模型配方
选择模型,检查运行时,并复制与你的机器匹配的命令。
270 models运行时数据来自仓库 manifest
AnimateDiffShanghai AI Laboratory / CUHK MMLabAnimateDiff is the classic motion-module approach to text-to-video: it animates existing Stable Diffusion image models by inserting a trained temporal module, rather than training a video model from scratch.Text to videoPython 3.11 · 12.8AstraOpen-source research release (institution not recorded)Astra is an interactive world model with camera control built over a Wan2.1 backbone: from a single image it rolls out navigable world video following camera actions.Image to videoPython 3.11 · 12.8Being-H0.5BeingBeyondBeing-H0.5 is BeingBeyond's 2B-parameter cross-embodiment VLA policy.VlaPython 3.10 · CUDA 11.8+CameraCtrlThe Chinese University of Hong Kong / Shanghai AI LaboratoryCameraCtrl pioneered plug-and-play camera control for text-to-video diffusion: it trains a camera encoder over Plücker embeddings so trajectories can steer generation without retraining the base model.WorldPython 3.11 · 11.3CogVideoXZhipu AI (THUDM)CogVideoX is Zhipu AI's open video generation family, one of the most widely used open T2V/I2V baselines. This catalog entry serves three runnable variants — 2B text-to-video, 5B text-to-video, and 5B image-to-video — through in-tree runti…Text to videoPython 3.11 · 12.8Cosmos Transfer 2.5NVIDIACosmos Transfer 2.5 does controlled world-to-world translation: it takes structured control inputs (e.g. edge maps or simulation renders) and generates photoreal world video that follows them — the sim-to-real arm of the Cosmos family.Controlled videoPython 3.11 · 12.8Depth Any PanoramasInsta360 (Research Team)Depth Any Panoramas (DAP) is Insta360's metric depth estimator for panoramic images — full 360° depth from a single equirectangular input.Panoramic depth estimationPython 3.11 · 12.8Depth Anything 3 Metric PriorByteDance (Seed)Depth Anything 3 Metric Prior is the metric-depth head of ByteDance Seed's Depth Anything 3, integrated as an independent geometry prior with verified runner status. The recorded checkpoint is depth-anything/DA3METRIC-LARGE, and it is the…Monocular depth estimationPython 3.11 · 12.8Depth Anything V1TikTok / The University of Hong KongDepth Anything V1 is the foundation model for relative monocular depth that popularized large-scale unlabeled-data depth training.DepthPython 3.11 · 12.8Depth Anything V2TikTok / The University of Hong KongDepth Anything V2 is the sharper, synthetic-data-trained successor to V1, integrated here as a metric/monocular depth prior with verified runner status. The recorded checkpoint is depth-anything/Depth-Anything-V2-Large; the shared priors f…Monocular depth estimationPython 3.11 · 12.8DIAMONDUniversity of Geneva / University of Edinburgh / Microsoft ResearchDIAMOND (DIffusion As a Model Of eNvironment Dreams) trains reinforcement-learning agents entirely inside a diffusion world model — its Atari agents and playable CS:GO world model made diffusion world models famous (NeurIPS 2024).Diffusion world modelPython 3.11 · 12.8DualCamCtrlThe Hong Kong University of Science and Technology (Guangzhou) (EnVision Lab)DualCamCtrl controls video generation with a camera trajectory plus a depth image: the RGB and depth conditions together pin down scene geometry so the generated camera move stays physically grounded.WorldPython 3.11 · 12.8DUSt3RNAVER Labs EuropeDUSt3R made unconstrained two-view 3D reconstruction easy: feed it two images — no calibration, no poses — and it directly regresses aligned pointmaps.Feed forward multi view 3D reconstruction and geometry estimationPython 3.11 · 