Perancangan dan Analisis Algoritma Penentuan Rute Spasial Temporal Autoregresif Pada Pembangkitan Video Panjang yang Konsisten

Madany, Fadhl Akmal (2026) Perancangan dan Analisis Algoritma Penentuan Rute Spasial Temporal Autoregresif Pada Pembangkitan Video Panjang yang Konsisten. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 5025221028-Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (14MB) | Request a copy

Abstract

Model difusi video autoregresif (AR) menghadapi tantangan besar dalam mempertahankan konsistensi visual pada generasi berdurasi panjang. Penggunaan kebijakan eviksi First-In-First-Out (FIFO) pada KV cache menyebabkan model melupakan konteks historis yang krusial, memicu penyimpangan visual, dan menghasilkan halusinasi saat terjadi perturbasi adegan perantara. Untuk mengatasi masalah tersebut, penelitian ini mengusulkan Algoritma Spatio-Temporal Autoregressive Routing (STAR), sebuah algoritma inferensi bebas pelatihan (training-free) yang mengoptimalkan memori secara dinamis. STAR merutekan atensi murni pada konteks spasial-temporal yang paling relevan secara semantik. Selain itu, sebuah mekanisme Bank Memori diimplementasikan untuk menyimpan dan memanggil kembali jangkar konteks selama pengosongan memori (flush), sehingga konsistensi identitas subjek tetap terjaga melintasi transisi adegan multikompleks. Hasil evaluasi kuantitatif menggunakan VBench menunjukkan bahwa STAR berhasil mempertahankan skor Dynamic Degree sempurna sebesar 100% pada generasi video hingga 240 detik, mengungguli model baseline state-of-the-art yang rentan mengalami stagnasi gerak. Secara kualitatif, pengujian preferensi manusia memvalidasi keunggulan STAR dengan skor di atas 53% pada metrik konsistensi subjek, latar belakang, dan keselarasan instruksi.
================================================================================================================================
Autoregressive (AR) video diffusion models face significant challenges in maintaining visual consistency during long-duration generation. Relying on a strict First-In-First-Out (FIFO) eviction policy within the KV cache causes these models to forget crucial historical context, triggering visual drift and hallucinations during intermediate scene perturbations. To address this, this research proposes Spatio-Temporal Autoregressive Routing (STAR) Algorithm, a fully training-free inference framework that dynamically optimizes memory allocation. STAR routes attention strictly to the most semantically relevant spatio-temporal contexts. Furthermore, a Memory Bank mechanism is implemented to save and load context anchors during memory flush operations, effectively preserving subject identity consistency across complex multi-scene transitions. Quantitative evaluation using the VBench framework demonstrates that STAR successfully maintains a perfect 100% Dynamic Degree score for video generation of up to 240 seconds, significantly outperforming state-of-the-art baseline models that are prone to motion stagnation. Qualitatively, human preference studies validate STAR's visual superiority, achieving winning scores of over 53% across subject consistency, background consistency, and prompt alignment metrics.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	Video Diffusion Models, Autoregressive Generation, KV Cache Optimization, Dynamic Routing, Scene Transitions, Model Difusi Video, Generasi Autoregresif, Optimasi KV Cache, Penentuan Rute, Transisi Adegan.
Subjects:	Q Science > QA Mathematics > QA336 Artificial Intelligence Q Science > QA Mathematics > QA76.87 Neural networks (Computer Science) Q Science > QA Mathematics > QA76.9 Computer algorithms. Virtual Reality. Computer simulation.
Divisions:	Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Informatics Engineering > 55201-(S1) Undergraduate Thesis
Depositing User:	Fadhl Akmal Madany
Date Deposited:	24 Jul 2026 02:03
Last Modified:	24 Jul 2026 02:03
URI:	http://repository.its.ac.id/id/eprint/136895

Actions (login required)

View Item