메뉴
Unify image understanding and generation with layerwise timestep experts and residual attention reuse, achieving 6x faster inference than comparable unified models while maintaining competitive performance.
기본적으로 소스를 먼저 확인하는 Prompt가 선택됩니다. 직접 명령으로 전환하거나 로컬 사본을 다운로드할 수도 있습니다.
설치 여부를 결정하기 전에 SKILL.md와 SkillsMP에 표시된 보조 파일을 읽어 보세요.
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
직접 명령은 검토 Prompt를 거치지 않습니다. 실행하기 전에 소스를 확인하세요.
npx skills add https://github.com/ADu2021/skillXiv --skill latte-flow-unified-multimodal명령은 한 줄로 유지됩니다. 복사하기 전에 가로로 스크롤해 전체 내용을 확인하세요.
로컬 사본을 원하시나요? SkillsMP에서 현재 제공할 수 있는 파일을 다운로드하세요.
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 직업 분류 기준
SKILL.md 표시 중
| name | latte-flow-unified-multimodal |
| title | LaTtE-Flow: Layerwise Timestep-Expert Flow-based Transformer |
| version | 0.0.2 |
| engine | skillxiv-v0.0.2-claude-opus-4.6 |
| license | MIT |
| url | https://arxiv.org/abs/2506.06952 |
| keywords | ["multimodal models","flow matching","efficient generation","image understanding","unified architecture"] |
| description | Unify image understanding and generation with layerwise timestep experts and residual attention reuse, achieving 6x faster inference than comparable unified models while maintaining competitive performance. |
LaTtE-Flow presents an efficient unified multimodal architecture combining image understanding and generation through flow-matching. The key innovation—Layerwise Timestep Experts—partitions transformer layers into timestep-specific groups, reducing inference complexity from O(L×T') to O(M×T') where M=L/K. Timestep-Conditioned Residual Attention reuses earlier layer computations, enabling 6x faster inference than competing unified models while maintaining competitive performance on both understanding and generation tasks.
import torch
import torch.nn as nn
class LayerwiseTimestepExpert(nn.Module):
"""
Partitions transformer layers into timestep-specific groups.
Each group specializes in specific diffusion timesteps.
Reduces complexity from O(L*T') to O((L/K)*T').
"""
def __init__(self, num_layers, num_experts, hidden_dim, num_heads):
super().__init__()
self.num_layers = num_layers
self.num_experts = num_experts
self.layers_per_expert = num_layers // num_experts
# Partition layers into timestep-expert groups
self.expert_groups = nn.ModuleList()
expert_idx (num_experts):
group_layers = nn.ModuleList()
layer_idx (.layers_per_expert):
layer = TransformerBlock(hidden_dim, num_heads)
group_layers.append(layer)
.expert_groups.append(group_layers)
.timestep_embedding = nn.Sequential(
nn.Linear(, ),
nn.SiLU(),
nn.Linear(, num_experts)
)
():
t_embed = .timestep_embedding(timestep.().unsqueeze(-))
expert_idx = torch.argmax(t_embed, dim=-).item()
normalized_t = ( - timestep.item()) * .num_experts
expert_idx = ((normalized_t), .num_experts - )
selected_expert = .expert_groups[expert_idx]
layer selected_expert:
x = layer(x)
x
(nn.Module):
():
().__init__()
.attention = nn.MultiheadAttention(hidden_dim, num_heads, batch_first=)
.ff = nn.Sequential(
nn.Linear(hidden_dim, * hidden_dim),
nn.GELU(),
nn.Linear( * hidden_dim, hidden_dim)
)
.norm1 = nn.LayerNorm(hidden_dim)
.norm2 = nn.LayerNorm(hidden_dim)
():
attn_out, _ = .attention(x, x, x)
x = x + attn_out
x = .norm1(x)
ff_out = .ff(x)
x = x + ff_out
x = .norm2(x)
x
class TimestepConditionedResidualAttention(nn.Module):
"""
Reuses attention maps from earlier layers, conditioned on timestep.
Enables parameter sharing and computation reuse across layers.
"""
def __init__(self, hidden_dim, num_heads, num_layers):
super().__init__()
self.hidden_dim = hidden_dim
self.num_heads = num_heads
self.num_layers = num_layers
# Store attention maps from earlier layers
self.attention_cache = {}
# Gating mechanism: modulate attention by timestep
self.timestep_gate = nn.Sequential(
nn.Linear(1, 64),
nn.SiLU(),
nn.Linear(64, num_heads) # Per-head gating
)
def forward(self, x, layer_idx, timestep, attention_cache=None):
"""
Forward pass with residual attention reuse.
Layer L reuses attention from layer L-1, gated by timestep.
