Upload liquid_diffusion/model.py
Browse files- liquid_diffusion/model.py +419 -0
liquid_diffusion/model.py
ADDED
|
@@ -0,0 +1,419 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""
|
| 2 |
+
LiquidDiffusion Model — A Novel Attention-Free Image Generation Architecture
|
| 3 |
+
|
| 4 |
+
Core Innovation: Parallel Liquid Neural Network blocks for image generation.
|
| 5 |
+
The CfC (Closed-form Continuous-depth) time-gating mechanism naturally bridges
|
| 6 |
+
with diffusion timesteps — the diffusion noise level IS the liquid time constant.
|
| 7 |
+
|
| 8 |
+
Mathematical Foundation:
|
| 9 |
+
CfC Eq.10: x(t) = σ(-f·t) ⊙ g + (1 - σ(-f·t)) ⊙ h
|
| 10 |
+
|
| 11 |
+
For image generation, we adapt this as:
|
| 12 |
+
φ'(t) = σ(-f(φ)·t_diff) ⊙ g(φ) + (1 - σ(-f(φ)·t_diff)) ⊙ h(φ)
|
| 13 |
+
|
| 14 |
+
Where t_diff is the diffusion timestep, f/g/h are spatial feature transforms.
|
| 15 |
+
This is FULLY PARALLEL — no ODE solver, no sequential scanning.
|
| 16 |
+
|
| 17 |
+
Additionally, we use learnable exponential relaxation (from LiquidTAD):
|
| 18 |
+
α = exp(-λ·t_diff), out = α·φ + (1-α)·S(φ)
|
| 19 |
+
This gives depth-dependent, time-aware residual connections.
|
| 20 |
+
|
| 21 |
+
Architecture:
|
| 22 |
+
Input (noisy image) → Conv stem → [Encoder: DownBlocks with LiquidCfC]
|
| 23 |
+
→ Bottleneck (LiquidCfC) → [Decoder: UpBlocks with LiquidCfC + skip]
|
| 24 |
+
→ Conv head → Velocity prediction (for rectified flow)
|
| 25 |
+
|
| 26 |
+
No attention anywhere. All spatial mixing via depthwise convolutions +
|
| 27 |
+
multi-scale parallel processing in liquid blocks.
|
| 28 |
+
|
| 29 |
+
References:
|
| 30 |
+
[1] Hasani et al., "Closed-form Continuous-time Neural Networks", Nature MI 2022 (CfC)
|
| 31 |
+
[2] arxiv 2604.18274 — LiquidTAD (parallel liquid relaxation)
|
| 32 |
+
[3] arxiv 2504.13499 — USM (U-Shape Mamba for diffusion)
|
| 33 |
+
[4] Liu et al., "Flow Straight and Fast: Rectified Flow", ICLR 2023
|
| 34 |
+
"""
|
| 35 |
+
|
| 36 |
+
import math
|
| 37 |
+
import torch
|
| 38 |
+
import torch.nn as nn
|
| 39 |
+
import torch.nn.functional as F
|
| 40 |
+
|
| 41 |
+
|
| 42 |
+
# =============================================================================
|
| 43 |
+
# 1. TIME EMBEDDING — Sinusoidal + MLP
|
| 44 |
+
# =============================================================================
|
| 45 |
+
|
| 46 |
+
class SinusoidalTimeEmbedding(nn.Module):
|
| 47 |
+
"""Maps scalar timestep t to a high-dimensional embedding.
|
| 48 |
+
Uses sinusoidal positional encoding followed by 2-layer MLP.
