Esta nota todavía no está traducida, así que se muestra la fuente en inglés.
Text-to-image conditioning and CFG
Text-to-image generation works by steering a visual denoising process with a text representation. The prompt does not draw the image; it biases each denoising step toward visual features associated with the text.
Conditioning path
- A tokenizer splits the prompt into text tokens.
- A text encoder turns tokens into embeddings.
- The denoising network attends to those embeddings while predicting noise.
- Optional negative prompts describe features to avoid.
- Other conditioning can include image, mask, depth, pose, edge map, style, or reference subject.
Classifier-free guidance
Classifier-free guidance compares two denoising predictions:
| Prediction | Meaning |
|---|---|
| Conditional | what the model predicts given the prompt |
| Unconditional | what the model predicts with no or empty prompt |
| Guided | push away from unconditional and toward prompt-conditioned direction |
Higher guidance can improve prompt adherence but may reduce realism, diversity, and color balance.
Prompt control levers
- Specific subject, medium, composition, lighting, and style.
- Negative prompt for recurring unwanted artifacts.
- Seed for reproducibility.
- Sampling steps and sampler choice.
- Guidance scale for adherence versus naturalness.
- Image size and aspect ratio.
Pitfall
Prompting is only one control channel. Precise layout, pose, identity, and product consistency usually need structural conditioning or fine-tuned adapters, not more adjectives.
Connects to: prompt anatomy · controlling image generation · tokenization