"""
# Compute local attention
local_attn = self._compute_attention(x) # [batch, seq, num_heads]
# Retrieve cached attention from previous layer
if attention_cache is not None and layer_idx > 0:
cached_attn = attention_cache.get(layer_idx - 1, None)
if cached_attn is not None:
# Compute timestep-dependent gate
t_embed = timestep.float().unsqueeze(-1) # [batch, 1]
gate = torch.sigmoid(self.timestep_gate(t_embed)) # [batch, num_heads]
# Gate:1.0 early (early timesteps favor local computation)
# Gate:0.0 late (late timesteps leverage cached attention)
gate = gate.unsqueeze(1) # [batch, 1, num_heads]
# Blend local and cached attention
blended_attn = gate * local_attn + (1 - gate) * cached_attn
# Store for next layer
attention_cache[layer_idx] = blended_attn
return blended_attn
# Store for next layer
if attention_cache is not None:
attention_cache[layer_idx] = local_attn
return local_attn
def _compute_attention(self, x):
"""Compute multi-head attention weights."""
# Simplified: actual implementation uses full attention mechanism
batch, seq_len, dim = x.shape
scores = torch.matmul(x, x.transpose(-2, -1)) / (dim ** 0.5)
attn_weights = torch.softmax(scores, dim=-1)
return attn_weights
class FlowMatchingGenerator(nn.Module):
"""
Generates images using flow-matching instead of traditional diffusion.
More stable training trajectory than reverse diffusion.
"""
def __init__(self, model_dim, vocab_size=256):
super().__init__()
self.model_dim = model_dim
self.vocab_size = vocab_size
# Transformer for flow prediction
self.flow_predictor = TransformerBlock(model_dim, num_heads=8)
# Output head: predict image tokens
self.output_head = nn.Linear(model_dim, vocab_size)
def forward(self, latent, prompt_embedding, timestep):
"""
Predict flow (vector field) that moves noise towards image.
Flow-matching: directly learn velocity field dX/dt.
"""
# Concatenate latent with prompt
x = torch.cat([latent, prompt_embedding], dim=-1)
# Predict flow
flow = self.flow_predictor(x)
# Predict image tokens from flow
logits = self.output_head(flow)
return logits
def generate(self, prompt_embedding, num_steps=50, latent_dim=512):
"""
Generate image via flow-matching.
Integration from t=0 (noise) to t=1 (image).
"""
# Start from random latent
x_t = torch.randn(1, latent_dim, self.model_dim)
# ODE solver: integrate flow
for step in range(num_steps):
t = torch.tensor([step / num_steps])
# Predict flow at this timestep
flow = self.flow_predictor(torch.cat([x_t, prompt_embedding], dim=-1))
# Simple Euler integration step
dt = 1.0 / num_steps
x_t = x_t + flow * dt
# Decode latent to image tokens
logits = self.output_head(x_t)
image_tokens = torch.argmax(logits, dim=-1)
# Decode tokens to image
image = self._decode_tokens(image_tokens)
return image
def _decode_tokens(self, tokens):
"""Decode image tokens to pixel values."""
# Would use learned codebook (VQVAE, etc.)
return tokens.float() / self.vocab_size
class LatteFlow(nn.Module):
"""
Unified multimodal model combining vision-language understanding
with efficient flow-based generation via layerwise timestep experts.
"""
def __init__(self, pretrained_vlm, model_dim=768, num_experts=4):
super().__init__()
# Frozen pretrained VLM (understanding)
self.vlm = pretrained_vlm
for param in self.vlm.parameters():
param.requires_grad = False
self.model_dim = model_dim
# Trainable generation components
self.timestep_experts = LayerwiseTimestepExpert(
num_layers=28, # From Qwen2-VL-2B
num_experts=num_experts,
hidden_dim=model_dim,
num_heads=12
)
self.residual_attention = TimestepConditionedResidualAttention(
hidden_dim=model_dim,
num_heads=12,
num_layers=28
)
self.flow_generator = FlowMatchingGenerator(model_dim)
# Image encoder (compression)
self.image_encoder = ImageEncoder(model_dim)
def forward_understanding(self, image, text):
"""Vision-language understanding using frozen VLM."""
with torch.no_grad():
understanding = self.vlm.encode(image, text)
return understanding
def forward_generation(self, prompt, image_resolution=(256, 256)):
"""Efficient image generation with layerwise experts."""
# Encode prompt
prompt_embedding = self.vlm.text_encoder(prompt)
# Generate via flow-matching with timestep experts
generated_image = self.flow_generator.generate(
prompt_embedding,
num_steps=50
)
return generated_image
def forward(self, image=None, text=None, prompt=None, task='understanding'):
"""
Unified forward pass supporting both understanding and generation.
"""
if task == 'understanding':
return self.forward_understanding(image, text)
elif task == 'generation':
return self.forward_generation(prompt)
class ImageEncoder(nn.Module):
"""Efficient image encoder with 32x downsampling."""
def __init__(self, latent_dim):
super().__init__()
# Simplified: DeepCompression Autoencoder
self.encoder = nn.Sequential(
nn.Conv2d(3, 32, 4, stride=2, padding=1),
nn.ReLU(),
nn.Conv2d(32, 64, 4, stride=2, padding=1),
nn.ReLU(),
nn.Conv2d(64, latent_dim, 4, stride=2, padding=1)
)
def forward(self, image):
return self.encoder(image)
Architecture Selection:
Training Configuration:
Inference Optimization:
Performance Targets:
When to Use LaTtE-Flow: