メニュー
Accelerate video generation through dual-expert consistency distillation, using separate denoisers for semantic layout/motion and detail refinement to resolve conflicting optimization gradients.
デフォルトでは、最初にソースを確認する Prompt が選択されています。直接コマンドに切り替えるか、ローカルコピーをダウンロードすることもできます。
インストールを決める前に、SKILL.md と SkillsMP に表示されている付属ファイルをお読みください。
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
直接コマンドでは確認用 Prompt が省略されます。実行前にソースを確認してください。
npx skills add https://github.com/ADu2021/skillXiv --skill dcm-dual-expert-consistencyコマンドは1行のまま表示されます。コピー前に横へスクロールして全体を確認してください。
ローカルで確認しますか？SkillsMP が現在取得できるファイルをダウンロードできます。
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.
SOC 職業分類に基づく
SKILL.md を表示中
| name | dcm-dual-expert-consistency |
| title | DCM: Dual-Expert Consistency Model for Efficient and High-Quality Video Generation |
| version | 0.0.2 |
| engine | skillxiv-v0.0.2-claude-opus-4.6 |
| license | MIT |
| url | https://arxiv.org/abs/2506.03123 |
| keywords | ["video-generation","consistency-distillation","dual-experts","efficiency","diffusion-models"] |
| description | Accelerate video generation through dual-expert consistency distillation, using separate denoisers for semantic layout/motion and detail refinement to resolve conflicting optimization gradients. |
Consistency distillation for video synthesis faces a fundamental optimization conflict: early diffusion steps require rapid semantic changes (layout, motion) while later steps need gradual detail refinement. A single student model cannot simultaneously optimize for both dynamics, creating conflicting gradients. DCM solves this by employing two specialized expert denoisers—one for semantic/motion reasoning, one for detail synthesis—enabling efficient 4-step video generation while maintaining quality near 50-step original models.
The approach maintains parameter efficiency through LoRA adapters and frozen semantic experts, achieving VBench scores of 83.83 (vs. 80.33 baseline) with 92% latency reduction.
# Expert training specification
def create_dual_expert_curriculum(diffusion_model, train_data):
"""
Partition training into high-noise (semantic) and low-noise (detail) phases.
Each expert specializes in different denoising dynamics.
"""
semantic_expert = copy_model(diffusion_model) # Will be frozen
detail_expert = copy_model(diffusion_model) # Will be adapted with LoRA
# High-noise timesteps for semantic learning (t: 1-25, high sigma)
semantic_batch = filter_timesteps(train_data, timestep_range=(1, 25))
# Low-noise timesteps for detail learning (t: 26-50, low sigma)
detail_batch = filter_timesteps(train_data, timestep_range=(26, ))
{
: semantic_expert,
: detail_expert,
: semantic_batch,
: detail_batch
}
def temporal_coherence_loss(semantic_expert, batch_frames, target_frames):
"""
Loss for semantic expert to maintain consistent motion across frames.
Ensures optical flow consistency and frame alignment.
"""
# Predict denoised output
denoised = semantic_expert(batch_frames, timestep=high_noise)
# Compute optical flow between consecutive frames
flow_target = compute_optical_flow(target_frames)
flow_denoised = compute_optical_flow(denoised)
# L2 loss on flow consistency
flow_loss = torch.nn.functional.mse_loss(flow_denoised, flow_target)
# Frame alignment loss for temporal coherence
alignment_loss = 0.0
for i in range(len(denoised) - 1):
# Warp frame[i+1] using flow[i] and compare to frame[i]
warped = warp_frame(denoised[i+1], flow_denoised[i])
alignment_loss += torch.nn.functional.mse_loss(warped, denoised[i])
total_coherence_loss = 0.7 * flow_loss + 0.3 * alignment_loss
return total_coherence_loss
def detail_expert_loss(detail_expert, detail_discriminator, batch_frames, target_frames):
"""
Adversarial loss for detail expert to synthesize high-frequency details.
Combines reconstruction fidelity with feature matching.
"""
# Forward through detail expert
denoised = detail_expert(batch_frames, timestep=low_noise)
# Reconstruction loss: match target quality
recon_loss = torch.nn.functional.mse_loss(denoised, target_frames)
# Adversarial loss: fool discriminator
fake_logits = detail_discriminator(denoised)
adversarial_loss = -torch.log(fake_logits + 1e-8).mean()
# Feature matching: align intermediate representations with target
# Extract features from intermediate layers of discriminator
fake_features = extract_features(detail_discriminator, denoised)
real_features = extract_features(detail_discriminator, target_frames)
feature_loss = 0.0
for fake_feat, real_feat in zip(fake_features, real_features):
feature_loss += torch.nn.functional.l1_loss(fake_feat, real_feat)
# Combined objective: reconstruction + adversarial + features
total_loss = 0.5 * recon_loss + 0.3 * adversarial_loss + 0.2 * feature_loss
return total_loss
def apply_lora_to_detail_expert(detail_expert, lora_rank=8):
"""
Add LoRA adapters to detail expert while keeping semantic expert frozen.
Maintains parameter efficiency—only ~1-5% additional parameters.
"""
for name, module in detail_expert.named_modules():
if isinstance(module, torch.nn.Linear) and 'attn' in name:
# Add LoRA matrices to attention layers
module.lora_A = torch.nn.Linear(module.in_features, lora_rank, bias=False)
module.lora_B = torch.nn.Linear(lora_rank, module.out_features, bias=False)
# Original forward: out = W @ x
# LoRA forward: out = W @ x + (B @ A @ x) * scale
original_forward = module.forward
def forward_with_lora(x):
out = original_forward(x)
lora_out = module.lora_B(module.lora_A(x))
return out + lora_out * (2.0 / lora_rank) # Scaling factor
module.forward = forward_with_lora
# Freeze semantic expert parameters
for param in detail_expert.parameters():
if 'lora' not in param.__dict__:
param.requires_grad = False
def dual_expert_denoising_step(semantic_expert, detail_expert,
noisy_frames, timestep, context):
"""
Route denoising to appropriate expert based on noise level.
Semantic expert: high noise (early steps)
Detail expert: low noise (late steps)
"""
sigma = get_sigma_from_timestep(timestep)
if sigma > threshold_sigma: # High noise → semantic expert
denoised = semantic_expert(noisy_frames, timestep, context)
else: # Low noise → detail expert
denoised = detail_expert(noisy_frames, timestep, context)
return denoised
When to Apply:
Setup Requirements:
Performance Expectations:
Key Tuning Parameters:
Training Strategy:
Common Pitfalls:
Implemented on HunyuanVideo base model. Achieves VBench 83.83 with 4-step inference. Training uses 8 A100 GPUs with mixed precision. Validated through human evaluation (82.67% preference) and automated metrics (PSNR, temporal consistency scores).