logbook.json

{
  "schema_version": 2,
  "title": "Reproduction: Both Semantics and Reconstruction Matter: Making Representation Encoders Ready for Text-to-Image Generation and Editing",
  "emoji": "🎯",
  "space_id": "rcgp/repro-psvae",
  "paper": {
    "arxiv_id": "2512.17909"
  },
  "tags": [
    "icml2026-repro",
    "paper-UmlXZ4t44h"
  ],
  "updated_at": "2026-07-22T13:38:56+00:00",
  "root": {
    "slug": "index",
    "title": "Reproduction: Both Semantics and Reconstruction Matter: Making Representation Encoders Ready for Text-to-Image Generation and Editing",
    "file": "pages/index.md",
    "children": [
      {
        "slug": "executive-summary",
        "title": "Executive summary",
        "file": "pages/executive-summary/page.md",
        "children": []
      },
      {
        "slug": "claim-1-ps-vae-maps-768-dimensional-dinov2-features-into-a-compact-96-channel-latent-space-at-16x16-spatial-downsampling-reducing-off-manifold-outliers-via-kl-regularization-in-the-semantic-vae-stage-section-3-3",
        "title": "Claim 1: PS-VAE maps 768-dimensional DINOv2 features into a compact 96-channel latent space at 16x16 spatial downsampling, reducing off-manifold outliers via KL regularization in the Semantic VAE stage (Section 3.3).",
        "file": "pages/claim-1-ps-vae-maps-768-dimensional-dinov2-features-into-a-compact-96-channel-latent-space-at-16x16-spatial-downsampling-reducing-off-manifold-outliers-via-kl-regularization-in-the-semantic-vae-stage-section-3-3/page.md",
        "children": []
      },
      {
        "slug": "claim-2-ps-vae-achieves-an-rfid-of-0-203-and-psnr-of-28-79-db-improving-substantially-over-rae-rfid-619-psnr-19-20-and-mar-vae-rfid-534-psnr-26-18-on-the-reconstruction-and-generation-benchmark-table-2",
        "title": "Claim 2: PS-VAE achieves an rFID of 0.203 and PSNR of 28.79 dB, improving substantially over RAE (rFID 619, PSNR 19.20) and MAR-VAE (rFID 534, PSNR 26.18) on the reconstruction and generation benchmark (Table 2).",
        "file": "pages/claim-2-ps-vae-achieves-an-rfid-of-0-203-and-psnr-of-28-79-db-improving-substantially-over-rae-rfid-619-psnr-19-20-and-mar-vae-rfid-534-psnr-26-18-on-the-reconstruction-and-generation-benchmark-table-2/page.md",
        "children": []
      },
      {
        "slug": "claim-3-ps-vae-96-channel-attains-a-geneval-score-of-76-56-and-editing-reward-of-0-222-exceeding-rae-geneval-71-27-editing-reward-0-059-and-mar-vae-geneval-75-75-editing-reward-0-056-table-2",
        "title": "Claim 3: PS-VAE (96-channel) attains a GenEval score of 76.56 and Editing Reward of 0.222, exceeding RAE (GenEval 71.27, Editing Reward 0.059) and MAR-VAE (GenEval 75.75, Editing Reward 0.056) (Table 2).",
        "file": "pages/claim-3-ps-vae-96-channel-attains-a-geneval-score-of-76-56-and-editing-reward-of-0-222-exceeding-rae-geneval-71-27-editing-reward-0-059-and-mar-vae-geneval-75-75-editing-reward-0-056-table-2/page.md",
        "children": []
      },
      {
        "slug": "claim-4-scaling-the-generative-model-from-653m-to-1708m-parameters-raises-ps-vae96c-s-geneval-score-from-76-56-to-78-14-with-faster-convergence-curves-than-baseline-feature-spaces-section-4-3-figure-10",
        "title": "Claim 4: Scaling the generative model from 653M to 1708M parameters raises PS-VAE96c's GenEval score from 76.56 to 78.14, with faster convergence curves than baseline feature spaces (Section 4.3, Figure 10).",
        "file": "pages/claim-4-scaling-the-generative-model-from-653m-to-1708m-parameters-raises-ps-vae96c-s-geneval-score-from-76-56-to-78-14-with-faster-convergence-curves-than-baseline-feature-spaces-section-4-3-figure-10/page.md",
        "children": []
      },
      {
        "slug": "claim-5-ps-vae-generalizes-to-a-siglip2-encoder-achieving-a-geneval-score-of-77-14-while-limiting-understanding-capability-degradation-on-mme-perception-1685-to-1652-section-4-4-table-3",
        "title": "Claim 5: PS-VAE generalizes to a SigLIP2 encoder, achieving a GenEval score of 77.14 while limiting understanding-capability degradation on MME-Perception (1685 to 1652) (Section 4.4, Table 3).",
        "file": "pages/claim-5-ps-vae-generalizes-to-a-siglip2-encoder-achieving-a-geneval-score-of-77-14-while-limiting-understanding-capability-degradation-on-mme-perception-1685-to-1652-section-4-4-table-3/page.md",
        "children": []
      },
      {
        "slug": "conclusion",
        "title": "Conclusion",
        "file": "pages/conclusion/page.md",
        "children": []
      }
    ]
  },
  "traces": [],
  "workspace": {
    "file": "workspace.json",
    "file_count": 0,
    "total_size": 0,
    "bucket_id": null
  },
  "agent_view_tokens": 6610,
  "trace_view_tokens": 148,
  "workspace_view_tokens": 8,
  "revision": "47195951face076cefe6",
  "traces_ref": {
    "repo_id": "rcgp/repro-psvae-traces",
    "repo_type": "dataset",
    "repo_url": "https://huggingface.co/datasets/rcgp/repro-psvae-traces",
    "private": true
  },
  "trace_dataset": "https://huggingface.co/datasets/rcgp/repro-psvae-traces"
}