IRIS-architecture / ARCHITECTURE.md
asdf98's picture
Add ARCHITECTURE.md
91e9e63 verified
# IRIS: Iterative Recurrent Image Synthesis
## A Novel Architecture for Mobile-First High-Quality Image Generation
### Version 1.0 | Architecture Design Document
---
## 1. Executive Summary
**IRIS** (Iterative Recurrent Image Synthesis) is a novel image generation architecture designed from first principles to achieve high visual quality on mobile devices (< 3-4GB RAM). It combines six key innovations drawn from cutting-edge research across multiple domains:
1. **Wavelet-Frequency Latent Space** β€” 16Γ— spatial compression via Haar DWT + learned VAE, operating in frequency-aware space
2. **Recurrent Depth Core** β€” Shared-weight denoising block iterated N times (inspired by Huginn), achieving deep model behavior from tiny parameter count
3. **Gated Recurrent Fourier Mixer (GRFM)** β€” Novel token mixing that combines RG-LRU gated recurrence with Adaptive Fourier Neural Operators, replacing O(NΒ²) attention with O(N log N) global mixing
4. **Manhattan Spatial Decay** β€” Learned per-head 2D spatial inductive bias via Manhattan distance exponential decay (from RMT)
5. **Rectified Flow with Consistency Distillation** β€” Straight ODE paths for few-step generation (1-4 steps)
6. **Adaptive Compute Budget** β€” Same model, variable quality: 4 iterations for mobile, 16 for quality
### Target Specifications
| Metric | Target | Achieved By |
|--------|--------|------------|
| Total Parameters | < 250M (generator) | Recurrent depth + efficient blocks |
| RAM (inference) | < 3GB total | ~600MB model + ~400MB VAE + ~200MB text encoder + buffers |
| Inference Steps | 1-4 | Rectified flow + consistency distillation |
| Core Iterations | 4 (fast) / 8-16 (quality) | Recurrent depth, shared weights |
| Image Quality | Competitive with SDXL at 512px | Frequency-aware latent + proper training |
| Prompt Adherence | Strong | CLIP-L/14 cross-attention conditioning |
| Training Cost | < 200 A100 GPU-hours (Stage 1) | Efficient architecture + progressive training |
---
## 2. Theoretical Foundations
### 2.1 Why Current Approaches Fail on Mobile
**Problem 1: Parameter Explosion in Transformers**
Standard DiT/UNet architectures use independent parameters for each layer. A 24-layer DiT-XL has ~675M params. Each self-attention layer stores O(dΒ²) params for Q,K,V projections Γ— number of layers.
**Problem 2: Quadratic Attention Complexity**
For 512Γ—512 images with 8Γ— VAE downsampling: 64Γ—64 = 4096 tokens. Self-attention requires 4096Β² Γ— d operations per layer. At d=768, that's ~12.9 GFLOPS per attention layer.
**Problem 3: Step Count**
Standard diffusion requires 20-50 neural function evaluations (NFE). Even a small model Γ— 50 steps = impractical.
### 2.2 Our Solution: Mathematical Framework
#### 2.2.1 Recurrent Depth as Implicit Neural ODE
The key insight from Huginn (arXiv:2502.05171): a shared-weight block `R` applied iteratively defines a discrete neural ODE:
```
s_0 = P(x) [Prelude: encode input]
s_{i+1} = R(s_i, c, t) [Core: iterate with conditioning c and timestep t]
y = C(s_r) [Coda: decode output]
```
This is mathematically equivalent to an Euler discretization of:
```
ds/dΟ„ = F_ΞΈ(s(Ο„), c, t) where Ο„ ∈ [0, 1], discretized into r steps
```
**Parameter efficiency**: If block R has P parameters, then r iterations give effective depth of rΓ—L layers (where L = layers in R) using only P parameters. A 6-layer block iterated 16 times = 96 effective layers.
**Connection to diffusion**: In standard diffusion, the denoiser f_ΞΈ is applied at each noise level t with the SAME parameters β€” this IS recurrent depth, but over the noise schedule axis. IRIS makes it recurrent over BOTH axes: noise schedule (outer loop, t) and computational depth (inner loop, Ο„).
#### 2.2.2 Gated Recurrent Fourier Mixer (GRFM) β€” Novel Contribution
We introduce GRFM, which processes the 2D token sequence through three parallel pathways merged multiplicatively:
**Pathway 1: Fourier Global Mixing (O(N log N))**
```
x_fourier = IRFFT2(SoftShrink(BlockMLP(RFFT2(x))))
```
From AFNO: captures global structure via frequency-domain mixing. The soft-shrinkage promotes sparsity in Fourier domain (images are naturally sparse in frequency).