12.8FantasyWorldAlibaba (Amap, Fantasy-AMAP)FantasyWorld does geometry-consistent world modeling by unifying video and 3D prediction: from an image it produces camera-controllable world video that stays geometrically coherent. Two variants are recorded — the verified Wan2.1 pipeline…Geometry consistent world modelingPython 3.11 · 12.8FlashWorldXiamen University / TencentFlashWorld generates 3D scenes in seconds: from one image (plus optional prompt) it produces a camera-navigable 3D scene rather than a flat clip.Fast 3D scene generation from image or textPython 3.11 · 12.8GeoCalibETH Zurich (CVG)GeoCalib estimates camera calibration and the gravity direction from a single image by combining geometric optimization with deep learning.Camera calibrationPython 3.11 · 12.8GigaBrain-0GigaAIGigaBrain-0 is GigaAI's 3.5B VLA foundation policy with subtask prediction, released as GigaBrain-0-3.5B-Base and the 0.1 refresh under Apache-2.0.VlaPython 3.11 · 12.8HunyuanWorld-MirrorTencent (Hunyuan)HunyuanWorld-Mirror is Tencent's universal 3D reconstruction model: feed it images and it returns point clouds, Gaussian splats, depth, normals, and camera parameters in one pass.Three dimensionPython 3.11 · 12.8HY-WorldPlayTencent (Hunyuan)HY-WorldPlay is Tencent's real-time interactive world video model: starting from an image, it follows pose/action input to roll the world forward — the interactive arm of the Hunyuan world stack.WorldPython 3.11 · 12.8Infinite-WorldMeituan (MeiGen-AI)Infinite-World is MeiGen-AI's long-horizon world video model: from one image it navigates and extends a world over long durations.WorldPython 3.11 · 12.8InSpatio World 1.5inspatioNative inference reuses the existing causal Wan graph, UMT5 encoder, Wan VAE, causal sampler and in-tree DA3.Image to videoPython 3.11 · 12.8InSpatio-WorldInSpatio (formal institution not recorded)InSpatio-World is a 1.3B video-conditioned world generator: given a video and a target trajectory it synthesizes trajectory-conditioned novel views of the same world.Video conditioned world generationPython 3.11 · 12.8InSpatio-WorldFMInSpatioInSpatio-WorldFM is a 3D world foundation model for spatial intelligence and frame generation, released with confirmed code and weights (inspatio/WorldFM on Hugging Face, Apache-2.0).WorldPython 3.11 · 12.8Kairos SensenovaSenseTime (SenseNova)Kairos Sensenova is SenseTime's embodied world-model video generator: a 4B model covering text-, image-, and text+image-to-video with a robot-oriented distilled recipe.Text to videoPython 3.10 · 12.8LingBot-WorldRobbyant (Ant Group)LingBot-World is Robbyant's camera-controlled world video model.WorldPython 3.10 · 12.1LoGeROpen-source research release (institution not recorded)LoGeR reconstructs 3D scenes with camera trajectories and point clouds from input views, and is one of the verified 3D routes in this catalog: checkpoint-backed Studio inference produced finite poses, depth, and a 601,570-vertex point clou…3d reconstructionPython 3.11 · 12.8LongVie 2Shanghai AI Laboratory (Vchitect)LongVie 2 is the second-generation controllable long-video framework from the Vchitect team, and the verified member of the pair in this catalog: Studio's default four-GPU torchrun USP route completed a checkpoint-backed run on 2026-07-28,…Image to videoPython 3.10 · 12.8LyraNVIDIALyra is NVIDIA's generative world model family for exploration: from an image or video (Lyra-1) or from actions (Lyra-2) it produces an explorable world video.Open generative 3D and 4D world modelsPython 3.10 · 12.4Matrix-Game 2.0Skywork AI (Kunlun Tech)Matrix-Game 2.0 is Skywork's real-time-oriented