SANA-Streaming — SANA Streaming¶

Development line: project:sana-streaming · thread sana-streaming
Last event: 2026-06-03 · 1 dated since 2026-06-03 · Researched: 2026-09-05 · confidence: high

What it is¶

SANA-Streaming is a 2B video-to-video editor for creators who need text-guided edits while retaining source motion and unedited regions.

• Long streaming editing: edits continuous video streams frame by frame.
• Short bidirectional editing: edits clips using both forward and backward context.

The reported real-time configuration runs 1280×704 at 24 end-to-end FPS on one RTX 5090. The public inference path uses BF16 checkpoints, not the paper’s MPQ deployment recipe. We use it for local long-form V2V work when its NVIDIA-oriented runtime and fixed 720p workflow fit the job.

Development line¶

• 2026-06-03 — SANA Streaming project page referenced. We linked the SANA Streaming development line to NVIDIA's SANA Streaming project page on 2026-06-03. This is a material reference point for the project's public development history. Available evidence does not establish a specific release, capability, or performance claim.

What changed¶

• 2026-06-03 — NVIDIA introduced SANA-Streaming as a 2B, 720p, minute-scale streaming video-to-video editor. The accompanying paper, dated 2026-05-28, reported 24 end-to-end FPS and 58 DiT FPS at 1280×704 on an RTX 5090.
• 2026-07 — NVIDIA released training support for bidirectional and distillation paths. This extended the initial inference-and-checkpoint release.

How to use this¶

As of 2026-06-03, start from the SANA Streaming project page to check official materials before relying on secondary summaries.

1. Install the repository environment, activate the sana Conda environment, and keep the documented pinned runtime versions for reproducible bidirectional inference. — https://github.com/NVlabs/Sana/blob/main/docs/sana_streaming.md
2. Run long_streaming with the released sana_streaming_ar.pth checkpoint for a source video, text instruction, and up to 969 decoded frames. Documented defaults use four denoising steps, CFG 1.0, two cached blocks, and sink-token caching. — https://github.com/NVlabs/Sana/blob/main/docs/sana_streaming.md
3. Use bidirectional_short with sana_bidirectional_short.pth when a five-second, 81-frame edit is sufficient. Documented defaults use 50 steps and CFG 6.0. — https://github.com/NVlabs/Sana/blob/main/docs/sana_streaming.md

Best practices¶

• Keep the mode-specific repository configuration and pinned package versions. Fused GDN kernels and the LTX-2 VAE path are version-sensitive. — https://github.com/NVlabs/Sana/blob/main/docs/sana_streaming.md
• Treat the RTX 5090 24-FPS figure as a system-co-designed MPQ result. It is not a guarantee for the released BF16 inference command. — https://arxiv.org/abs/2605.30409
• Use long streaming for sequence length and bidirectional mode for short edits. Released checkpoints, frame counts, sampling settings, and temporal assumptions differ between them. — https://github.com/NVlabs/Sana/blob/main/docs/sana_streaming.md

Superseded by this¶

• 2026-07 — The June inference-only understanding is incomplete. NVIDIA later released bidirectional and distillation training paths.

Still unknown¶

• The dated June event names no release commit and no day-specific announcement. We cannot independently pin the exact 2026-06-03 release scope beyond the June release notice and the 2026-05-28 paper date.
• The supplied response schema lacks event_findings and new_events fields. We represent their verified content in what_changed and supersedes.

Sources¶

Agent brief¶

• Subject: project:sana-streaming, thread sana-streaming, 1 dated events 2026-06-03 → 2026-06-03.
• Practical note: As of 2026-06-03, practitioners should use the SANA Streaming project page as the primary starting point for identifying the project's public materials before relying on secondary summaries.
• Confidence: high. Dated supersedes above are the authority for what is obsolete.