菜单
Replaces iterative diffusion with single-step decoding for image compression. Combines VAE latents with fidelity guidance and rate annealing training. Achieves 20× decoding speedup with improved perceptual quality.
默认使用会先检查来源的 Prompt；你也可以切换为直接命令，或下载本地副本。
决定是否安装前，请先阅读 SKILL.md，以及 SkillsMP 当前展示的配套文件。
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
直接命令不会经过审查 Prompt；运行前请先检查来源。
npx skills add https://github.com/ADu2021/skillXiv --skill sodec-diffusion-compression命令会保持在同一行。复制前请横向滚动并检查完整内容。
想先保存到本地？可下载 SkillsMP 当前能够提供的文件。
Convert arXiv papers into ready-to-use agent skills using category-aware extraction. First classifies the paper into one or more of 11 research categories, then applies a specialized extraction pipeline for each category — because different types of papers produce different types of usable knowledge. A single paper can yield multiple skills if it spans categories. Use this skill whenever the user wants to turn a paper into a skill, extract practical techniques from research, build a skill library from papers, convert arXiv papers into reusable agent instructions, or batch-process multiple papers into skills. Also trigger when someone asks about extracting actionable knowledge from papers, making research practical for LLM agents, or systematically converting academic contributions into structured agent capabilities.
Establish regret bounds for behavior cloning with discretized actions combining statistical error and quantization error terms. Prove smoothness requirements for safe quantizer design, show that learning-based quantizers fail these requirements, and propose model-based augmentation to reduce error dependence from H² to H.
Dynamically allocate LoRA ranks per-layer during fine-tuning instead of using fixed uniform ranks. Learn optimal rank for each layer and subject via variational framework with discretized exponential distribution, reducing memory footprint while maintaining fidelity and text-alignment.
基于 SOC 职业分类
正在显示 SKILL.md
| name | sodec-diffusion-compression |
| title | SODEC - Steering One-Step Diffusion for Fast Image Compression |
| version | 0.0.2 |
| engine | skillxiv-v0.0.2-claude-opus-4.6 |
| license | MIT |
| url | https://arxiv.org/abs/2508.04979 |
| keywords | ["image-compression","diffusion-models","generative-models","single-step-decoding"] |
| description | Replaces iterative diffusion with single-step decoding for image compression. Combines VAE latents with fidelity guidance and rate annealing training. Achieves 20× decoding speedup with improved perceptual quality. |
Traditional generative compression uses iterative diffusion refinement that requires dozens of denoising steps, making inference prohibitively slow. SODEC demonstrates that information-rich latent representations eliminate the need for iteration. By combining pre-trained VAE encodings with a fidelity guidance module and rate-annealing training strategy, the approach achieves single-step decoding with 20× speedup while maintaining or improving perceptual quality.
Create VAE-based latent representation that captures sufficient information for reconstruction.
import torch
import torch.nn as nn
from typing import Tuple
class InformationRichVAE(nn.Module):
"""
VAE encoder that creates information-rich latents for compression.
