Latent Diffusion: Classifier-free Guidance
Advanced
The problem, from first principles
Classifier-free guidance sharpens how well samples match the text prompt. The model is run twice, once with and once without the condition, and the difference is amplified by a guidance weight.
From theory to code
Implement cfg_combine(eps_u, eps_c, w), returning the unconditional prediction pushed toward the conditional one by weight w.
Constraints
wcan exceed one for stronger guidance.
Hints
Hint 1
Compute the difference eps_c - eps_u, scale it by w, and add to eps_u.
Keywords
research-paperscomputer-visiondiffusionlatent
Python
Run your code to see results
No results yet
Run your code to check it against the examples, or submit to run every test.