**Pathway 2: Gated Linear Recurrence (O(N))**
```
a_t = Οƒ(Ξ›)^(cΒ·Οƒ(W_a Β· x_t)) [decay gate, per-element]
i_t = Οƒ(W_x Β· x_t) [input gate]
h_t = a_t βŠ™ h_{t-1} + √(1 - a_tΒ²) βŠ™ (i_t βŠ™ x_t) [RG-LRU update]
x_recurrent = W_o Β· h_T
```
From Griffin (arXiv:2402.19427): captures sequential dependencies with O(1) state per token position. Bidirectional (forward + backward scan).
**Pathway 3: Manhattan Spatial Gate**
```
D_{nm} = Ξ³_head^(|x_n - x_m| + |y_n - y_m|) [Manhattan decay matrix]
gate = Οƒ(W_g Β· x) βŠ™ (D Β· (W_v Β· x))
```
From RMT (arXiv:2309.11523): per-head learnable spatial decay provides multi-scale locality bias.
**Fusion (Novel)**:
```
output = LayerNorm(x_fourier βŠ™ Οƒ(gate) + x_recurrent βŠ™ (1 - Οƒ(gate)))
```
The gate adaptively selects between global Fourier features (textures, patterns) and local recurrent features (edges, fine details) based on spatial context. This is NOT a simple concatenation β€” it's a learned, spatially-varying interpolation.
#### 2.2.3 Wavelet-Frequency Latent Space
Instead of standard VAE operating on pixels, we first apply Haar DWT:
```
x ∈ R^{3Γ—HΓ—W} β†’ DWT β†’ y ∈ R^{12Γ—H/2Γ—W/2}
```
Then a lightweight VAE encoder compresses to:
```
z ∈ R^{CΓ—H/8Γ—W/8} (effective 16Γ— total spatial compression from original)
```
The VAE operates on wavelet coefficients, preserving frequency structure. The LL (low-low) subband carries global structure; LH, HL, HH carry directional high-frequency details. This means the latent space is inherently frequency-aware.
**Benefit**: The denoiser operates on a latent that already separates structure from detail, making the learning problem easier for a small model.
#### 2.2.4 Rectified Flow + Consistency Distillation
**Training Phase 1 (Rectified Flow)**:
```
x_t = (1-t) Β· x_0 + t Β· Ξ΅ [linear interpolation]
v_target = Ξ΅ - x_0 [velocity field]
L = w(t) Β· ||v_ΞΈ(x_t, t, c) - v_target||Β²
w(t) = t/(1-t+Ξ΅) [SNR reweighting]
t ~ LogitNormal(0, 1) [concentrate on hard timesteps]
```
**Training Phase 2 (Consistency Distillation)**:
```
f_ΞΈ(x_t, t) = c_skip(t)Β·x_t + c_out(t)Β·F_ΞΈ(x_t, t)
L_CD = d(f_ΞΈ(x_{t_{n+1}}, t_{n+1}), f_{θ⁻}(xΜ‚_{t_n}, t_n))
```
Where θ⁻ is EMA of θ. This enables 1-4 step generation.
---
## 3. Architecture Details
### 3.1 Overall Pipeline
```
Text β†’ CLIP-L/14 β†’ c ∈ R^{77Γ—768}
β”Œβ”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”
Image β†’ Haar DWT β†’ WaveletVAE Encode β†’ zβ‚€ ∈ R^{CΓ—hΓ—w} β”‚
β”‚ β”‚
β”‚ Noise schedule (RF): β”‚
β”‚ z_t = (1-t)zβ‚€ + tΒ·Ξ΅ β”‚
β”‚ β”‚
β–Ό β”‚
β”Œβ”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β” β”‚
β”‚ PRELUDE β”‚ β”‚
β”‚ (2 blocks) β”‚ β”‚
β”‚ PatchEmbed + β”‚ β”‚
β”‚ Initial mixing β”‚ β”‚
β””β”€β”€β”€β”€β”€β”€β”€β”€β”¬β”€β”€β”€β”€β”€β”€β”€β”€β”€β”˜ β”‚
β”‚ β”‚
β–Ό β”‚
β”Œβ”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β” β”‚
β”‚ CORE (shared) │◄── Iterate r β”‚
β”‚ GRFM Block β”‚ times β”‚
β”‚ + FFN β”‚ (4-16) β”‚
β”‚ + adaLN-Zero β”‚ β”‚
β””β”€β”€β”€β”€β”€β”€β”€β”€β”¬β”€β”€β”€β”€β”€β”€β”€β”€β”€β”˜ β”‚
β”‚ β”‚
β–Ό β”‚
β”Œβ”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β” β”‚
β”‚ CODA β”‚ β”‚
β”‚ (2 blocks) β”‚ β”‚
β”‚ Final refine + β”‚ β”‚
β”‚ Unpatchify β”‚ β”‚
β””β”€β”€β”€β”€β”€β”€β”€β”€β”¬β”€β”€β”€β”€β”€β”€β”€β”€β”€β”˜ β”‚
β”‚ β”‚
β–Ό β”‚
vΜ‚ = predicted velocity β”‚
zβ‚€_pred = z_t - tΒ·vΜ‚ β”‚
β”‚ β”‚
β–Ό β”‚
WaveletVAE Decode β†’ Haar IDWT β†’ Imageβ”‚
β””β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”˜
```
### 3.2 Detailed Block Design
#### Prelude (2 blocks, unique weights)
```python
class Prelude:
