ShotStream¶

Development line: project:shotstream · thread shotstream
Last event: 2026-03-31 · 1 dated since 2026-03-31 · Researched: 2026-09-06 · confidence: medium

What it is¶

ShotStream is a CUDA research implementation for creators and video-model developers who need to steer a multi-shot text-to-video narrative while it runs.

• Next-shot generation from historical shots and streaming prompts.
• Execution on Wan2.1-T2V-1.3B with local checkpoints.
• Separate global and local caches for cross-shot and within-shot continuity.

The authors report 15.95 FPS at 832×480 on one NVIDIA H200. Demos show 5–6 shots and nearly 500 frames. This is a local research reproduction path, not a hosted video service. Public demos do not guarantee parity with the internal training setup.

Development line¶

• 2026-03-31 — ShotStream public project, code, and model pages were linked. On 2026-03-31, a dated post pointed readers to the public project page, GitHub repository, and Hugging Face page. The evidence provides a public reference point for the development line. It does not establish a release version, capabilities, authorship, or the exact text of the original message.

What changed¶

• 2026-03-25 — ShotStream’s Hugging Face history records configuration and three checkpoint uploads.
• 2026-03-26 — ShotStream’s arXiv v1 was submitted.
• 2026-03-27 — ShotStream released its project page, arXiv version, training/inference code, and checkpoints.
• 2026-03-31 — ShotStream’s model card received a text-to-video tag plus expanded documentation, usage, and citation material.
• 2026-06-18 — ShotStream’s repository reported acceptance to ECCV 2026.

How to use this¶

1. Clone the repository and create the documented Python 3.10, CUDA, PyTorch, requirements, and FlashAttention environment. — https://github.com/KlingAIResearch/ShotStream
2. Download both the Wan2.1-T2V-1.3B base model and ShotStream checkpoints with Git LFS into the documented directories. — https://github.com/KlingAIResearch/ShotStream
3. Prepare prompt JSON with one global_caption and shot1 through shotN descriptions for the sequence. — https://github.com/KlingAIResearch/ShotStream/blob/main/demo/testdata/json/01.json
4. Run the documented autoregressive four-step multi-shot inference launcher. — https://github.com/KlingAIResearch/ShotStream

Superseded by this¶

• 2026-03-27 — Released training code, inference code, and checkpoints supersede the pre-release paper state.
• 2026-03-31 — Model-card documentation and pipeline metadata updated; this date brought no new weights or code.
• 2026-06-18 — ECCV 2026 acceptance supersedes pre-acceptance venue status without announcing a new runtime artifact.

Still unknown¶

• The links alone do not establish whether the 2026-03-31 event specifically referred to the same-day model-card update.
• Primary sources list no supported hosted endpoint, consumer-GPU minimum, VRAM requirement, or independent reproduction benchmark. The reported throughput is an author H200 measurement.
• Primary sources contain no first-party Simplified-Chinese operating guidance or independent practitioner discussion. Secondary summaries were not used as evidence.

Sources¶

Agent brief¶

• Subject: project:shotstream, thread shotstream, 1 dated events 2026-03-31 → 2026-03-31.
• 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.