[{"data":1,"prerenderedAt":55},["Reactive",2],{"news-story-news-1792311983093-h8zy9":3},{"id":4,"title_zh":5,"title_en":6,"summary_zh":7,"summary_en":8,"takeaways_zh":9,"takeaways_en":14,"content_zh":19,"content_en":20,"external_links":21,"source":32,"category":39,"impact_score":40,"views_count":41,"likes_count":42,"related_entities":43,"tags":47,"image_url":54},"news-1792311983093-h8zy9","4D/3D 视觉与自回归视频扩散加速突破：Gen4P 单次前馈 4D 感知、SUPRA-TEX 超对齐材质生成与港科大 ForcingBooster 削减 50% 去噪计算","4D/3D Vision and Fast Video Diffusion Breakthroughs: Gen4P Single-Pass 4D Perceiver, SUPRA-TEX 32.08 dB PBR Texturing, and HKUST ForcingBooster (-50.19% Patch NFE)","三维与四维视觉生成迎来三项底层革新：Gen4P（arXiv:2610.22371）将预训练视频生成模型改造为无需迭代去噪的单次前馈 4D 感知器，在 WorldTrack 基准全面超越 MonST3R 与 Any4D；华科与华为团队推出 SUPRA-TEX（arXiv:2610.22352），将 DTC-Tex 渲染 PBR PSNR 推高至 32.08 dB（超越 TRELLIS.2 的 28.60 dB）；港科大 ForcingBooster（arXiv:2610.12625）通过自适应分块去噪将自回归视频扩散 Patch NFE 砍半（-50.19%），实现无损画质下 1.30 倍端到端加速。","Three new architectures advance 4D perception, 3D PBR texturing, and real-time autoregressive video diffusion: Gen4P (arXiv:2610.22371) repurposes a video generator into a single-pass feed-forward 4D perceiver without iterative denoising; HUST and Huawei's SUPRA-TEX (arXiv:2610.22352) reaches 32.08 dB Shaded PBR PSNR on DTC-Tex; and HKUST's ForcingBooster (arXiv:2610.12625) cuts patch denoising NFE by 50.19% for a 1.30x end-to-end speedup across Wan2.1 14B and LongLive.",[10,11,12,13],"Gen4P（arXiv:2610.22371）将预训练视频生成器转化为无需迭代去噪的单次前馈 4D 感知器，在 WorldTrack 基准（ADT 91.93 APD / 0.112 EPE）全面超越 MonST3R、Any4D 与 OpenD4RT","华科与华为 SUPRA-TEX（arXiv:2610.22352）结合 MLLM 光照剥离、只读 3D VAE 隐空间锚点与交错流 Transformer，在 DTC-Tex 上斩获 32.08 dB PBR PSNR（超越 TRELLIS.2 的 28.60 dB 与 Hunyuan3D 2.1 的 26.80 dB）","港科大 ForcingBooster（arXiv:2610.12625）提出自适应分块去噪，在 Causal Forcing 上削减 50.19% Patch NFE，实现 1.30 倍端到端加速（11.15s 降至 8.56s）且语义分数无损（72.20 vs 72.21）","ForcingBooster 已在 Wan2.1 14B、LingBot-World-V2 与 LongLive 等主流自回归视频扩散架构上完成通用性验证",[15,16,17,18],"Gen4P (arXiv:2610.22371) converts a pretrained video generator into a single-pass feed-forward 4D perceiver without iterative denoising, hitting 91.93 APD / 0.112 EPE on WorldTrack ADT","HUST and Huawei's SUPRA-TEX (arXiv:2610.22352) reaches 32.08 dB Shaded PBR PSNR on DTC-Tex, beating Pixal3D (29.19 dB), TRELLIS.2 (28.60 dB), and Hunyuan3D 2.1 (26.80 dB)","HKUST's ForcingBooster (arXiv:2610.12625) reduces patch denoising NFE by 50.19% in autoregressive video diffusion for a 1.30x end-to-end speedup (11.15s to 8.56s) with lossless VBench semantics (72.20 vs 72.21)","ForcingBooster generalizes across Causal Forcing, Wan2.1 14B, LingBot-World-V2, and LongLive interactive world models","### 核心背景与行业痛点\n当前 3D/4D 视觉理解与实时交互视频生成受困于三大工程瓶颈：第一，在大规模生成数据上预训练的视频扩散模型虽蕴含丰富的时空几何先验，但将其用于 4D 点云重建与 3D 轨迹追踪时，传统多步去噪管线极慢且难以直接输出高精度度量几何；第二，现有 3D 资产材质生成（如 `TRELLIS.2`、`Hunyuan3D 2.1`）在多视角重投影或原生 3D 生成时常出现纹理接缝、模糊或几何-纹理错位；第三，自回归实时视频扩散（如 `Causal Forcing`、`Wan2.1`）对每一帧的所有图像块（Patch）均匀执行 4 步去噪，在静态背景区域浪费了海量算力。**`Gen4P`（`arXiv:2610.22371`）**、华中科技大学与华为联合研发的 **`SUPRA-TEX`（`arXiv:2610.22352`）** 以及香港科技大学提出的 **`ForcingBooster`（`arXiv:2610.12625`）** 分别攻克了上述难题。\n\n### 架构亮点与底层机制\n1. **`Gen4P`：无需迭代去噪的单次前馈连续时间 4D 感知器（`arXiv:2610.22371`）**：\n   - 摒弃缓慢的多步扩散采样，将预训练视频生成器重构为**单次前馈（Single-Pass Feed-Forward）**架构。通过引入连续时间按需查询机制（Continuous-Time On-Demand Queries）实现空间-时间特征解耦，并挂载冻结的视频深度 LoRA（Frozen Video-Depth LoRA）显式唤醒扩散隐空间中的三维度量几何先验，单次前向传播即可完成任意长序列 4D 重建与 3D 点追踪。\n2. **`SUPRA-TEX`：只读几何隐空间锚定与交错流 Transformer 材质补全（`arXiv:2610.22352`）**：\n   - 将 3D 纹理生成解耦为三步：首先利用微调的多模态大模型（MLLM）从参考图中分离环境光照、直接估算物理渲染（PBR）贴图；随后将可见表面属性直接重投影至原生 3D VAE 隐空间作为**只读几何锚点（Read-Only Anchors）**；最后通过交错流匹配 Transformer（Interleaved Flow Transformer）协同全局图像注意力与局部 3D 拓扑补全不可见区域。\n3. **`ForcingBooster`：块级自适应时间步截断加速自回归视频扩散（`arXiv:2610.12625`）**：\n   - 针对自回归视频生成中不同空间区域的帧间运动差异，引入轻量级块级退出门控（Patch-Wise Early Stopping）：静态或简单纹理 Patch 在早期去噪步即提前完成并退出计算，仅对高频动态前景 Patch 执行完整 4 步精炼。\n\n### 权威 Benchmark 与实测跑分对比\n- **`Gen4P` (`WorldTrack` 3D 点追踪权威基准)**：在 `ADT`、`Dynamic Replica (DR)` 与 `Panoptic Studio` 上，`Gen4P` (`Ours-II`) 的 `APD ↑ / EPE ↓` 分别达到 **91.93 / 0.112**、**88.18 / 0.130**（动态物体达 **92.52 / 0.090**）与 **82.11 / 0.149**，全面超越 `MonST3R`、`STv2`、`St4RTrack`、`Any4D` 与 `OpenD4RT`。\n- **`SUPRA-TEX` (`DTC-Tex` 3D PBR 材质生成基准)**：在 Shaded PBR 指标上斩获 **32.08 dB 全图 PSNR**、**23.67 dB 前景 PSNR**、**0.042 LPIPS** 与 **16.31 `FD_DINOv3`**，显著击败 `Pixal3D`（`29.19 dB`）、`TRELLIS.2`（`28.60 dB`）、`LumiTex`（`28.29 dB`）与 `Hunyuan3D 2.1`（`26.80 dB`）。\n- **`ForcingBooster` (`VBench` 自回归视频生成加速)**：在 `Causal Forcing` 上**削减 50.19% 的块级去噪函数评估次数（Patch NFE）**，包含 VAE 解码在内的端到端生成耗时从 `11.15s` 降至 **`8.56s`（1.30x 加速）**，同时保持 `VBench` 语义得分完全无损（**72.20** vs 原版 4 步基线 `72.21`），并已兼容 `Wan2.1 14B`、`LingBot-World-V2` 与 `LongLive`。\n\n### 开发者实战落地与开箱指南\n三项工作均已开源代码仓库（`yitongx/Gen4P`、`Supra-3D/SUPRA-Tex`、`IGL-HKUST/ForcingBooster`）：\n- 构建交互式世界模型或实时云渲染引擎的团队，可将 `ForcingBooster` 的块级动态步数分配插件直接集成至 `Wan2.1 14B` 或 `LongLive` 推理管线，立省一半 DiT 块算力；\n- 3D 游戏资产管线可利用 `SUPRA-TEX` 的只读隐空间重投影机制，解决正面角色贴图在 3D 化时的五官偏移与高光烘焙污染问题。","### 核心背景与行业痛点\n3D/4D spatial intelligence and real-time interactive video diffusion face three pressing bottlenecks: first, repurposing pretrained video diffusion models for 4D reconstruction and 3D point tracking typically relies on slow multi-step denoising; second, 3D texture generators (`TRELLIS.2`, `Hunyuan3D 2.1`) struggle with multi-view seams, baked-in lighting artifacts, and texture-geometry misalignment; third, autoregressive video diffusion models (`Causal Forcing`, `Wan2.1`) apply uniform 4-step denoising across every spatial patch, wasting compute on static backgrounds. **`Gen4P` (`arXiv:2610.22371`)**, HUST and Huawei's **`SUPRA-TEX` (`arXiv:2610.22352`)**, and HKUST's **`ForcingBooster` (`arXiv:2610.12625`)** tackle each of these bottlenecks directly.\n\n### 架构亮点与底层机制\n1. **`Gen4P`: Single-Pass Feed-Forward Continuous-Time 4D Perceiver (`arXiv:2610.22371`)**:\n   - Converts a pretrained video generator into a **single-pass feed-forward** 4D perception model without iterative denoising. It decouples spatial and temporal aggregation via continuous-time on-demand queries and injects a **frozen video-depth LoRA** to surface latent geometric priors for simultaneous 4D point cloud reconstruction and 3D point tracking.