patch_embed: Conv2d(C_latent, D, kernel_size=2, stride=2) # 2Γ— spatial reduce
pos_embed: learned R^{(h/2 Γ— w/2) Γ— D}
blocks: [PreludeBlock Γ— 2]
class PreludeBlock:
norm1 β†’ DepthwiseSepConv3x3 β†’ GELU β†’ PointwiseConv β†’ norm2 β†’ FFN
# Uses conv instead of attention β€” cheap local feature extraction
```
#### Core (shared weights, iterated r times)
```python
class CoreBlock:
# adaLN-Zero conditioning on (timestep t, iteration i, text_global)
adaln_modulation: Linear(D_cond, 6*D) # scale, shift, gate for norm1, norm2
norm1 β†’ GRFM β†’ gate1 β†’ residual
norm2 β†’ CrossAttention(q=x, kv=text_tokens) β†’ gate2 β†’ residual # Only 77 text tokens
norm3 β†’ FFN(SiLU) β†’ gate3 β†’ residual
```
Cross-attention with only 77 text tokens is cheap: O(N Γ— 77 Γ— d) β‰ˆ O(NΒ·d).
#### GRFM (Gated Recurrent Fourier Mixer) β€” The Core Innovation
```python
class GRFM:
def forward(x, spatial_shape):
B, N, D = x.shape
H, W = spatial_shape
x_2d = x.reshape(B, H, W, D)
# Pathway 1: Fourier Global (O(N log N))
x_freq = rfft2(x_2d, dim=(1,2))
x_freq = block_mlp(x_freq) # Block-diagonal MLP in freq domain
x_freq = soft_shrink(x_freq, lambd=self.sparsity_threshold)
x_fourier = irfft2(x_freq, dim=(1,2))
# Pathway 2: Bidirectional Gated Recurrence (O(N))
x_flat_fwd = x # N tokens in raster order
x_flat_bwd = x.flip(1) # Reversed
h_fwd = gated_linear_recurrence(x_flat_fwd, self.decay_fwd, self.gate_fwd)
h_bwd = gated_linear_recurrence(x_flat_bwd, self.decay_bwd, self.gate_bwd)
x_recurrent = linear(concat(h_fwd, h_bwd.flip(1)))
# Pathway 3: Manhattan Spatial Gate
manhattan_dist = compute_manhattan(H, W) # Precomputed
gamma = sigmoid(self.gamma_param) # Per-head
spatial_decay = gamma.pow(manhattan_dist) # [heads, N, N] β€” sparse/windowed
x_gated = einsum('hnn,bnd->bnd', spatial_decay[:, :K, :K], value_proj(x))
gate = sigmoid(gate_proj(x))
# Adaptive Fusion
output = x_fourier * gate + x_recurrent * (1 - gate)
output = output + 0.1 * x_gated # Small residual from spatial
return output_proj(output)
```
#### Coda (2 blocks, unique weights)
```python
class Coda:
blocks: [CodaBlock Γ— 2]
unpatchify: ConvTranspose2d(D, C_latent, kernel_size=2, stride=2)
final_norm: LayerNorm(D)
class CodaBlock:
norm1 β†’ LocalWindowAttention(window=8) β†’ residual # Small window, efficient
norm2 β†’ FFN β†’ residual
```
### 3.3 Parameter Budget
| Component | Parameters | Notes |
|-----------|-----------|-------|
| WaveletVAE Encoder | ~15M | Lightweight (LiteVAE-style) |
| WaveletVAE Decoder | ~8M | Tiny decoder (SnapGen-style) |
| CLIP-L/14 Text Encoder | ~39M | Frozen, not counted for training |
| Prelude (2 blocks) | ~12M | Conv-based, cheap |
| Core Block (shared) | ~45M | GRFM + CrossAttn + FFN |
| Coda (2 blocks) | ~15M | Local attention + FFN |
| Embeddings/conditioning | ~3M | Time, iteration, position |
| **Total Generator** | **~75M unique** | Core shared across iterations |
| **Effective depth** | **75M β†’ behaves like 400M+** | At r=8 iterations |
| **Total system** | **~137M** | Including VAE + text encoder |
### 3.4 Memory Analysis (Inference at 512Γ—512)
```
CLIP-L/14 text encoder: ~156 MB (fp16)
WaveletVAE Decoder: ~16 MB (fp16)
IRIS Generator: ~150 MB (fp16)
Latent tensor: ~2 MB (32Γ—32Γ—16, fp16)
KV cache (text cross-attn): ~12 MB
Intermediate activations: ~100 MB (single block, not accumulated)
OS/framework overhead: ~500 MB
─────────────────────────────────────────
Total: ~936 MB βœ“ (well under 3GB)
```
**Key insight**: Because Core block weights are shared, we don't accumulate layer-by-layer activations. Each iteration reuses the same memory buffer.