|
| 49 |
+
"""
|
| 50 |
+
def __init__(self, dim: int, max_period: int = 10000):
|
| 51 |
+
super().__init__()
|
| 52 |
+
self.dim = dim
|
| 53 |
+
self.max_period = max_period
|
| 54 |
+
self.mlp = nn.Sequential(
|
| 55 |
+
nn.Linear(dim, dim * 4),
|
| 56 |
+
nn.SiLU(),
|
| 57 |
+
nn.Linear(dim * 4, dim),
|
| 58 |
+
)
|
| 59 |
+
|
| 60 |
+
def forward(self, t: torch.Tensor) -> torch.Tensor:
|
| 61 |
+
"""t: [B] timestep values in [0, 1] → [B, dim] embeddings"""
|
| 62 |
+
half = self.dim // 2
|
| 63 |
+
freqs = torch.exp(
|
| 64 |
+
-math.log(self.max_period) * torch.arange(half, device=t.device, dtype=t.dtype) / half
|
| 65 |
+
)
|
| 66 |
+
args = t[:, None] * freqs[None, :]
|
| 67 |
+
emb = torch.cat([torch.cos(args), torch.sin(args)], dim=-1)
|
| 68 |
+
if self.dim % 2 == 1:
|
| 69 |
+
emb = F.pad(emb, (0, 1))
|
| 70 |
+
return self.mlp(emb)
|
| 71 |
+
|
| 72 |
+
|
| 73 |
+
# =============================================================================
|
| 74 |
+
# 2. ADAPTIVE LAYER NORM (AdaLN) — Timestep conditioning via scale/shift
|
| 75 |
+
# =============================================================================
|
| 76 |
+
|
| 77 |
+
class AdaLN(nn.Module):
|
| 78 |
+
"""Adaptive Layer Normalization: out = norm(x) * (1 + scale(t)) + shift(t)"""
|
| 79 |
+
def __init__(self, dim: int, cond_dim: int):
|
| 80 |
+
super().__init__()
|
| 81 |
+
self.norm = nn.GroupNorm(num_groups=min(32, dim), num_channels=dim, affine=False)
|
| 82 |
+
self.proj = nn.Sequential(nn.SiLU(), nn.Linear(cond_dim, dim * 2))
|
| 83 |
+
|
| 84 |
+
def forward(self, x: torch.Tensor, t_emb: torch.Tensor) -> torch.Tensor:
|
| 85 |
+
"""x: [B,C,H,W], t_emb: [B, cond_dim] → [B,C,H,W]"""
|
| 86 |
+
scale, shift = self.proj(t_emb).chunk(2, dim=1)
|
| 87 |
+
return self.norm(x) * (1 + scale[:, :, None, None]) + shift[:, :, None, None]
|
| 88 |
+
|
| 89 |
+
|
| 90 |
+
# =============================================================================
|
| 91 |
+
# 3. PARALLEL CfC BLOCK — Core liquid neural network layer
|
| 92 |
+
# =============================================================================
|
| 93 |
+
|
| 94 |
+
class ParallelCfCBlock(nn.Module):
|
| 95 |
+
"""Parallel Closed-form Continuous-depth block for spatial features.
|
| 96 |
+
|
| 97 |
+
CfC Eq.10: x(t) = σ(-f·t) ⊙ g + (1 - σ(-f·t)) ⊙ h
|
| 98 |
+
|
| 99 |
+
Adaptations for image generation:
|
| 100 |
+
1. f/g/h heads operate on 2D feature maps via conv layers
|
| 101 |
+
2. Diffusion timestep t IS the CfC time parameter
|
| 102 |
+
3. Multi-directional depthwise convolutions for spatial context
|
| 103 |
+
4. No recurrence — each spatial position computed independently
|
| 104 |
+
5. Liquid relaxation residual: α·input + (1-α)·CfC_output
|
| 105 |
+
where α = exp(-λ·t_diff) adapts residual strength to noise level
|
| 106 |
+
"""
|
| 107 |
+
def __init__(self, dim: int, t_dim: int, expand_ratio: float = 2.0,
|
| 108 |
+
kernel_size: int = 7, dropout: float = 0.0):
|
| 109 |
+
super().__init__()
|
| 110 |
+
hidden = int(dim * expand_ratio)
|
| 111 |
+
|
| 112 |
+
# Shared backbone: depthwise + pointwise for local spatial context
|
