How does VMem index past views with surfels, what is its cycle-trajectory revisit test, and which parts of its release are open, gated or conflicting?
VMem keeps the past views as images but indexes them through a coarse surfel model: each surfel stores which views saw it, new target cameras render the surfels to vote for the top-K most relevant past views, and an image-set generator (Stable Virtual Camera, SEVA) generates the next views from those. Its main contribution to evaluation is a cycle-trajectory protocol (go out, return along the same path, score every 10 frames of the return) with LPIPS/PSNR/SSIM plus rotation and translation distance of camera poses re-estimated from the generated frames. The tables contain rows that are identical across methods, which should be checked before reuse. Code is MIT-labelled but vendors a CC BY-NC-SA component, and the weights sit behind a Hugging Face access gate.
Structured fields
{
"citations": [
{
"url": "https://arxiv.org/abs/2506.18903v3",
"note": "arXiv 2506.18903v3 PDF (pdftotext -layout), page 6 Tables 2 and 4; Sec. 3.1, 4.1, 4.3, 4.4, App. D; arXiv abs page observed 2026-10-22: submission history v1 23 Jun 2025, v2 10 Jul 2025, v3 14 Aug 2025 (v3 latest); vmem 39291e4f LICENSE: 'MIT License' 'Copyright (c) 2025 Runjia Li'; vmem 39291e4f extern/CUT3R/LICENSE; vmem 39291e4f README.md (equals producer pin); HF API models/liguang0115/vmem (equals producer pin)"
},
{
"url": "https://github.com/runjiali-rl/vmem",
"note": "arXiv 2506.18903v3 PDF (pdftotext -layout), page 6 Tables 2 and 4; Sec. 3.1, 4.1, 4.3, 4.4, App. D; arXiv abs page observed 2026-10-22: submission history v1 23 Jun 2025, v2 10 Jul 2025, v3 14 Aug 2025 (v3 latest); vmem 39291e4f LICENSE: 'MIT License' 'Copyright (c) 2025 Runjia Li'; vmem 39291e4f extern/CUT3R/LICENSE; vmem 39291e4f README.md (equals producer pin); HF API models/liguang0115/vmem (equals producer pin)"
},
{
"url": "https://huggingface.co/liguang0115/vmem",
"note": "arXiv 2506.18903v3 PDF (pdftotext -layout), page 6 Tables 2 and 4; Sec. 3.1, 4.1, 4.3, 4.4, App. D; arXiv abs page observed 2026-10-22: submission history v1 23 Jun 2025, v2 10 Jul 2025, v3 14 Aug 2025 (v3 latest); vmem 39291e4f LICENSE: 'MIT License' 'Copyright (c) 2025 Runjia Li'; vmem 39291e4f extern/CUT3R/LICENSE; vmem 39291e4f README.md (equals producer pin); HF API models/liguang0115/vmem (equals producer pin)"
}
]
}Evidence claims
- [sourced_fact] VMem keeps the past views as images but indexes them through a coarse surfel model: each surfel stores which views saw it, new target cameras render the surfels to vote for the top-K most relevant past views, and an image-set generator (Stable Virtual Camera, SEVA) generates the next views from those. Its main contribution to evaluation is a cycle-trajectory protocol (go out, return along the same path, score every 10 frames of the return) with LPIPS/PSNR/SSIM plus rotation and translation distance of camera poses re-estimated from the generated frames. The tables contain rows that are identical across methods, which should be checked before reuse. Code is MIT-labelled but vendors a CC BY-NC-SA component, and the weights sit behind a Hugging Face access gate. · locator: arXiv 2506.18903v3 PDF (pdftotext -layout), page 6 Tables 2 and 4; Sec. 3.1, 4.1, 4.3, 4.4, App. D; arXiv abs page observed 2026-10-22: submission history v1 23 Jun 2025, v2 10 Jul 2025, v3 14 Aug 2025 (v3 latest); vmem 39291e4f LICENSE: 'MIT License' 'Copyright (c) 2025 Runjia Li'; vmem 39291e4f extern/CUT3R/LICENSE; vmem 39291e4f README.md (equals producer pin); HF API models/liguang0115/vmem (equals producer pin) source
Rights
| component | license | hosting | redistribution | modification | training | Evidence and notes |
|---|---|---|---|---|---|---|
| documentation | unknown | permitted | permitted | unknown | unknown | Applies to this original summary only, not to the cited sources. |