advanced-variants
Advanced denoising-diffusion-pytorch variants: Karras UNets, continuous-time, EDM, simple diffusion, learned/weighted objectives, and attention backend compatibility.
- License 许可
- MIT
- 最近更新
- 2026-10-08
- Stars
- ★ 11
- Forks
- 3
01 前言 Front Matter
- name
- advanced-variants
- license
- MIT
- metadata
- { "disco-role": "operating" }
- description
- Advanced denoising-diffusion-pytorch variants: Karras UNets, continuous-time, EDM, simple diffusion, learned/weighted objectives, and attention backend compatibility.
- disable-model-invocation
- true
02 资源文件 4
- 参考文档references/api-reference.md3.7 KB
- 参考文档references/compatibility.md3.5 KB
- 参考文档references/troubleshooting.md2.9 KB
- 参考文档references/workflows.md4.9 KB
03 SKILL.md 正文
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, folderTrainer, FID, DDIM, or RePaint work: ../image-diffusion/SKILL.md. - Common 1D sequence
Unet1D,GaussianDiffusion1D,Dataset1D, orTrainer1Dwork: ../sequence-diffusion/SKILL.md. - Classifier-free guidance, external classifier
cond_fn, orXMWrapper: ../conditioning-guidance/SKILL.md.
What this sub-skill covers
- Karras architectures:
KarrasUnet,KarrasUnet1D,KarrasUnet3D, andInvSqrtDecayLRSched. - Advanced diffusion wrappers:
ContinuousTimeGaussianDiffusion,VParamContinuousTimeGaussianDiffusion,ElucidatedDiffusion,LearnedGaussianDiffusion, andWeightedObjectiveGaussianDiffusion. - Simple diffusion APIs:
denoising_diffusion_pytorch.simple_diffusion.UViTand itsGaussianDiffusion. - Attention compatibility:
Attend,flash_attnconstructor flags, PyTorch SDPA behavior, and CPU/CUDA expectations. - A bundled tiny smoke helper: scripts/smoke_advanced_variants.py.
Required reference order
- Read references/api-reference.md for import paths, signatures, tensor shapes, and constructor constraints.
- Read references/workflows.md for safe setup recipes and cross-skill routing.
- Read references/compatibility.md before changing devices, enabling
flash_attn, using Karras 3D downsampling, or selecting a CPU/CUDA verification plan. - 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.
04 同仓库更多 Skills
obsidian-clipper-template-creator
Guide for creating templates for the Obsidian Web Clipper. Use when you want to create a new clipping template, understand available variables, or format clipped content.
obsidian-clipper-template-creator
Guide for creating templates for the Obsidian Web Clipper. Use when you want to create a new clipping template, understand available variables, or format clipped content.
python-api
Programmatic Object-Detection-Metrics API guidance for BoundingBox, BoundingBoxes, Evaluator, enums, coordinate conversion, AP/mAP semantics, and source-style import diagnostics.
file-evaluation
Folder-based PASCAL VOC AP/mAP evaluation from ground-truth and detection text files.
object-detection-metrics
Guides legacy Object-Detection-Metrics PASCAL VOC AP/mAP evaluation from detection text folders and direct Python API classes.
nx-workspace-patterns
Configure and optimize Nx monorepo workspaces. Use when setting up Nx, configuring project boundaries, optimizing build caching, or implementing affected commands.