---
## 4. Training Recipe
### Stage 1: Wavelet VAE Training (Standalone)
```
Data: ImageNet (1.2M images) + CC3M (3M images)
Resolution: 256Γ—256
Objective: Reconstruction loss + KL + Perceptual (LPIPS) + Wavelet frequency loss
Batch: 32
LR: 1e-4, cosine decay
Duration: ~20 GPU-hours on A100
```
### Stage 2: Class-Conditional Pretraining
```
Data: ImageNet 256Γ—256 (class labels)
Objective: Rectified Flow velocity matching
Batch: 256
LR: 1e-4, warmup 5000 steps, cosine decay
Core iterations: r=8 (randomly sample r ∈ {4,6,8,10,12} for robustness)
Duration: ~100 GPU-hours on A100
```
### Stage 3: Text-Image Alignment
```
Data: CC3M + CC12M (15M images with captions, re-captioned by VLM)
Resolution: 256β†’512 progressive
Objective: Rectified Flow + cross-attention on CLIP-L text tokens
Batch: 128
LR: 2e-5, constant
Duration: ~200 GPU-hours on A100
```
### Stage 4: Aesthetic Fine-tuning
```
Data: JourneyDB + high-aesthetic LAION subset (1M images, aesthetic score > 6.0)
Resolution: 512Γ—512
Batch: 64
LR: 5e-6
Duration: ~50 GPU-hours
```
### Stage 5: Consistency Distillation
```
Teacher: Trained IRIS model from Stage 4
Student: Same architecture, initialized from teacher
Objective: Consistency loss (CD) + optional LADD (adversarial)
Target: 1-4 step generation
Duration: ~30 GPU-hours
```
**Total estimated cost: ~400 A100 GPU-hours β‰ˆ $1,600 at cloud prices**
**Colab/Kaggle feasible**: Stage 1-2 can run on T4/A100 free tier
---
## 5. Novel Contributions Summary
1. **GRFM (Gated Recurrent Fourier Mixer)**: First architecture to fuse Fourier global mixing, gated linear recurrence, and Manhattan spatial decay in a single differentiable block with learned gating
2. **Recurrent Depth for Image Generation**: First application of the Huginn prelude-core-coda pattern to image generation, enabling budget-adaptive compute
3. **Wavelet-Frequency Latent Space**: DWT preprocessing before VAE encoding preserves frequency structure in the latent space
4. **Iteration-Aware Conditioning**: The core block receives both timestep t and iteration index i via adaLN, allowing it to learn different behavior at different depths
5. **Dual-Axis Recurrence**: Recurrence over both noise schedule (diffusion steps) and computational depth (core iterations) β€” a new paradigm for efficient generation
---
## 6. Extensions
### 6.1 Image Editing (Inpainting, Super-Resolution)
The iterative nature of IRIS makes it natural for editing:
- **Inpainting**: Mask latent tokens, condition core iterations on unmasked context
- **Super-Resolution**: Encode low-res image via WaveletVAE, condition generation on LL subband
- **Prompt-based editing**: Encode source image, modify text conditioning, run partial denoising (SDEdit-style)
### 6.2 ControlNet-like Conditioning
Add lightweight adapter to Prelude that injects spatial control signals (edges, depth, pose) into the latent, then the shared Core naturally propagates this through iterations.
---