\n2. **`SUPRA-TEX`: Read-Only Native Latent Anchors & Interleaved Flow Transformer (`arXiv:2610.22352`)**:\n   - Decouples illumination removal and unseen surface synthesis: a fine-tuned MLLM first estimates de-lit PBR maps directly from the reference image; visible surface attributes are projected into the native 3D VAE latent space as **read-only geometric anchors**; and an **Interleaved Flow Transformer** completes occluded regions by combining global image attention with local 3D structural attention.\n3. **`ForcingBooster`: Adaptive Patch-Wise Denoising for Autoregressive Video Diffusion (`arXiv:2610.12625`)**:\n   - Dynamically allocates denoising steps per spatial patch in autoregressive video diffusion so static or low-motion patches exit early while complex dynamic patches receive full multi-step refinement.\n\n### 权威 Benchmark 与实测跑分对比\n- **`Gen4P` (`WorldTrack` 3D Point Tracking Benchmark)**: Achieves SOTA `APD ↑ / EPE ↓` of **91.93 / 0.112** on `ADT`, **88.18 / 0.130** (`92.52 / 0.090` dynamic) on `Dynamic Replica`, and **82.11 / 0.149** on `Panoptic Studio`, outperforming `MonST3R`, `STv2`, `St4RTrack`, `Any4D`, and `OpenD4RT`.\n- **`SUPRA-TEX` (`DTC-Tex` 3D PBR Benchmark)**: Reaches **32.08 dB Shaded PBR PSNR**, **23.67 dB foreground PSNR**, **0.042 LPIPS**, and **16.31 `FD_DINOv3`**, decisively beating `Pixal3D` (`29.19 dB`), `TRELLIS.2` (`28.60 dB`), `LumiTex` (`28.29 dB`), and `Hunyuan3D 2.1` (`26.80 dB`).\n- **`ForcingBooster` (`VBench` Autoregressive Video Acceleration)**: Cuts **patch denoising NFE by 50.19%** and reduces end-to-end latency (including VAE decoding) from `11.15s` to **`8.56s` (1.30x speedup)** while preserving the `Causal Forcing` Semantic score (**72.20** vs. `72.21` 4-step baseline), verified across `Wan2.1 14B`, `LingBot-World-V2`, and `LongLive`.\n\n### 开发者实战落地与开箱指南\nAll three projects provide open-source repositories (`yitongx/Gen4P`, `Supra-3D/SUPRA-Tex`, and `IGL-HKUST/ForcingBooster`):\n- Teams serving interactive world models or streaming video generators (`Wan2.1 14B`, `LongLive`, `LingBot-World-V2`) can plug in `ForcingBooster`'s patch-wise early exit schedule to halve patch denoising FLOPs without degrading visual semantics;\n- 3D asset pipelines can adopt `SUPRA-TEX` to preserve pixel-exact frontal textures while synthesizing relightable PBR materials for occluded regions.",[22,25,27,29],{"title":23,"url":23,"type":24},"https://github.com/yitongx/Gen4P","github",{"title":26,"url":26,"type":24},"https://github.com/Supra-3D/SUPRA-Tex",{"title":28,"url":28,"type":24},"https://github.com/IGL-HKUST/ForcingBooster",{"title":30,"url":30,"type":31},"https://gen4p.github.io/","other",{"platform":33,"author_name":34,"author_handle":35,"author_avatar":36,"tweet_url":37,"published_at":38},"twitter","Gen4P / HUST & Huawei / HKUST IGL","@huggingface","https://unavatar.io/x/huggingface","","2026-10-18T08:26:23.093Z","Research",91,43,0,{"models":44,"agents":45,"vendors":46},[],[],[],[48,49,50,51,52,53],"4D Vision","3D Generation","Video Diffusion","Gen4P","SUPRA-TEX","ForcingBooster","https://cdn-thumbnails.huggingface.co/social-thumbnails/papers/2610.22371.png",1792886318524]