SAMA¶
★★★★★ Basic
Development line: project:sama · thread sama
Last event: 2026-03-20 · 1 dated since 2026-03-20 · Researched: 2026-09-05 · confidence: high
What it is¶
An open-source instruction-guided video editing model for Wan users who change objects, style, or text in a video while keeping motion.
- Semantic anchoring plans edits across keyframes.
- Motion alignment preserves temporal dynamics.
- SAMA-14B checkpoint and an official ComfyUI workflow are available.
Requires Linux, an NVIDIA GPU, Python 3.10, a CUDA 12.1-compatible environment, and Wan2.1-T2V-14B. This is a local inference stack on top of Wan, not a hosted service.
Development line¶
- 2026-03-20 — SAMA public project, source, and model resources were linked. On 2026-03-20, a dated SAMA message linked the project's public website, source repository, and the Hugging Face page for SAMA-14B. These links establish that the three public resources were associated with SAMA on that date, without establishing further claims about the model or project.
What changed¶
- 2026-03-20 — SAMA paper published; preprint submitted on 2026-03-19 with 24 pages and 12 figures.
- 2026-03-21 — SAMA-14B checkpoint released; SAMA-5B still marked as Coming soon.
- 2026-03-24 — Official SAMA-ComfyUI workflow opened.
- 2026-06-20 — SAMA accepted to ECCV 2026.
- 2026-06-26 — SAMA-edit-filtered-1M metadata set published.
How to use this¶
As of 2026-03-20, we use the linked SAMA website, source repository, and SAMA-14B model page as starting points for evaluation, verifying capabilities, licensing, and usage requirements from those resources before adoption.
- Clone the repository, create a Python 3.10 environment, and install dependencies. — https://github.com/Cynthiazxy123/SAMA
- Download SAMA-14B and prepare the full local Wan2.1-T2V-14B directory. — https://huggingface.co/syxbb/SAMA-14B
- Set MODEL_ROOT, STATE_DICT, SRC_VIDEO, PROMPT, and OUTPUT_DIR in infer_sh/run_sama.sh, then run the script. — https://github.com/Cynthiazxy123/SAMA
- For a node-based workflow, use the official ComfyUI integration with the Wan base model and SAMA-14B. — https://github.com/Cynthiazxy123/SAMA
Best practices¶
- Verify the base Wan2.1-T2V-14B directory is complete: the script intentionally stops if files are missing. — https://github.com/Cynthiazxy123/SAMA
- Use the source FPS; set --fps explicitly when it is missing. — https://github.com/Cynthiazxy123/SAMA
- Account for automatic padding of input frames to the Wan 4k+1 requirement. — https://github.com/Cynthiazxy123/SAMA
Superseded by this¶
- 2026-03-20 — paper-only state: SAMA-14B is available as of 2026-03-21, and the official ComfyUI workflow is available as of 2026-03-24.
Still unknown¶
- Official materials do not provide measured VRAM, inference speed, or supported resolutions. Hardware planning requires separate testing of the configuration and workflow.
Sources¶
| source | title | read |
|---|---|---|
| https://cynthiazxy123.github.io/SAMA/ | SAMA: Factorized Semantic Anchoring and Motion Alignment for Instruction-Guided Video Editing | 2026-09-05 |
| https://github.com/Cynthiazxy123/SAMA | Cynthiazxy123/SAMA — official inference code | 2026-09-05 |
| https://huggingface.co/syxbb/SAMA-14B | syxbb/SAMA-14B | 2026-09-05 |
| https://arxiv.org/abs/2603.19228 | SAMA: Factorized Semantic Anchoring and Motion Alignment for Instruction-Guided Video Editing | 2026-09-05 |
Agent brief¶
- Subject:
project:sama, threadsama, 1 dated events 2026-03-20 → 2026-03-20. - Practical note: As of 2026-03-20, practitioners can use the linked SAMA website, source repository, and SAMA-14B model page as the primary starting points for evaluation, while verifying capabilities, licensing, and usage requirements from those resources before adoption.
- Confidence: high. Dated supersedes above are the authority for what is obsolete.