| 113 |
+
self.backbone_dw = nn.Conv2d(dim, dim, kernel_size, padding=kernel_size // 2, groups=dim)
|
| 114 |
+
self.backbone_pw = nn.Conv2d(dim, hidden, 1)
|
| 115 |
+
self.backbone_act = nn.SiLU()
|
| 116 |
+
|
| 117 |
+
# Three CfC heads
|
| 118 |
+
self.f_head = nn.Conv2d(hidden, dim, 1) # time-constant gate
|
| 119 |
+
self.g_head = nn.Sequential( # "from" state
|
| 120 |
+
nn.Conv2d(hidden, hidden, kernel_size, padding=kernel_size // 2, groups=hidden),
|
| 121 |
+
nn.SiLU(),
|
| 122 |
+
nn.Conv2d(hidden, dim, 1),
|
| 123 |
+
)
|
| 124 |
+
self.h_head = nn.Sequential( # "to" state (attractor)
|
| 125 |
+
nn.Conv2d(hidden, hidden, kernel_size, padding=kernel_size // 2, groups=hidden),
|
| 126 |
+
nn.SiLU(),
|
| 127 |
+
nn.Conv2d(hidden, dim, 1),
|
| 128 |
+
)
|
| 129 |
+
|
| 130 |
+
# CfC time parameters: maps t_emb to per-channel gate modulation
|
| 131 |
+
self.time_a = nn.Linear(t_dim, dim)
|
| 132 |
+
self.time_b = nn.Linear(t_dim, dim)
|
| 133 |
+
|
| 134 |
+
# Liquid relaxation decay (LiquidTAD-inspired)
|
| 135 |
+
self.rho = nn.Parameter(torch.zeros(1, dim, 1, 1))
|
| 136 |
+
|
| 137 |
+
# Output gate conditioned on timestep
|
| 138 |
+
self.output_gate = nn.Sequential(nn.SiLU(), nn.Linear(t_dim, dim))
|
| 139 |
+
|
| 140 |
+
self.dropout = nn.Dropout(dropout) if dropout > 0 else nn.Identity()
|
| 141 |
+
|
| 142 |
+
def forward(self, x: torch.Tensor, t_emb: torch.Tensor) -> torch.Tensor:
|
| 143 |
+
"""x: [B,C,H,W], t_emb: [B, t_dim] → [B,C,H,W]"""
|
| 144 |
+
residual = x
|
| 145 |
+
|
| 146 |
+
# Shared backbone
|
| 147 |
+
backbone = self.backbone_act(self.backbone_pw(self.backbone_dw(x)))
|
| 148 |
+
|
| 149 |
+
# Three CfC heads
|
| 150 |
+
f = self.f_head(backbone) # time constant logits
|
| 151 |
+
g = self.g_head(backbone) # "from" state
|
| 152 |
+
h = self.h_head(backbone) # "to" state
|
| 153 |
+
|
| 154 |
+
# CfC time-gating: σ(time_a(t) · f - time_b(t))
|
| 155 |
+
ta = self.time_a(t_emb)[:, :, None, None]
|
| 156 |
+
tb = self.time_b(t_emb)[:, :, None, None]
|
| 157 |
+
gate = torch.sigmoid(ta * f - tb)
|
| 158 |
+
|
| 159 |
+
# CfC interpolation: gate*g + (1-gate)*h
|
| 160 |
+
cfc_out = gate * g + (1.0 - gate) * h
|
| 161 |
+
cfc_out = self.dropout(cfc_out)
|
| 162 |
+
|
| 163 |
+
# Liquid relaxation: α = exp(-λ · |t_mean|)
|
| 164 |
+
t_scalar = t_emb.mean(dim=1, keepdim=True)[:, :, None, None]
|
| 165 |
+
lam = F.softplus(self.rho) + 1e-6
|
| 166 |
+
alpha = torch.exp(-lam * t_scalar.abs().clamp(min=0.01))
|
| 167 |
+
|
| 168 |
+
out = alpha * residual + (1.0 - alpha) * cfc_out
|
| 169 |
+
|
| 170 |
+
# Output gate
|
| 171 |
+
out_gate = torch.sigmoid(self.output_gate(t_emb))[:, :, None, None]
|
| 172 |
+
return out * out_gate
|
| 173 |
+
|
| 174 |
+
|
| 175 |
+
# =============================================================================
|
| 176 |
+
# 4. MULTI-SCALE SPATIAL MIXING — Global context without attention
|
| 177 |
+
# =============================================================================
|
| 178 |
+
|
| 179 |
+
class MultiScaleSpatialMix(nn.Module):
|
| 180 |
+
"""Multi-scale depthwise conv + global pooling for spatial context.