"""
def __init__(self, in_channels=3, latent_dim=16, hidden_dim=128):
super().__init__()
self.latent_dim = latent_dim
# Encoder: image -> latent
self.encoder = nn.Sequential(
nn.Conv2d(in_channels, hidden_dim, 4, stride=2, padding=1),
nn.ReLU(),
nn.Conv2d(hidden_dim, hidden_dim * 2, , stride=, padding=),
nn.ReLU(),
nn.Conv2d(hidden_dim * , hidden_dim * , , stride=, padding=),
nn.ReLU(),
nn.AdaptiveAvgPool2d((, )),
nn.Flatten()
)
.fc_mu = nn.Linear(hidden_dim * * , latent_dim)
.fc_logvar = nn.Linear(hidden_dim * * , latent_dim)
.fc_decode = nn.Linear(latent_dim, hidden_dim * * * )
.decoder = nn.Sequential(
nn.ConvTranspose2d(hidden_dim * , hidden_dim * , , stride=, padding=),
nn.ReLU(),
nn.ConvTranspose2d(hidden_dim * , hidden_dim, , stride=, padding=),
nn.ReLU(),
nn.ConvTranspose2d(hidden_dim, in_channels, , stride=, padding=),
nn.Tanh()
)
() -> [torch.Tensor, torch.Tensor]:
h = .encoder(x)
mu = .fc_mu(h)
logvar = .fc_logvar(h)
mu, logvar
() -> torch.Tensor:
std = torch.exp( * logvar)
eps = torch.randn_like(std)
mu + eps * std
() -> torch.Tensor:
h = .fc_decode(z)
h = h.view(h.shape[], -, , )
x_recon = .decoder(h)
x_recon
() -> [torch.Tensor, torch.Tensor, torch.Tensor]:
mu, logvar = .encode(x)
z = .reparameterize(mu, logvar)
x_recon = .decode(z)
x_recon, mu, logvar
Create module that ensures generated images remain faithful to original.
class FidelityGuidanceModule(nn.Module):
"""
Steers generation toward fidelity to original image.
"""
def __init__(self, hidden_dim=128):
super().__init__()
# Learned guidance function
self.guidance_net = nn.Sequential(
nn.Linear(16, hidden_dim), # latent_dim
nn.ReLU(),
nn.Linear(hidden_dim, hidden_dim),
nn.ReLU(),
nn.Linear(hidden_dim, 3), # Guidance for RGB channels
)
def compute_guidance(
self,
latent: torch.Tensor,
generated_image: torch.Tensor,
original_image: torch.Tensor,
guidance_strength: float = 1.0
) -> torch.Tensor:
"""
Compute fidelity guidance signal.
Args:
latent: Latent code
generated_image: Generated image from latent
original_image: Original image to stay faithful to
guidance_strength: Strength of fidelity guidance
Returns:
Guidance tensor to apply to generation
"""
# Compute pixel-level difference
pixel_diff = (generated_image - original_image).abs().mean(dim=(2, 3)) # [batch, channels]
# Network-based guidance
net_guidance = self.guidance_net(latent)
# Combine pixel and network guidance
guidance = pixel_diff + guidance_strength * net_guidance
return guidance
def apply_guidance(
self,
generated_image: torch.Tensor,
guidance: torch.Tensor,
learning_rate: float = 0.01
) -> torch.Tensor:
"""
Apply guidance to adjust generation.
Args:
generated_image: Current generated image
guidance: Guidance signal
learning_rate: Strength of adjustment
Returns:
Adjusted image
"""
# Gradient-based adjustment toward fidelity
guided_image = generated_image - learning_rate * guidance.view(-1, 3, 1, 1)
return guided_image
Create decoder that generates image in one step from latent.
class SingleStepDecoder(nn.Module):
"""
Direct latent-to-image decoding without iterative refinement.
"""
def __init__(self, latent_dim=16, hidden_dim=256):
super().__init__()
# Direct mapping from latent to image
self.latent_to_image = nn.Sequential(
nn.Linear(latent_dim, hidden_dim * 8),
nn.ReLU(),
nn.Linear(hidden_dim * 8, hidden_dim * 4),
nn.ReLU(),
nn.Linear(hidden_dim * 4, 3 * 256 * 256), # 256x256 RGB image
nn.Tanh()
)
# Confidence estimation network
self.confidence_net = nn.Sequential(
nn.Linear(latent_dim, hidden_dim),
nn.ReLU(),
nn.Linear(hidden_dim, 1),
nn.Sigmoid()
)
def forward(self, latent: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
"""
Generate image from latent in single step.
