🖼️ AI Image generation

Tools to run existing models

• AUTOMATIC1111 stable-diffusion-webui
  • Really simple to use, friendly interface but very limited
• ComfyUI
  • Way complex, hard configuration but large community and content

Sites to get existing workflows and loras

• runninghub
• openart
• civitai

Good models

• Wan 2.2

  • I2V = Image to Video
  • T2V = Text to Video

• Flux dev

  • T2I = Text to Image

Configurations that can really improve the generation

• Steps = Number of denoising iterations.

  • The number could variate depending the model this is how it works
    • Low -> fast, but less detail.
    • Medium -> sweet spot.
    • High -> can improve realism, but diminishing returns (and slower). Too high sometimes makes images look "overcooked."
  • The number can vary between (1 and N)

• Sampler = The algorithm that controls how noise is removed step by step.

  • Euler a -> fast, creative, less deterministic.
  • DPM++ 2M Karras -> smooth, detailed, highly recommended.
  • DDIM -> older, predictable, fast for testing.

• Seed = The random number that initializes noise.

  • Same seed + prompt + config = exact same image.
  • Changing seed = different variation.
  • Good for reproducing or exploring variations systematically.

• CFG = Classifier-Free Guidance.

  • Used to control how strongly the model follows your text prompt versus just producing random plausible images.
  • The number could variate depending the model this is how it works
    • Low CFG -> The model has a lot of freedom.
      • Output may look more "natural" but not match your text closely.
    • Medium CFG -> Balanced. This is the sweet spot for most models.
    • High CFG -> Model is forced to follow the prompt exactly.
      • Can cause overbaked, distorted, or ugly results (like too many extra fingers, harsh edges).
  • The number can vary between (1 and N)

• Negative Prompt = As the name say defines what you don’t want.

  • Super important for Stable Diffusion!
  • Example prompts: low quality, blurry, bad anatomy, extra fingers, text, watermark
    • It is common to find it in chinese characters here
      • Training Artifacts
      • Prompt Embeddings
      • Copy-Paste Culture from CivitAI
      • "Noise" Tokens for Prompt Weighting

How generative models work

• Overview
  • https://www.youtube.com/watch?v=F1X4fHzF4mQ
  • https://www.youtube.com/watch?v=iv-5mZ_9CPY

Next step

• Deep understanding about diffusion model
  • https://github.com/CompVis/stable-diffusion/blob/main/main.py