Figure 4: Illustration of key reference frame retrieval framework. Given query frames from current query chunk to be out-painted and candidate frames from both history chunk and future chunk, we aim to find the most suitable reference frames to out-paint current chunk. Specifically, we calculate the registration maps and scores for all query-candidate pairs, and then input them into a VLM, Qwen3.5-VL-27B [ 32 ] . The VLM is requested to understand the whole video, speculate the missing contents ( i.e. , contents to out-paint like the two women and the indoor background), analyze which candidate frames may contain such missing contents, finally select the most suitable reference frames.
Figure 5: Qualitative comparison. Our LongSCOP could achieve stable video out-painting results and maintain the semantic consistency over a long time horizon. The region circled by the yellow box indicates the original input region. Zoom in for better views.
批判点评:收益依赖长视频里确实有可借的参考帧;遮挡区域从未露出时,检索无法提供缺失外观。短片 DAVIS 上 LongSCOP 的 PSNR 20.04 dB 低于 Infinite-Canvas 的 20.80 dB,不能把长视频优势泛化为所有扩画。
2. SceneShifter:换场景却不换人
SceneShifter: Training-free Multi-Scene Temporal Control for Audio-driven Human Animation | 悉尼大学;南洋理工大学;威斯康星大学麦迪逊分校 | arXiv:2611.09960
Figure 5: Scene-structured attention across model architectures. Rows denote queries and columns denote keys. Purple denotes scene-local interactions; orange denotes selected foreign video K/V from a shared memory. The example shows one head; text queries remain local.
Figure 10: Minute-long generation with HunyuanVideo-1.5. OneReel captures the requested scene changes, while dense generation remains visually repetitive.
本文贡献:MPEG-Diffusion 先以低分辨率跑前 k 步,再用冻结主干特征训练的线性头选高清区域;双分辨率注意力填补缺失键值,混合分辨率解码器合并潜变量。k 是开始分辨率分叉的去噪步,不是生成帧数。
Figure 2 : Overview of MPEG-Diffusion. The first k k of T T steps run at LR. At step k k , a linear head selects the HR region M M (part 2), and the remaining steps denoise M M at HR and the rest at LR. In dual-resolution attention (part 1), each query reads the K/V of its own resolution, with missing entries (dashed) filled from the other resolution by linear maps. The mixed-resolution decoder (part 3) merges the HR and LR latents into one HR video. Flames mark trained parts.
Figure 8 : Qualitative comparison on faces (Wan2.1 14B, k = 25 k=25 , same seed). Prompt: a young man steps into a surprise party. Both baselines show clear artifacts where the HR and LR regions meet, with one eye sharp and the other ghosted (arrows). Ours shows no visible degradation and is close to HR (more examples in Appendix A ).
Figure 3 : Overview of the proposed V2C-Edit pipeline. (1) Coupling-Controlled Velocities: Construct four editing velocities with different coupling behaviors. (2) Instruction-Level VLM Steering: Select the optimal velocity at each timestep using ILS. (3) Two-Stage Adaptive Masking: Apply spatial masking when semantic alignment is sufficient and the masked update maintains ILS. The VLM query templates are provided in Appendix F .
Figure 4 : Qualitative comparisons of text-based image editing methods. Our method achieves the highest fidelity to the editing instruction while maintaining strong background consistency with the source image.
本文贡献:STEER(Source-selection Training via Early-trajectory Emphasis and Recovery)分两阶段后训练:Emphasis 强化早期关键状态,Recovery 用自生成轨迹学习纠错;得到 SteerSound,并建立同时检查目标改到位和非目标保留的 Selective Audio Editing Benchmark。
Figure 1: Strengthening selective audio editing with STEER. Left: Existing editors may leave the target unchanged or modify non-target content. Middle: Early states shape final source selection, while later denoising tends to follow the established editing direction rather than revise it. Right: STEER addresses these failures through early-trajectory emphasis (Emphasis) and recovery training (Recovery), yielding SteerSound with more reliable selective editing.
Figure 3 : Bidirectional early-trajectory interventions on source selection. Corrective (top) and corruptive (bottom) prefix interventions are evaluated on 100 validation examples per task. Bars show per-step corrective or corruptive counts, while curves show the corresponding cumulative effects. For each task, k 80 k_{80} denotes the shortest intervention prefix length at which the cumulative effect reaches 80%. Removal consistently exhibits a shorter k 80 k_{80} than replacement.
Figure 2 : Overall framework of FADE. FADE consists of an offline profiling stage and an online inference stage. The offline stage calibrates and stores the optimal fractional orders in a static LUT, while the online stage alternates between full-compute and prediction steps using LUT-retrieved orders for feature forecasting. The upper-right panel illustrates the geometric contrast between the FADE trajectory and the integer-order trajectory in the latent space.
Figure 2 : Overview of SoftPAT3D. Each card shows the state of one prompt at that point of the pipeline, and its label names the step that produced it. (a) The prompt is expanded into a reference image, segmented and lifted as in PAT3D, and the support hidden under each deformable object is inpainted, regenerated and converted into a collision mesh (Sec. 3.4 ). (b) Each deformable object receives a rest shape from its manufacturing process under a size prior (Sec. 3.2 ), a material from the material library and a placement above the part the prompt names (Sec. 3.4 ). (c) One simulation brings the scene to equilibrium under a single contact barrier (Sec. 3.3 ), and the result passes the geometric checks and the prompt check (Sec. 3.5 ); a scene that fails enters placement optimization (Sec. 3.6 ).
Figure 5 : Vertical sections, cloth in red, furniture in grey, cloth inside the furniture in dark red, and a dashed line where the cloth floats above its support. Top: GPT-6 Astra; the blanket and the towel hang above their seats, the quilt runs through the mattress, the pillow sits inside the backrest, and the skirt ripples in mid-air. Bottom: ours; the cloth rests on the support, wraps its edges and hangs from them.
评论 (0)