Args:
latent: Latent code [batch, latent_dim]
Returns:
(generated_image, confidence)
"""
# Direct generation
image_flat = self.latent_to_image(latent)
image = image_flat.view(-1, 3, 256, 256)
# Confidence score
confidence = self.confidence_net(latent)
return image, confidence
def decode_batch(
self,
latents: torch.Tensor,
apply_guidance: bool = False,
guidance_module: FidelityGuidanceModule = None,
original_images: torch.Tensor = None
) -> Tuple[torch.Tensor, torch.Tensor]:
"""
Decode batch of latents with optional guidance.
Args:
latents: Batch of latent codes
apply_guidance: Whether to apply fidelity guidance
guidance_module: Guidance module if applying guidance
original_images: Original images for guidance
Returns:
(generated_images, confidences)
"""
images, confidences = self.forward(latents)
if apply_guidance and guidance_module is not None:
for _ in range(3): # 3 refinement steps with guidance
guidance = guidance_module.compute_guidance(
latents,
images,
original_images,
guidance_strength=0.1
)
images = guidance_module.apply_guidance(images, guidance)
return images, confidences
Create training strategy that progressively tightens compression constraints.
class RateAnnealingTrainer:
"""
Train with rate annealing: progressively increase compression.
"""
def __init__(self, vae, decoder, guidance_module):
self.vae = vae
self.decoder = decoder
self.guidance = guidance_module
def train_step(
self,
images: torch.Tensor,
current_rate: float,
target_rate: float,
num_steps: int
):
"""
Single training step with rate annealing.
Args:
images: Batch of images
current_rate: Current compression rate (bits/pixel)
target_rate: Target compression rate
num_steps: Current step number
Returns:
Loss value
"""
# Encode images
mu, logvar = self.vae.encode(images)
latent = self.vae.reparameterize(mu, logvar)
# Decode
decoded_images, confidence = self.decoder(latent)
# Apply guidance
guided_images = self.guidance.compute_guidance(latent, decoded_images, images)
# Reconstruction loss
recon_loss = torch.mean((decoded_images - images) ** 2)
# Rate loss: encourage lower bitrate as training progresses
# Current rate decreases as we train
rate_schedule = target_rate + (current_rate - target_rate) * (1.0 - num_steps / 10000)
# KL divergence for compression
kl_loss = -0.5 * torch.sum(1 + logvar - mu.pow(2) - logvar.exp())
kl_loss = kl_loss / images.shape[0]
# Annealed loss: increase rate penalty over time
annealing_factor = min(1.0, num_steps / 5000)
rate_loss = annealing_factor * kl_loss
# Total loss
total_loss = recon_loss + 0.1 * rate_loss
return total_loss
def train_epoch(
self,
dataloader,
num_epochs: int = 10,
target_rate: float = 0.5
):
"""
Train for one epoch with rate annealing.
Args:
dataloader: Training data
num_epochs: Number of epochs
target_rate: Target compression rate
"""
optimizer = torch.optim.Adam(
list(self.vae.parameters()) +
list(self.decoder.parameters()) +
list(self.guidance.parameters()),
lr=1e-4
)
total_steps = 0
for epoch in range(num_epochs):
epoch_loss = 0
for batch_images in dataloader:
# Get current rate based on progress
current_rate = 2.0 # Start with 2 bits/pixel
# Training step
loss = self.train_step(
batch_images,
current_rate=current_rate,
target_rate=target_rate,
num_steps=total_steps
)
optimizer.zero_grad()
loss.backward()
optimizer.step()
epoch_loss += loss.item()
total_steps += 1
print(f"Epoch {epoch}: Loss={epoch_loss / len(dataloader):.4f}")
The critical insight is that information-rich VAE latents reduce the need for iterative refinement. By starting from a high-quality latent representation and applying single-step decoding with guidance, the approach maintains quality while achieving massive speedup. Rate annealing prevents early mode collapse during training.
Steering One-Step Diffusion for Fast Image Compression (arXiv:2508.04979)
Replaces iterative diffusion decoding with single-step generation for image compression. Achieves 20× decoding speedup through VAE latents, fidelity guidance, and rate-annealing training while maintaining or improving perceptual quality.