet可以指定姿态、深度和结构。
ControlNet Tutorial 2026: Control AI Images With Poses, Depth and Edges
ControlNet, released by Lvmin Zhang and Maneesh Agrawala in February 2023, lets you condition image generation on structural inputs: a pose skeleton, a depth map, an edge drawing, or a segmentation mask. It remains the standard technique in 2026 for getting AI images to follow a composition.
What ControlNet Controls
| Condition | Input | Use case |
|---|---|---|
| OpenPose | Skeleton pose image | Keep a character's pose exactly |
| Depth | Depth map | Consistent 3D layout |
| Canny/soft edges | Edge drawing | Match an outline or sketch |
| Segmentation | Labeled mask | Control object placement |
| Normal map | Surface normals | Consistent lighting direction |
Setup (ComfyUI - the 2026 standard)
- Install ComfyUI (123,000 stars)
- Download a ControlNet model (e.g., control_v11p_sd15_openpose) from Hugging Face
- In the workflow, add a ControlNet loader node between the checkpoint and the sampler
The Workflow
Load Checkpoint -> CLIP Text Encode (prompt)
|
Load Image (pose) -> ControlNet Loader -> Apply ControlNet
|
KSampler -> VAE Decode -> Save
The Apply ControlNet node mixes the text prompt and the structural condition. The control_strength parameter (0-1) decides how strictly the structure is followed.
Key Parameters
| Parameter | Effect |
|---|---|
| control_strength | 1.0 = strict structure, 0.3 = loose suggestion |
| start_percent / end_percent | When during denoising the control applies (early = composition, late = details) |
| Model choice | Match ControlNet version to your base model (SD1.5 vs SDXL vs FLUX) |
Getting a Pose Image (OpenPose)
No drawing skills needed: use an OpenPose estimator (e.g., via ComfyUI's pose node or ControlNet's bundled preprocessor) to extract a skeleton from any photo, then reuse that skeleton with a completely different subject.
FAQ
ControlNet vs img2img? img2img changes the whole image; ControlNet changes only the specified structure (pose, edges) while the model fills everything else from the prompt.
Does ControlNet work with FLUX? Yes - Flux ControlNet models (e.g., Flux.1-dev-ControlNet-Canny) appeared through 2025 and are the current standard for FLUX conditioning.
Is ControlNet free? Yes - open source, runs locally in ComfyUI or Automatic1111 (164,300 stars).
