---

name: advanced-variants description: "Advanced denoising-diffusion-pytorch variants: Karras UNets, continuous-time, EDM, simple diffusion, learned/weighted objectives, and attention backend compatibility." disable-model-invocation: true metadata: disco-role: operating license: MIT

Advanced Variants

Use this sub-skill when a task goes beyond common DDPM/DDIM image or 1D sequence paths and asks for Karras magnitude-preserving UNets, continuous-time or v-parameterized diffusion, EDM / elucidated diffusion, simple diffusion UViT, learned variance, weighted objectives, video-shaped Karras 3D models, or flash / SDPA attention behavior in denoising-diffusion-pytorch 2.3.1.

Route elsewhere

• Common 2D image Unet, GaussianDiffusion, folder Trainer, FID, DDIM, or RePaint work: ../image-diffusion/SKILL.md.
• Common 1D sequence Unet1D, GaussianDiffusion1D, Dataset1D, or Trainer1D work: ../sequence-diffusion/SKILL.md.
• Classifier-free guidance, external classifier cond_fn, or XMWrapper: ../conditioning-guidance/SKILL.md.

What this sub-skill covers

• Karras architectures: KarrasUnet, KarrasUnet1D, KarrasUnet3D, and InvSqrtDecayLRSched.
• Advanced diffusion wrappers: ContinuousTimeGaussianDiffusion, VParamContinuousTimeGaussianDiffusion, ElucidatedDiffusion, LearnedGaussianDiffusion, and WeightedObjectiveGaussianDiffusion.
• Simple diffusion APIs: denoising_diffusion_pytorch.simple_diffusion.UViT and its GaussianDiffusion.
• Attention compatibility: Attend, flash_attn constructor flags, PyTorch SDPA behavior, and CPU/CUDA expectations.
• A bundled tiny smoke helper: scripts/smoke_advanced_variants.py.

Required reference order

1. Read references/api-reference.md for import paths, signatures, tensor shapes, and constructor constraints.
2. Read references/workflows.md for safe setup recipes and cross-skill routing.
3. Read references/compatibility.md before changing devices, enabling flash_attn, using Karras 3D downsampling, or selecting a CPU/CUDA verification plan.
4. Use references/troubleshooting.md when an assertion mentions sinusoidal conditioning, self-conditioning, out_dim, DDIM, simple diffusion noise dimensions, Karras video divisibility, or SDPA.

Fast verification

From the generated skill root:

python sub-skills/advanced-variants/scripts/smoke_advanced_variants.py --quick --device cpu

Add --include-3d only when a tiny video-shaped Karras 3D check is desired.