|
| 181 |
+
|
| 182 |
+
Uses parallel depthwise convolutions at 3x3, 5x5, 7x7 scales
|
| 183 |
+
plus adaptive average pooling for global receptive field.
|
| 184 |
+
This replaces self-attention's global spatial mixing.
|
| 185 |
+
"""
|
| 186 |
+
def __init__(self, dim: int, t_dim: int):
|
| 187 |
+
super().__init__()
|
| 188 |
+
self.dw3 = nn.Conv2d(dim, dim, 3, padding=1, groups=dim)
|
| 189 |
+
self.dw5 = nn.Conv2d(dim, dim, 5, padding=2, groups=dim)
|
| 190 |
+
self.dw7 = nn.Conv2d(dim, dim, 7, padding=3, groups=dim)
|
| 191 |
+
self.global_pool = nn.AdaptiveAvgPool2d(1)
|
| 192 |
+
self.global_proj = nn.Conv2d(dim, dim, 1)
|
| 193 |
+
self.merge = nn.Conv2d(dim * 4, dim, 1)
|
| 194 |
+
self.act = nn.SiLU()
|
| 195 |
+
self.adaln = AdaLN(dim, t_dim)
|
| 196 |
+
|
| 197 |
+
def forward(self, x: torch.Tensor, t_emb: torch.Tensor) -> torch.Tensor:
|
| 198 |
+
x_norm = self.adaln(x, t_emb)
|
| 199 |
+
s3 = self.dw3(x_norm)
|
| 200 |
+
s5 = self.dw5(x_norm)
|
| 201 |
+
s7 = self.dw7(x_norm)
|
| 202 |
+
sg = self.global_proj(self.global_pool(x_norm)).expand_as(x_norm)
|
| 203 |
+
return x + self.act(self.merge(torch.cat([s3, s5, s7, sg], dim=1)))
|
| 204 |
+
|
| 205 |
+
|
| 206 |
+
# =============================================================================
|
| 207 |
+
# 5. LIQUID DIFFUSION BLOCK — Complete processing unit
|
| 208 |
+
# =============================================================================
|
| 209 |
+
|
| 210 |
+
class LiquidDiffusionBlock(nn.Module):
|
| 211 |
+
"""One complete LiquidDiffusion block:
|
| 212 |
+
AdaLN → ParallelCfC → MultiScaleSpatialMix → FeedForward
|
| 213 |
+
"""
|
| 214 |
+
def __init__(self, dim: int, t_dim: int, expand_ratio: float = 2.0,
|
| 215 |
+
kernel_size: int = 7, dropout: float = 0.0):
|
| 216 |
+
super().__init__()
|
| 217 |
+
self.adaln1 = AdaLN(dim, t_dim)
|
| 218 |
+
self.cfc = ParallelCfCBlock(dim, t_dim, expand_ratio, kernel_size, dropout)
|
| 219 |
+
self.spatial_mix = MultiScaleSpatialMix(dim, t_dim)
|
| 220 |
+
self.adaln2 = AdaLN(dim, t_dim)
|
| 221 |
+
ff_dim = int(dim * expand_ratio)
|
| 222 |
+
self.ff = nn.Sequential(
|
| 223 |
+
nn.Conv2d(dim, ff_dim, 1), nn.SiLU(), nn.Conv2d(ff_dim, dim, 1),
|
| 224 |
+
)
|
| 225 |
+
self.res_scale = nn.Parameter(torch.ones(1) * 0.1)
|
| 226 |
+
|
| 227 |
+
def forward(self, x: torch.Tensor, t_emb: torch.Tensor) -> torch.Tensor:
|
| 228 |
+
x = x + self.res_scale * self.cfc(self.adaln1(x, t_emb), t_emb)
|
| 229 |
+
x = self.spatial_mix(x, t_emb)
|
| 230 |
+
x = x + self.res_scale * self.ff(self.adaln2(x, t_emb))
|
| 231 |
+
return x
|
| 232 |
+
|
| 233 |
+
|
| 234 |
+
# =============================================================================
|
| 235 |
+
# 6. DOWN/UP SAMPLE + SKIP FUSION
|
| 236 |
+
# =============================================================================
|
| 237 |
+
|
| 238 |
+
class DownSample(nn.Module):
|
| 239 |
+
"""Strided convolution downsampling (2x)."""
