SCoPE¶

Development line: project:scope · thread scope-public-resources
Last event: 2026-08-14 · 1 dated since 2026-08-14 · Researched: 2026-09-07 · confidence: medium

What it is¶

SCoPE — an image-to-video research release for teams that need camera-path control on Wan2.2-I2V-A14B.

• Inputs: takes a first frame, text prompt, and target camera trajectory.
• Positional coordinate: adds each token’s camera ray as a second coordinate without a separate control branch.
• Inference weights: ships self-contained weights for its Wan2.2-I2V-A14B base.

The checkpoint is about 67 GB. Local inference requires Python 3.11 and a CUDA-capable GPU. The documented default example runs 81 frames at 480×832. Use it when you can supply compatible camera poses; inaccurate geometry, out-of-distribution paths, occlusion, and fast motion can reduce quality.

Development line¶

• 2026-08-14 — SCoPE project resources were linked. On 2026-08-14, public links connected the SCoPE project site, TencentARC source repository, and Hugging Face page. Together, these links provided public entry points for project information, source code, and model-hosting materials. They do not establish a release date, version, authorship, or capabilities.

What changed¶

• 2026-06-25 — SCoPE paper v1 established ray-based positional encoding for video diffusion transformers.
• 2026-08-12 — arXiv v3 became the latest paper revision before the public links. The initial public code drop added inference and training code, examples, documentation, and a license.
• 2026-08-14 — The linked materials described an already-public SCoPE release. No distinct first-party release or paper change was found for this date.
• 2026-08-15 — The README documented the official Hugging Face demo.
• 2026-08-16 — The repository improved trajectory selection and demonstration workflows.
• 2026-08-22 — The repository restored missing vendored DiffSynth model sources. This repaired source layout rather than establishing new model performance.

How to use this¶

1. Use the official browser demo to inspect the image-to-video camera-control workflow before a local setup. — https://huggingface.co/TencentARC/SCoPE
2. For reproducible local inference, clone the repository and create its pinned environment with uv sync. The documented release uses Python 3.11, PyTorch 2.9.1, and CUDA 12.8. — https://github.com/TencentARC/SCoPE
3. Download TencentARC/SCoPE into a local checkpoint directory. Keep room for the approximately 67 GB checkpoint and cache. — https://huggingface.co/TencentARC/SCoPE
4. Run a bundled case with an explicit trajectory, such as omni-misty-forest and truck_right, for a repeatable baseline. — https://github.com/TencentARC/SCoPE
5. For a custom run, supply a first image, prompt, OpenCV camera-to-world pose array, and horizontal field of view. — https://github.com/TencentARC/SCoPE/blob/main/MODEL_CARD.md

Best practices¶

• Match the documented Python 3.11, PyTorch 2.9.1, and CUDA 12.8 environment when reproducing outputs. The project warns that PyTorch-version changes can alter numerical results. — https://github.com/TencentARC/SCoPE
• Use an explicit --trajectory for automated or repeatable jobs, and do not combine it with --camera_path. — https://github.com/TencentARC/SCoPE
• Validate custom poses as finite OpenCV camera-to-world arrays shaped [81, 3, 4] or [81, 4, 4]. Provide x_fov in horizontal radians and use xi=0 for a pinhole camera. — https://github.com/TencentARC/SCoPE/blob/main/MODEL_CARD.md
• Keep both checkpoint and Hugging Face cache on local storage. Use --vram_limit_gb when VRAM management or offloading is needed. — https://github.com/TencentARC/SCoPE
• Use source at or after commit 449343d for local checkout work because it restores omitted vendored DiffSynth model sources. Do not treat that repair as end-to-end inference proof. — https://github.com/TencentARC/SCoPE/commit/449343d217c54b98cb182cbc00f318f892bfecef

Superseded by this¶

• 2026-08-22 — Guidance based on a pre-449343d checkout is superseded for a complete source layout. The later commit restores omitted vendored DiffSynth model sources. This is a repository repair, not a new model release.

Still unknown¶

• No first-party page establishes an independent SCoPE release or paper change on 2026-08-14; the public code release and latest pre-event arXiv revision are dated 2026-08-12.
• Current README instructions include post-2026-08-14 improvements, so they are current operating guidance rather than proof of the exact initial workflow.
• No independent English- or Simplified-Chinese operator report was sufficient to elevate setup advice beyond the official documentation.
• SCoPE has unrelated name collisions in research and model search. The exact TencentARC repository, model ID, and arXiv ID define this entry.

Sources¶

Agent brief¶

• Subject: project:scope, thread scope-public-resources, 1 dated events 2026-08-14 → 2026-08-14.
• Practical note: See the sourced usage and practice sections above, including their limits.
• Confidence: medium. Dated supersedes above are the authority for what is obsolete.