interactive world model generation: an auto-regressive diffusion design that follows keyboard/mouse actions across game scenes. The in-tree runner was checked with a 15-latent-frame short run…WorldPython 3.11 · 12.8Matrix-Game 3.0Skywork AI (Kunlun Tech)Matrix-Game 3.0 is the third Skywork generation, pushing interactive world video to higher resolution with few-step inference.WorldPython 3.11 · 12.8Metric3DDJI / Zhejiang UniversityMetric3D recovers metric-scale depth from a single image by normalizing images into a canonical camera space, which lets one model generalize across cameras.Monocular depth estimationPython 3.11 · 12.8MolmoAct2Allen Institute for AI (AI2)MolmoAct2 is AI2's second-generation action reasoning model with an unusually broad Apache-2.0 checkpoint zoo: base, Think (depth-reasoning), DROID (Franka), Bimanual YAM, SO100/101, and LIBERO variants are all confirmed. WorldFoundry port…VlaPython 3.11 · 12.8MonST3RGoogle DeepMind / UC BerkeleyMonST3R extends the DUSt3R pointmap approach to dynamic scenes: it estimates per-timestep geometry directly from video, recovering shape and motion without a static-scene assumption.Geometry estimation in the presence of motionPython 3.11 · 12.8MotionCtrlTencent (ARC Lab)MotionCtrl is a unified, flexible motion controller for video generation: it controls camera motion and object motion independently in one model (SIGGRAPH 2024).Python 3.11 · 12.8NeoVerseCreateAI (formerly TuSimple)NeoVerse is a 4D world model built as a native assembly over canonical Wan components plus model-specific control and reconstruction roles: it enhances world generation with 4D-aware conditioning.4D world modeling from in the wild monocular videosPython 3.11 · 12.8OpenVLAOpen-source research release (institution not recorded)OpenVLA is the widely used open 7B vision-language-action model (MIT license): a Prismatic-based VLM fine-tuned on Open X-Embodiment robot demonstrations to predict discretized actions. Both openvla-7b and the LIBERO-10 fine-tune are confi…VlaPython 3.11 · 11.3Prior Depth AnythingOpen-source research release (institution not recorded)Prior Depth Anything is a depth-completion prior: it takes any coarse or sparse metric-depth prior plus an image and refines them into a dense, sharper metric depth map.Depth completionPython 3.11 · 12.8ReCamMasterKuaishou (Kling Team)ReCamMaster re-renders an existing video along a new camera trajectory — camera-controlled generative rendering from a single video.Camera controlled video rerenderingPython 3.11 · 12.8RoboFlamingoOpen-source research release (institution not recorded)RoboFlamingo adapts the OpenFlamingo vision-language model into a robot imitation policy: a lightweight policy head turns the VLM's visual-language features into actions.VlaPython 3.11 · Prepare-onlySCOPEOpen-source research release (institution not recorded)SCOPE is an interactive game world model over a Wan backbone: image plus per-step actions in, game-world video out (paper arXiv 2605.23345).Interactive world modelPython 3.11 · 12.8Show-oNational University of Singapore (Show Lab)Show-o (Xie et al., Show Lab at NUS with ByteDance, arXiv 2408.12528) is a single 1.3B transformer that unifies multimodal understanding and generation. It is initialized from Phi-1.5 and, unlike systems that attach a separate diffusion ex…Multimodal generationPython 3.11 · 11.3SolarisNew York University (VisionX)Solaris is a multi-agent interactive video world model from NYU's VisionX group (paper arXiv 2602.22208): multiple agents act inside one generated world.Video generationPython 3.11 · 