---

name: conditioning-guidance description: "Use denoising-diffusion-pytorch conditioning and guidance APIs: classifier-free guidance, classifier-gradient cond_fn guidance, and XMWrapper multi-candidate loss." disable-model-invocation: true metadata: disco-role: operating license: MIT

Conditioning Guidance

Use this sub-skill for conditioning and guidance surfaces in denoising-diffusion-pytorch: classifier-free guidance, class-label sampling controls, external classifier-gradient cond_fn guidance, and XMWrapper multi-candidate loss selection.

Route here when

• The task mentions classifier-free guidance, class conditioning, num_classes, class labels, classes, cond_drop_prob, cond_scale, rescaled_phi, or CFG++ / use_cfg_plus_plus.
• The task asks for an external classifier cond_fn, guidance_kwargs, classifier-gradient sampling, label-conditioned classifier logits, or guidance shape/device debugging.
• The task asks about XMWrapper, explorative / multi-candidate loss, candidates, max_batch_size, random_time_method, or random_time_kwarg.

Route elsewhere for base setup

• 2D image Unet, GaussianDiffusion, image folders, Trainer, FID, DDIM/DDPM basics, interpolation without classes, or RePaint setup: ../image-diffusion/SKILL.md.
• 1D tensors, Unet1D, GaussianDiffusion1D, Dataset1D, Trainer1D, sequence layout, sequence sampling, or channel_first: ../sequence-diffusion/SKILL.md.
• Karras models, continuous-time wrappers, learned variance, weighted objectives, simple diffusion, flash/SDPA attention, or other advanced variants: ../advanced-variants/SKILL.md.

Operating map

1. Read references/api-reference.md for import paths, signatures, tensor contracts, and guidance-specific semantics.
2. Read references/workflows.md for safe recipes: CFG loss/sample, classifier-gradient cond_fn, XMWrapper, and candidate chunking.
3. Read references/troubleshooting.md for label shape errors, CFG schedule assertions, guidance strength issues, cond_fn gradient failures, and XMWrapper failures.
4. Run scripts/smoke_conditioning_guidance.py after installing the package to verify tiny CPU or CUDA guidance checks without training, downloads, or data files.

Safe defaults

For quick local checks, keep models tiny: dim=8, dim_mults=(1,), channels=1, image_size=8 or seq_length=8, timesteps=8, sampling_timesteps=4, random tensors in [0, 1], candidates=2, and max_batch_size no larger than batch * candidates.