Glossary
ControlNet Explained
How ControlNet works and which control types — pose, depth, canny, normal map — to use for different creative goals.
ControlNet is a neural network architecture (introduced by Lvmin Zhang et al., 2023) that attaches to a base diffusion model and accepts a spatial conditioning signal — a pose skeleton, depth map, edge map, or normal map — in addition to the text prompt. This lets you control the composition and structure of the output precisely, overriding the model's default tendency to invent its own layout.
Control types and their use cases: Canny Edge detects sharp edges in a reference image and re-renders them in the new style; good for architecture and product shots. Depth Map preserves the 3D depth structure of a scene; useful for scene composition. OpenPose/DWPose extracts human body skeleton keypoints; essential for character poses. Normal Map encodes surface orientation; best for lighting and material consistency. Scribble converts rough sketches into detailed renders.
ControlNet weight: most UIs expose a weight slider (0.5–1.5). At 1.0 the control is respected strictly; at 0.5 the control is a loose suggestion that the prompt can override. Reducing weight is useful when you want pose guidance but still want the model to interpret the scene freely.
ControlNet start/end steps: advanced UIs let you apply control only for the first N% of denoising steps and then let the model freestyle for the remainder. This tends to produce more natural-looking images while still respecting the structural constraint.
Combining multiple ControlNets: you can stack multiple control signals — for example, pose + depth — and weight them independently. This is how Synexa's pose-clone feature works: it extracts pose from your reference image and applies it to your LoRA character's body.
Frequently Asked Questions
- What is the most useful ControlNet type for AI influencer content?
- OpenPose/DWPose — it lets you copy a specific body pose from any reference photo and apply it to your character.
- Does ControlNet slow down generation?
- Yes, slightly — each control type adds an additional forward pass. On modern hardware the overhead is 10–30% per control type.
- Can I use ControlNet with Flux?
- Yes. Flux-compatible ControlNet models (flux-controlnet-canny, flux-controlnet-depth) are available, though the ecosystem is smaller than for SD. Synexa uses Flux ControlNet for pose cloning.
- What is IP-Adapter and how does it differ from ControlNet?
- IP-Adapter uses a reference image as a style/identity condition encoded through the image encoder (CLIP), rather than as a spatial map. It controls style and content coherence rather than structure — complementary to, not a replacement for, ControlNet.