12.8Splatt3RUniversity of Oxford (Active Vision Lab)Splatt3R predicts a 3D Gaussian splat scene directly from an uncalibrated image pair — zero shot, no camera parameters, no per-scene optimization.3d reconstructionPython 3.11 · 11.3StarVLAOpen-source research release (institution not recorded)StarVLA is a Lego-like modular VLA framework; this catalog entry is GPU-validated: the official QwenOFT predict_action ran on an 80GB CUDA GPU with the local Qwen3-VL-OFT-LIBERO-4in1 checkpoint and Qwen3-VL-4B base VLM. A WM4A world-action…VlaPython 3.11 · 12.8Step-Video-T2VStepFunStep-Video-T2V is StepFun's large open text-to-video model (30B parameters upstream, MIT weights at stepfun-ai/stepvideo-t2v). In WorldFoundry the public pipeline is assembled from native role-based components with the shared diffusion run…Text to videoPython 3.11 · 12.8VGGT-OmegaMeta AIVGGT-Omega is the extended VGGT release published under facebook/VGGT-Omega, covering 3D reconstruction, Gaussian splatting, and novel view synthesis.3d reconstructionPython 3.11 · 12.8VMemUniversity of OxfordVMem is an interactive video scene generation model with a surfel-indexed view memory: as you navigate a scene, it remembers which surfaces were seen from where, keeping new views consistent with the past.Video generationPython 3.11 · 12.8Wan2.1Alibaba (Tongyi Lab)Wan2.1 is Alibaba Tongyi Lab's open video generation family and the backbone of a large part of this catalog — many world-model and control entries build on it. Four variants are recorded here: T2V 1.3B, T2V 14B, and I2V 14B at 480p and 72…Text to videoPython 3.11 · 12.8Wan2.2Alibaba (Tongyi Lab)Wan2.2 is the next Wan generation, tracked separately from Wan2.1 because upstream versions code and checkpoints separately. The recorded runnable surface is the TI2V 5B variant — a compact text+image-to-video model. WorldFoundry wraps the…Text to videoPython 3.11 · 12.8Warp-as-HistoryOpen-source research release (institution not recorded)Warp-as-History (WAH) is a camera-controlled image-to-video method: it warps the input image along the requested camera path and treats those warps as the generation history that a video model then refines.Image to videoPython 3.10 · 12.1WorldCamKAIST / Adobe Research / MAUM AIWorldCam, from KAIST's CV Lab with Adobe Research and MAUM AI, is an interactive world model with camera-pose control for long-horizon game video.WorldPython 3.11 · 12.8WorldCrafterTencentarcReuses shared Helios transformer, scheduler, latent preparation and VGGT layers.Text to videoPython 3.11 · 12.8WorldCrafter BaseTencentarcWorldCrafter Base is a world model from Tencentarc. The catalog records it for text-to-video, image-to-video, camera-controlled-video, and world-model.Text to videoPython 3.11 · 12.8WorldCrafter FastTencentarcWorldCrafter Fast is a world model from Tencentarc. The catalog records it for text-to-video, image-to-video, camera-controlled-video, and world-model.Text to videoPython 3.11 · 12.8YUMEOpen-source research release (institution not recorded)YUME is an interactive world-video generator: from an image (and for YUME-1.5, text) it produces explorable first-person world videos that respond to keyboard-style motion input. Both recorded variants are verified in this catalog — YUME I…Image/video conditioned interactive world video generationPython 3.11 · 12.8ZeroScopeOpen-source community (cerspense)ZeroScope is a community fine-tune of ModelScopeT2V optimized for watermark-free 16:9 video, released in 576w and XL upscale checkpoints by the creator cerspense.Text to videoPython 3.11 · 12.8AC3DSnap Inc. (Snap Research)AC3D analyzes