🎨 PMA-VAE: Parallel Mobile Artistic Variational Autoencoder

A novel attention-free architecture for image generation, super-resolution, artifact removal, and artistic style transfer.

🏗️ Architecture

Image → PixelUnshuffle stem → MobileConv stages → Parallel 2D Mamba blocks
  → Multi-scale latent (z_base H/16, z_detail H/8, z_style global)
  → Light parallel decoder with FiLM style modulation → Reconstructed image

Key Design Principles

Component	Choice	Why
Backbone	MobileConv + Parallel 2D Mamba	Fast, efficient, attention-free
Downsampling	PixelUnshuffle → stride-2 conv	Lossless initial features
Upsampling	PixelShuffle (sub-pixel)	Mobile-friendly, no checkerboard
Latent	Multi-scale (base/detail/style)	Controllable, prevents collapse
Style control	FiLM conditioning	Lightweight, multiplicative
Global context	4-dir cross-scan SSM	O(n) complexity, no attention
Local context	Depthwise separable conv + SE	Standard mobile building block
Training	Progressive resolution + KL warmup	Stable convergence
Loss	L1 + VGG + PatchGAN + edge + KL	Comprehensive quality

✨ Features

• Attention-free: Uses parallel 2D Mamba/SSM blocks instead of self-attention
• Mobile-deployable: Lightweight decoder (~4-8M params) using depthwise separable convolutions
• Multi-scale latent space: z_base (structure), z_detail (texture), z_style (global style)
• No sequential pixel loops: Fully parallel training AND inference via Blelloch parallel scan
• Anti-collapse: KL warmup + free bits + progressive resolution training
• FiLM conditioning: Style modulation throughout the decoder for artist style transfer
• Pure PyTorch: No custom CUDA kernels needed — works on Colab free tier T4

📊 Model Variants

Config	Encoder	Decoder	Total	Target
pmavae_tiny	0.56M	1.91M	2.47M	Testing
pmavae_small	2.00M	4.27M	6.27M	Free Colab T4
pmavae_base	5.18M	9.83M	15.01M	Colab Pro / better GPU

🔧 Multi-Scale Latent Space

z_base   : H/16 × W/16 × 24-32  → Structure, composition, objects
z_detail : H/8  × W/8  × 6-8    → Texture, brush strokes, edges
z_style  : 1 × 1 × 96-128       → Global style vector

This separation enables:

• Style transfer: Swap z_style between images
• Super-resolution: Enhance z_detail while keeping z_base
• Artifact removal: Clean up z_detail while preserving structure
• Image generation: Sample from learned distributions

🏋️ Training

Loss Function

Loss = L1 + 0.5 × VGG_perceptual + 0.1 × edge_sobel + β × KL_free_bits + λ × PatchGAN

• KL warmup: β linearly increases from 0 → 1e-6 over 5000 steps
• Discriminator cold start: PatchGAN activates after 10000 steps
• Adaptive disc weight: Gradient magnitude balancing (taming-transformers trick)
• Free bits: 0.25 nats per latent dimension prevents posterior collapse

Progressive Resolution

Phase 1: 256×256 → Learn structure
Phase 2: 384×384 → Refine texture
Phase 3: 512×512 → Full detail
Phase 4: FHD tiled → High-resolution fine-tuning

📱 Mobile Deployment

The decoder is designed for mobile:

• Depthwise separable convolutions (MobileNet-style)
• PixelShuffle upsampling (no transpose conv artifacts)
• Squeeze-Excitation for channel attention
• FiLM style modulation (single MLP)
• Exportable to ONNX → Core ML / TFLite / ONNX Runtime Mobile

📁 Files

• model.py — Full PMA-VAE architecture (encoder, decoder, SSM blocks)
• losses.py — Loss functions (VGG perceptual, PatchGAN, KL free bits, edge loss)
• train.py — Training script with progressive resolution, checkpoint management
• PMA_VAE_Colab_Training.ipynb — Complete Colab notebook

📚 References

• Mamba: Gu & Dao, 2023. arxiv:2312.00752
• VMamba: Liu et al., 2024. arxiv:2401.10166
• Latent Diffusion: Rombach et al., 2022. arxiv:2112.10752
• VAE: Kingma & Welling, 2013. arxiv:1312.6114
• Taming Transformers: Esser et al., 2021. arxiv:2012.09841
• EfficientViT: Cai et al., 2023. arxiv:2305.07027
• MobileOne: Vasu et al., 2022. arxiv:2206.04040

License

MIT