Work

Adaptive Spatial Generative Architecture (ASGA v0.7): Formally Verified Multiscale Residual Flow with Bidirectional Revision and Block-Factored Attention

Year: 2026
Type: preprint
Abstract: Prevailing visual generative models suffer from one of two structural pathologies: uniform-grid diffusion and rectified-flow transformers (DiTs) expend identical floating-point operations on featureless background regions and intricate foreground structures across dozens of full-grid trajectory steps, whereas unidirectional multiscale cascades (such as Visual Autoregressive Modeling, VAR) enforce a rigid coarse-to-fine schedule that cannot revise early coarse structural errors once committed. We introduce the Adaptive Spatial Generative Architecture (ASGA v0.7), a dynamic, non-monotonic, multi-region generative paradigm that reformulates image synthesis as an adaptive sequential decision process over a Shared Multiscale Residual Latent Canvas equipped with Local Flow Maturity Maps and Versioned Support Counters. First, we subject the mathematical foundations of ASGA to a machine-checked formal verification audit in Lean 4 (v4.36.0), uncovering and repairing 10 subtle mathematical failure modes and proving 34 machine-checked theorems (0 errors, 0 warnings, 0 sorry axioms) covering state-space invariants, parallel batch commutativity, cross-scale descendant invalidation, trust-region adversarial drift bounds, and deterministic finite-time Lyapunov termination. Second, we design Block-Factored Unmasked FlashAttention (delivering a 21.0x wall-clock CUDA kernel speedup with < 2.1e-7 numerical divergence) alongside a Sparse O(L_spatial) Full-Rank 1D-Row x 1D-Col Kronecker Spatial Conditioning Field. Third, across five empirical evaluations in the pretrained Stable Diffusion TAESD latent space—including 512x512 real Kodak natural photograph generation from pure Gaussian noise, zero-shot generalization across all 24 Kodak photographs (16 held-out unseen test images), and a 4-attack white-box red-team adversarial suite—ASGA v0.7 outperforms a matched-parameter Dense 64x64 Rectified Flow DiT by +12.36 to +13.84 dB PSNR while consuming 6.56x to 13.13x fewer active token-steps, generalizes zero-shot to unseen photographs within 0.18 dB of the ground-truth Oracle, and demonstrates +8.44 dB autonomous STRUCT_REVISE self-healing on real spliced photographs.
Language: English
Open Access status: green
OpenAlex ID: W7236861259
Cites:
Cited by:
Related to: 10