|
| 240 |
+
def __init__(self, in_dim: int, out_dim: int):
|
| 241 |
+
super().__init__()
|
| 242 |
+
self.conv = nn.Conv2d(in_dim, out_dim, 3, stride=2, padding=1)
|
| 243 |
+
def forward(self, x):
|
| 244 |
+
return self.conv(x)
|
| 245 |
+
|
| 246 |
+
|
| 247 |
+
class UpSample(nn.Module):
|
| 248 |
+
"""Nearest-neighbor interpolation + conv upsampling (2x)."""
|
| 249 |
+
def __init__(self, in_dim: int, out_dim: int):
|
| 250 |
+
super().__init__()
|
| 251 |
+
self.conv = nn.Conv2d(in_dim, out_dim, 3, padding=1)
|
| 252 |
+
def forward(self, x):
|
| 253 |
+
return self.conv(F.interpolate(x, scale_factor=2, mode='nearest'))
|
| 254 |
+
|
| 255 |
+
|
| 256 |
+
class SkipFusion(nn.Module):
|
| 257 |
+
"""Timestep-gated skip connection fusion."""
|
| 258 |
+
def __init__(self, dim: int, t_dim: int):
|
| 259 |
+
super().__init__()
|
| 260 |
+
self.proj = nn.Conv2d(dim * 2, dim, 1)
|
| 261 |
+
self.gate = nn.Sequential(nn.SiLU(), nn.Linear(t_dim, dim), nn.Sigmoid())
|
| 262 |
+
|
| 263 |
+
def forward(self, x, skip, t_emb):
|
| 264 |
+
merged = self.proj(torch.cat([x, skip], dim=1))
|
| 265 |
+
g = self.gate(t_emb)[:, :, None, None]
|
| 266 |
+
return merged * g + x * (1 - g)
|
| 267 |
+
|
| 268 |
+
|
| 269 |
+
# =============================================================================
|
| 270 |
+
# 7. LIQUID DIFFUSION U-NET — The complete denoiser
|
| 271 |
+
# =============================================================================
|
| 272 |
+
|
| 273 |
+
class LiquidDiffusionUNet(nn.Module):
|
| 274 |
+
"""LiquidDiffusion: Attention-Free Image Generation with Liquid Neural Networks.
|
| 275 |
+
|
| 276 |
+
U-Net where every processing block uses Parallel CfC layers instead of attention.
