How does WorldMem store and retrieve earlier frames so a generated Minecraft or RealEstate10K world stays consistent after the camera leaves and returns, what revisit numbers does it report, and what does it cost?

WorldMem keeps every generated frame as a token-level memory unit with a 5-number pose and a timestamp, picks up to a handful of units per step by field-of-view overlap and time difference (greedy, with a similarity filter), and lets the generating tokens cross-attend to them with pose/time embeddings. Its revisit tests compare regenerated frames against ground-truth frames with PSNR, LPIPS and reconstruction FID: 100 frames generated after a 600-frame memory bank on a custom Minecraft set, and five start=end trajectories over 100 RealEstate10K scenes. These are paper_result numbers on simulator poses; the paper's own tables disagree with each other and with the repository's expected-results block on some cells. Curator qualifications: Excerpts present (E2 differs only by an extraction hyphen 'field-ofview'; E6 present in layout line 415 'DF 17.32 0.4376 51.28'). Inconsistencies verified: the configuration with LPIPS 0.1429 / rFID 15.37 reports PSNR 23.98 in Tables 1, 2 and 3 but 25.32 in Table 7 (length 8) and Table 8 (ground-truth poses); the README 'Expected Results' give PSNR 24.01, LPIPS 0.1667, FID 15.13 ('FID is computed over 5000 frames'). Additional anomaly: Table 7 length 4 reports PSNR 14.68 with LPIPS 0.1568 and rFID 16.54, better than length 1 (22.18 / 0.1899 / 20.47) on LPIPS and rFID but far worse on PSNR, which is likely a transcription error (not established). Cost: Supplementary says 'The generation cost (20 denoising steps) is ∼0.9s per frame' without naming hardware in that sentence; the H200 baseline comparison gives inference 'Speed (it/s)' 1.03 without memory and 0.89 with memory (unit 'it' not defined), so 'about 0.9 s per frame ... on one H200' combines two statements. Memory bank 21 MB for 600 tokens confirmed. Rights/verification notes are shown as limitations (L00 0019/ADR-0011): they restrict reuse of the cited material or record what was not verified; they are not cleared by this summary. Limitations: [component_reuse_restriction] Weights and dataset licence terms absent; base Oasis weights terms not read; .ckpt loading risk (C3R-0095). Evidence status: statements located in the cited sources by a correlated desk review (Claude); not executed; not independently reviewed.

{
  "citations": [
    {
      "url": "https://arxiv.org/abs/2504.12369v3",
      "note": "arXiv 2504.12369v3 PDF (pdftotext and -layout): Sec. 3.3-3.4, 4.1, Tables 1-3 (layout lines 402-416), Tables 7-8 (lines 909-921), Supplementary 7.1-7.2; WorldMem README (equals producer pin) 'Expected Results' lines 133-141; LICENSE.md S-Lab License 1.0; HF API zeqixiao/worldmem_checkpoints"
    },
    {
      "url": "https://github.com/xizaoqu/WorldMem",
      "note": "arXiv 2504.12369v3 PDF (pdftotext and -layout): Sec. 3.3-3.4, 4.1, Tables 1-3 (layout lines 402-416), Tables 7-8 (lines 909-921), Supplementary 7.1-7.2; WorldMem README (equals producer pin) 'Expected Results' lines 133-141; LICENSE.md S-Lab License 1.0; HF API zeqixiao/worldmem_checkpoints"
    },
    {
      "url": "https://huggingface.co/zeqixiao/worldmem_checkpoints",
      "note": "arXiv 2504.12369v3 PDF (pdftotext and -layout): Sec. 3.3-3.4, 4.1, Tables 1-3 (layout lines 402-416), Tables 7-8 (lines 909-921), Supplementary 7.1-7.2; WorldMem README (equals producer pin) 'Expected Results' lines 133-141; LICENSE.md S-Lab License 1.0; HF API zeqixiao/worldmem_checkpoints"
    }
  ]
}