Content-Consistent Start-to-End Frame Video Generation via Manifold-Guided Feature Intervention and Lightweight Adaptation

Abstract

Generating coherent video between semantically distant endpoint images requires both endpoint adherence and valid intermediate structure. We present TACT, a two-stage framework for endpoint-conditioned transition video generation. Stage 1 is training-free: it fuses start and end conditions and intervenes in early temporal features, using linear interpolation for same-object animation and spherical linear interpolation for cross-semantic metamorphosis while retaining native late denoising. Stage 2 distills the resulting trajectories into rank-64 LoRA adapters with an annealing-inspired perceptual-weight schedule. On held-out MorphBench, Stage 1 reduces LPIPS from 0.2897 to 0.1460, raises SSIM from 0.6830 to 0.8995, and improves PSNR by 8.40 dB over DynamiCrafter; it also outperforms TVG on these metrics. Stage 2 retains gains over the baseline while recovering sharper local texture, revealing a controlled smoothness-detail trade-off.