and improves camera control in video diffusion: it is a ControlNet-style plugin for CogVideoX. WorldFoundry includes the official inference adapter; checkpoint-backed inference requires the AC3D ControlNet and matching CogVid…Camera control videoPython 3.10 · 12.4Action Chunking TransformerOpen-source research release (institution not recorded)ACT (Action Chunking Transformer) is the classic imitation-learning policy behind the ALOHA line of work: instead of predicting one action at a time it predicts chunks of future actions, smoothing control and reducing compounding error. Wo…Action chunking policyPython 3.10 · CUDA 11.8+AHA-WAMOpen-source research release (institution not recorded)AHA-WAM is a world-action model for manipulation: paired video and action experts share a mixture cache so the policy benefits from predictive world modeling. The MIT-licensed SereneC/AHA-WAM-RoboTwin2.0 checkpoint (robotwin_ahawam.pt) is…VlaPython 3.11 · 12.8AlayaWorldOpen-source research release (institution not recorded)AlayaWorld is a camera-controlled interactive world model for long video: from an image it rolls out world video autoregressively with action control and DA3-backed spatial memory.Image to videoPython 3.11 · 12.8
如何读懂就绪度
每个模型都有一个就绪度徽章。徽章由相互独立的 manifest 记录计算得出,缺失的证据永远不会被四舍五入成更强的声明。
- Runner verified — 至少一条路径已记录 WorldFoundry runner parity 产物。请选择证据对应的 variant 和配置。
- Integrated — 已绑定可运行 Pipeline,但端到端 GPU parity 仍在验证。可以运行;在 parity 记录之前,输出应视为未验证。
- Runtime ported — 运行时代码已在仓库中,路径还在接线。检查 manifest;运行可能需要手动步骤。
- Profile only — 只记录了 runtime profile,没有绑定 Pipeline。页面可用于溯源与规划;不会显示运行命令。
- Planned — 条目只记录了意向与上游来源。可以跟踪;目前没有可运行内容。
- Blocked — 已记录的 blocker(许可、权重、上游)阻止了集成。阅读 manifest 备注了解具体 blocker。
如何选择模型
- 按家族和任务过滤。 使用分类标签(视频、世界模型、3D 与 4D、具身智能、托管 API)和搜索框——搜索会匹配模型名、别名、提供方和任务 id。
- 优先选择 verified 或 integrated 条目。 这些条目已绑定 Pipeline;其余条目只是参考资料。
- 检查运行时信息。 独立环境需要单独安装;统一环境的模型共享一个 conda 环境。Python 与 CUDA 可以在安装前告诉你驱动栈是否匹配。
- 打开模型页面。 每个页面都会解释模型是什么、来自哪里、输入/输出契约、精确的环境、带固定 revision 的权重仓库,以及为所选 variant 生成可复制运行命令的命令构建器。
每个模型页面包含什么
模型页面由自动生成的运行时记录与模型专属文章组成。docs/fumadocs/scripts/generate-model-recipes.py 刷新 recipe 数据和标记为 pageSource: generated 的页面;手写文章单独维护。每个页面会把下面几类声明分开呈现:
- Catalog
- 稳定 model ID、任务、官方来源(论文、项目、代码、权重)、别名和集成状态。
- 发表机构
- 正式发表机构(企业优先,其次大学或实验室)及其类型。GitHub 账号永远不会被当作品牌展示;若仓库未记录正式机构,页面会明确说明。
- Runtime profile
- 输入契约、artifact 契约、backend stage 和运行时就绪状态。
- Environment
- Python、CUDA tier、固定依赖、验证 import,以及独立/统一环境路由。
- Pipeline binding
- WorldFoundry 使用的 runner 与每个 variant 的具体 pipeline target。
- Benchmarks
- Manifest 推荐或其证据记录中提到的评测,链接到 Benchmark Hub。
- Evidence
- 原生 demo 与 WorldFoundry runner 的 parity;只有 catalog 记录并不代表 GPU 已验证。
缺失值会明确显示为未记录。Catalog manifest 可以携带人工撰写的双语 docs: 块,记录概述、架构、发表机构与推荐评测;没有此块的条目使用已有 catalog 和 runtime 元数据。使用说明与 blocker 写在每个模型页的运行时契约中。Catalog tasks 描述模型能力;命令构建器使用 CLI 实际接受的 task-profile ID 和参数。验证记录只适用于所注明的路径与配置。
常用命令
# 检查完整 catalog manifest。
worldfoundry-eval zoo model-show \
--model-id <model-id> \
--include-manifest \
--json
# 不下载,只解析 checkpoint、凭据与本地资产 blocker。
worldfoundry-eval zoo model-download \
--model-id <model-id> \
--check-local \
--json
# 为一个 runtime profile 解析并安装环境 manifest。
bash scripts/setup/model_env_install.sh --model <model-id>
# 为一个 runtime profile 下载声明的 checkpoint。
bash scripts/inference/prepare_model_infer.sh <model-id> --download
# 运行一个已记录的 task profile(每个模型页面会生成精确的参数)。
worldfoundry-eval run <model-id> --pipeline.task-profile <task> --json在文档中的位置
- 运行推理——选定模型后完整的 request 文件与 artifact 流程。
- 环境参考——统一 vs. 独立环境路由的决策方式。
- 评测与 Benchmark Hub——为模型输出打分;模型页面会链接到其 manifest 推荐或提及的评测。
- 添加模型——如何编写新的 catalog manifest(及其可选
docs:块)。





















