Skip to main content Startseite Ersteller adu2021 skillxiv latte-flow-unified-multimodal
latte-flow-unified-multimodal 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.
Zur Installation springen Skills Marktplatz Entdecken und erkunden Sie KI-Skills, die von der Community erstellt wurden.
Mit Codex oder Claude installieren Kopieren Sie diesen Prompt, fügen Sie ihn in Codex, Claude oder einen anderen Assistant ein und lassen Sie die Skill-Seite prüfen und installieren.
Prompt-Details anzeigen Ein direkter Befehl überspringt den Prüf-Prompt. Prüfen Sie die Quelle, bevor Sie ihn ausführen.
npx skills add https://github.com/ADu2021/skillXiv --skill latte-flow-unified-multimodalDer Befehl bleibt in einer Zeile. Scrollen Sie horizontal, um ihn vor dem Kopieren vollständig zu prüfen.
Sie bevorzugen eine lokale Kopie? Laden Sie die Dateien herunter, die SkillsMP derzeit vorliegen.
ZIP herunterladen Herunterladen... 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: Layerwise Timestep-Expert Flow-based Transformer
Core Concept
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.
Architecture Overview
Unified Multimodal Design : Integrates frozen pretrained vision-language model with trainable generation pathways for tight understanding-generation couplingLayerwise Timestep Experts : Partitions L transformer layers into K non-overlapping groups, each specializing in distinct timestep intervals during diffusionTimestep-Conditioned Residual Attention : Later layers reuse self-attention maps from earlier layers, modulated by timestep embeddings via gatingFlow-Matching Generation : Replaces traditional diffusion with more stable flow-matching formulation for image generationDual Architecture Variants : "Couple" preserves frozen VLM; "Blend" shares transformer layers for tighter integration
Implementation
Step 1: Layerwise Timestep Expert Architecture
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
self .expert_groups = nn.ModuleList()
for
in
range
for
in
range
self
self
self
1
128
128
def
forward
self, x, timestep
"""
Route to appropriate expert group based on timestep.
Only execute M=L/K layers instead of all L layers.
"""
self
float
1
1
1.0
self
min
int
self
1
self
for
in
return
class
TransformerBlock
"""Single transformer block."""
def
__init__
self, hidden_dim, num_heads
super
self
True
self
4
4
self
self
def
forward
self, x
self
self
self
self
return
Step 2: Timestep-Conditioned Residual Attention 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
self .attention_cache = {}
self .timestep_gate = nn.Sequential(
nn.Linear(1 , 64 ),
nn.SiLU(),
nn.Linear(64 , num_heads)
)
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.
"""
local_attn = self ._compute_attention(x)
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 :
t_embed = timestep.float ().unsqueeze(-1 )
gate = torch.sigmoid(self .timestep_gate(t_embed))
gate = gate.unsqueeze(1 )
blended_attn = gate * local_attn + (1 - gate) * cached_attn
attention_cache[layer_idx] = blended_attn
return blended_attn
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."""
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
Step 3: Flow-Matching Image Generation 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
self .flow_predictor = TransformerBlock(model_dim, num_heads=8 )
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.
"""
x = torch.cat([latent, prompt_embedding], dim=-1 )
flow = self .flow_predictor(x)
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).
"""
x_t = torch.randn(1 , latent_dim, self .model_dim)
for step in range (num_steps):
t = torch.tensor([step / num_steps])
flow = self .flow_predictor(torch.cat([x_t, prompt_embedding], dim=-1 ))
dt = 1.0 / num_steps
x_t = x_t + flow * dt
logits = self .output_head(x_t)
image_tokens = torch.argmax(logits, dim=-1 )
image = self ._decode_tokens(image_tokens)
return image
def _decode_tokens (self, tokens ):
"""Decode image tokens to pixel values."""
return tokens.float () / self .vocab_size
Step 4: Unified LaTtE-Flow Model 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__()
self .vlm = pretrained_vlm
for param in self .vlm.parameters():
param.requires_grad = False
self .model_dim = model_dim
self .timestep_experts = LayerwiseTimestepExpert(
num_layers=28 ,
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)
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."""
prompt_embedding = self .vlm.text_encoder(prompt)
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__()
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)
Practical Guidance
Couple variant : Preserve frozen VLM; simpler, less coupled understanding-generationBlend variant : Share transformer layers; tighter integration, more parameters
Dataset: 1.2M ImageNet images at 256×256 resolution
Batch size: 2,048 for stable training
Steps: 240K total (warm-up + main + fine-tune)
Learning rate: 1e-4 with cosine annealing
Layerwise experts: Execute only 7 layers per timestep (vs 28 for standard models)
Residual attention: Reuse 60% of attention maps from previous layers
Flow-matching: Replaces 50 diffusion steps with 50 ODE integration steps (comparable cost, better quality)
Speedup: 6x faster than comparable unified models (Unified-7B, etc.)
Understanding: Competitive with frozen Qwen2-VL-2B on multimodal benchmarks
Generation: FID 28-32 on ImageNet-50K (competitive with recent diffusion models)
Efficiency: <2 seconds inference on single GPU (512x512 resolution)
Joint understanding-generation applications
Edge deployment (efficiency critical)
Multimodal reasoning (image analysis + generation)
Real-time applications (low latency requirement)
Reference
Layerwise experts: Mixture-of-experts variant specialized by timestep rather than task
Flow-matching: Learned velocity field replaces reverse diffusion scheduling
Residual attention: Parameter reduction via attention map reuse across layers
Timestep conditioning: Gating mechanisms adapt computation to diffusion stage
Mehr aus diesem Repository
Verwandte Berufe SOC
Basierend auf der SOC-Berufsklassifikation