|
| 277 |
+
The diffusion timestep serves dual purpose:
|
| 278 |
+
1. Conditions the denoiser via AdaLN scale/shift
|
| 279 |
+
2. Acts as CfC "time parameter" — controlling liquid neuron interpolation
|
| 280 |
+
|
| 281 |
+
Scales:
|
| 282 |
+
tiny: channels=[64,128,256], blocks=[2,2,4], ~8M (256px, fast)
|
| 283 |
+
small: channels=[96,192,384], blocks=[2,3,6], ~25M (256px, quality)
|
| 284 |
+
base: channels=[128,256,512], blocks=[2,4,8], ~65M (512px)
|
| 285 |
+
large: channels=[128,256,512,768],blocks=[2,4,8,4], ~120M (512px HQ)
|
| 286 |
+
"""
|
| 287 |
+
def __init__(self, in_channels=3, channels=None, blocks_per_stage=None,
|
| 288 |
+
t_dim=256, expand_ratio=2.0, kernel_size=7, dropout=0.0):
|
| 289 |
+
super().__init__()
|
| 290 |
+
if channels is None:
|
| 291 |
+
channels = [64, 128, 256]
|
| 292 |
+
if blocks_per_stage is None:
|
| 293 |
+
blocks_per_stage = [2, 2, 4]
|
| 294 |
+
|
| 295 |
+
assert len(channels) == len(blocks_per_stage)
|
| 296 |
+
self.channels = channels
|
| 297 |
+
self.num_stages = len(channels)
|
| 298 |
+
|
| 299 |
+
# Time embedding
|
| 300 |
+
self.time_embed = SinusoidalTimeEmbedding(t_dim)
|
| 301 |
+
|
| 302 |
+
# Input stem
|
| 303 |
+
self.stem = nn.Sequential(
|
| 304 |
+
nn.Conv2d(in_channels, channels[0], 3, padding=1),
|
| 305 |
+
nn.SiLU(),
|
| 306 |
+
nn.Conv2d(channels[0], channels[0], 3, padding=1),
|
| 307 |
+
)
|
| 308 |
+
|
| 309 |
+
# Encoder
|
| 310 |
+
self.encoder_blocks = nn.ModuleList()
|
| 311 |
+
self.downsamplers = nn.ModuleList()
|
| 312 |
+
for i in range(self.num_stages):
|
| 313 |
+
stage = nn.ModuleList()
|
| 314 |
+
for _ in range(blocks_per_stage[i]):
|
| 315 |
+
stage.append(LiquidDiffusionBlock(
|
| 316 |
+
channels[i], t_dim, expand_ratio, kernel_size, dropout))
|
| 317 |
+
self.encoder_blocks.append(stage)
|
| 318 |
+
if i < self.num_stages - 1:
|
| 319 |
+
self.downsamplers.append(DownSample(channels[i], channels[i + 1]))
|
| 320 |
+
|
| 321 |
+
# Bottleneck
|
| 322 |
+
self.bottleneck = nn.ModuleList([
|
| 323 |
+
LiquidDiffusionBlock(channels[-1], t_dim, expand_ratio, kernel_size, dropout),
|
| 324 |
+
LiquidDiffusionBlock(channels[-1], t_dim, expand_ratio, kernel_size, dropout),
|
| 325 |
+
])
|
| 326 |
+
|
| 327 |
+
# Decoder
|
| 328 |
+
self.decoder_blocks = nn.ModuleList()
|
| 329 |
+
self.upsamplers = nn.ModuleList()
|
| 330 |
+
self.skip_fusions = nn.ModuleList()
|
| 331 |
+
for i in range(self.num_stages - 1, -1, -1):
|
| 332 |
+
if i < self.num_stages - 1:
|
| 333 |
+
self.upsamplers.append(UpSample(channels[i + 1], channels[i]))
|
| 334 |
+
self.skip_fusions.append(SkipFusion(channels[i], t_dim))
|
| 335 |
+
stage = nn.ModuleList()
|
| 336 |
+
for _ in range(blocks_per_stage[i]):
|
| 337 |
+
stage.append(LiquidDiffusionBlock(
|
| 338 |
+
channels[i], t_dim, expand_ratio, kernel_size, dropout))
|
| 339 |
+
self.decoder_blocks.append(stage)
|
| 340 |
+
|
| 341 |
+
# Output head (initialized to zero for stable start)
|
| 342 |
+
self.head = nn.Sequential(
|
| 343 |
+
nn.GroupNorm(min(32, channels[0]), channels[0]),
