AI Image Generation

Local Generation with Open Models

35 分钟

Running open models locally gives unlimited generations, privacy and deep control, at the cost of setup and hardware.

Hardware

A GPU with 8 GB of VRAM can run many models with optimisations; 12–24 GB is comfortable for larger ones. Apple Silicon Macs also work with suitable software. Cloud GPUs are an option if you lack hardware.

Interfaces

  • ComfyUI: node-based workflows, very flexible and shareable.
  • Automatic1111 / Forge: classic web interfaces with many extensions.
  • InvokeAI: polished interface with a canvas for editing.

Download models only from reputable sources, prefer safetensors files and check each model's licence.

Key settings

Setting Effect
Steps more denoising steps: more detail, slower (diminishing returns)
CFG / guidance scale how strictly the prompt is followed; too high looks harsh
Sampler / scheduler denoising algorithm; affects style and speed
Seed reproducibility; same seed + settings = same image
Resolution use the model's native resolution, then upscale

Control

  • ControlNet and similar guide composition with edges, depth maps, poses or sketches, so the layout is exactly what you want.
  • LoRAs add styles, characters or products.
  • IP-Adapter-style references transfer the look of a reference image.

Workflows as assets

Save ComfyUI workflows as files in version control. A reusable "product on pedestal, studio lighting" workflow can produce consistent images for a whole catalogue.

Exercise

Install ComfyUI (or a hosted equivalent), generate an image with a fixed seed, then use a ControlNet depth or edge map from a photo you took to control composition. Save the workflow and reproduce the image exactly.

Check yourself

  • What does the seed control?
  • What does ControlNet add?
  • Why save workflows as files?

你的进度只保存在此浏览器中。

提交审核（10 积分）