Guidance-distilled models like FLUX.1-dev replaced classic dual-pass CFG with an embedded guidance scale parameter (halving compute per step in ComfyUI), though standard CFG and negative prompts remain widely used in Video Generation Models (Sora, Veo, Kling, Wan) and SDXL workflows.
Classifier-Free Guidance (CFG)
Also known as: CFG · Guidance Scale · Negative Prompt
Plain English
The 'prompt adherence slider' in image and video generators. Turning CFG up forces the model to follow your text prompt more strictly—and enables negative prompts—while turning it too high oversaturates or 'burns' the image.
Technical Overview
At each denoising step in Diffusion Models, CFG runs two forward passes—one conditioned on the text prompt and one unconditional (or conditioned on a negative prompt)—and extrapolates along their difference vector scaled by weight w. Because two passes double inference cost, guidance-distilled DiT models like FLUX.1-dev bake guidance directly into a single-pass conditioning input.
Chronological Entries (2)
Ho and Salimans proposed Classifier-Free Guidance, showing that randomly dropping text conditioning during training lets a single diffusion model steer sample fidelity vs. diversity at inference time without an external classifier.
Direct Relations (3)
- Part ofDiffusion Models
- UsesSampling & Schedulers
- Alternative toDistillation & Synthetic Data