|
| 344 |
+
nn.SiLU(),
|
| 345 |
+
nn.Conv2d(channels[0], in_channels, 3, padding=1),
|
| 346 |
+
)
|
| 347 |
+
nn.init.zeros_(self.head[-1].weight)
|
| 348 |
+
nn.init.zeros_(self.head[-1].bias)
|
| 349 |
+
|
| 350 |
+
def forward(self, x: torch.Tensor, t: torch.Tensor) -> torch.Tensor:
|
| 351 |
+
"""
|
| 352 |
+
Args:
|
| 353 |
+
x: [B, C, H, W] noisy image
|
| 354 |
+
t: [B] timestep values in [0, 1]
|
| 355 |
+
Returns:
|
| 356 |
+
[B, C, H, W] predicted velocity
|
| 357 |
+
"""
|
| 358 |
+
t_emb = self.time_embed(t)
|
| 359 |
+
h = self.stem(x)
|
| 360 |
+
|
| 361 |
+
# Encoder
|
| 362 |
+
skips = []
|
| 363 |
+
for i in range(self.num_stages):
|
| 364 |
+
for block in self.encoder_blocks[i]:
|
| 365 |
+
h = block(h, t_emb)
|
| 366 |
+
skips.append(h)
|
| 367 |
+
if i < self.num_stages - 1:
|
| 368 |
+
h = self.downsamplers[i](h)
|
| 369 |
+
|
| 370 |
+
# Bottleneck
|
| 371 |
+
for block in self.bottleneck:
|
| 372 |
+
h = block(h, t_emb)
|
| 373 |
+
|
| 374 |
+
# Decoder
|
| 375 |
+
up_idx = 0
|
| 376 |
+
for dec_i in range(self.num_stages):
|
| 377 |
+
stage_idx = self.num_stages - 1 - dec_i
|
| 378 |
+
if dec_i > 0:
|
| 379 |
+
h = self.upsamplers[up_idx](h)
|
| 380 |
+
h = self.skip_fusions[up_idx](h, skips[stage_idx], t_emb)
|
| 381 |
+
up_idx += 1
|
| 382 |
+
for block in self.decoder_blocks[dec_i]:
|
| 383 |
+
h = block(h, t_emb)
|
| 384 |
+
|
| 385 |
+
return self.head(h)
|
| 386 |
+
|
| 387 |
+
def count_params(self):
|
| 388 |
+
total = sum(p.numel() for p in self.parameters())
|
| 389 |
+
trainable = sum(p.numel() for p in self.parameters() if p.requires_grad)
|
| 390 |
+
return total, trainable
|
| 391 |
+
|
| 392 |
+
|
| 393 |
+
# =============================================================================
|
| 394 |
+
# 8. MODEL CONFIGS
|
| 395 |
+
# =============================================================================
|
| 396 |
+
|
| 397 |
+
def liquid_diffusion_tiny(**kwargs):
|
| 398 |
+
"""~8M params, 256px, fits ~4GB VRAM."""
|
| 399 |
+
return LiquidDiffusionUNet(
|
| 400 |
+
channels=[64, 128, 256], blocks_per_stage=[2, 2, 4],
|
| 401 |
+
t_dim=256, expand_ratio=2.0, kernel_size=7, **kwargs)
|
| 402 |
+
|
| 403 |
+
def liquid_diffusion_small(**kwargs):
|
| 404 |
+
"""~25M params, 256px, fits ~8GB VRAM."""
|
| 405 |
+
return LiquidDiffusionUNet(
|
| 406 |
+
channels=[96, 192, 384], blocks_per_stage=[2, 3, 6],
|
| 407 |
+
t_dim=384, expand_ratio=2.0, kernel_size=7, **kwargs)
|
| 408 |
+
|
| 409 |
+
def liquid_diffusion_base(**kwargs):
|
| 410 |
+
"""~65M params, 512px, fits ~14GB VRAM."""
|
| 411 |
+
return LiquidDiffusionUNet(
|
| 412 |
+
channels=[128, 256, 512], blocks_per_stage=[2, 4, 8],
|
| 413 |
+
t_dim=512, expand_ratio=2.0, kernel_size=7, **kwargs)
|
| 414 |
+
|
| 415 |
+
def liquid_diffusion_large(**kwargs):
|
| 416 |
+
"""~120M params, 512px, needs ~24GB VRAM."""
|
| 417 |
+
return LiquidDiffusionUNet(
|
| 418 |
+
channels=[128, 256, 512, 768], blocks_per_stage=[2, 4, 8, 4],
|
| 419 |
+
t_dim=512, expand_ratio=2.0, kernel_size=7, **kwargs)
|