Add files using upload-large-folder tool

public_pr102_intake_20260505_auto/INTAKE_LOG.md

@@ -0,0 +1,16 @@
+# PR #102 intake — hnerv_lc_v2_scale095_rplus1 (score 0.195)
+Started UTC: 2026-05-05T22:44:44.008380+00:00
+
+✓ pr_metadata.json (head_sha=1e330ec5633539c48278ce3cc96d2b15ea7a9eac, author=EthanYangTW)
+✓ pr_body.md (1872 chars)
+  head_repo release asset: https://github.com/EthanYangTW/comma_video_compression_challenge/releases/download/qpose14-r55-segactions-minp/archive.zip
+  head_repo release asset: https://github.com/EthanYangTW/comma_video_compression_challenge/releases/download/v0.2/archive_220.zip
+  head_repo release asset: https://github.com/EthanYangTW/comma_video_compression_challenge/releases/download/v2-hnerv-lc-scale095/archive.zip
+  head_repo release asset: https://github.com/EthanYangTW/comma_video_compression_challenge/releases/download/v1-hnerv-finetuned/archive.zip
+  head_repo release asset: https://github.com/EthanYangTW/comma_video_compression_challenge/releases/download/roi_v2_archive/archive.zip
+  head_repo release asset: https://github.com/EthanYangTW/comma_video_compression_challenge/releases/download/roi_v2/archive.zip
+  head_repo release asset: https://github.com/EthanYangTW/comma_video_compression_challenge/releases/download/qpose14-r55-segactions-minp-v2/archive.zip
+  head_repo release asset: https://github.com/EthanYangTW/comma_video_compression_challenge/releases/download/qpose14-r55/archive.zip
+✓ archive.zip downloaded (276481 bytes, sha256 03a2afd5fe92c93a...)
+
+Status: archive=downloaded source=skipped_exists

public_pr102_intake_20260505_auto/archive_provenance.json

@@ -0,0 +1,19 @@
+{
+  "candidates_tried": [
+    "https://github.com/EthanYangTW/comma_video_compression_challenge/releases/download/qpose14-r55-segactions-minp/archive.zip"
+  ],
+  "attempts": [
+    {
+      "url": "https://github.com/EthanYangTW/comma_video_compression_challenge/releases/download/qpose14-r55-segactions-minp/archive.zip",
+      "success": true,
+      "result": "03a2afd5fe92c93a9b7b7e43625158a73b455f0cfbca82d278008a728db78746",
+      "zip_members": [
+        "p"
+      ]
+    }
+  ],
+  "archive_sha256": "03a2afd5fe92c93a9b7b7e43625158a73b455f0cfbca82d278008a728db78746",
+  "archive_size_bytes": 276481,
+  "status": "downloaded",
+  "fetched_at_utc": "2026-05-05T22:44:45.122186+00:00"
+}

public_pr102_intake_20260505_auto/pr_body.md

@@ -0,0 +1,45 @@
+# submission name:
+hnerv_lc_v2_scale095_rplus1
+
+# upload zipped `archive.zip`
+Included in this PR under `submissions/hnerv_lc_v2_scale095_rplus1/archive.zip`.
+
+# report.txt
+CPU-only run with `CUDA_VISIBLE_DEVICES=''` so inflation also used CPU:
+```
+=== Evaluation config ===
+  batch_size: 16
+  device: cpu
+  num_threads: 2
+  prefetch_queue_depth: 4
+  report: submissions/hnerv_lc_v2_scale095_rplus1/report.txt
+  seed: 1234
+  submission_dir: submissions/hnerv_lc_v2_scale095_rplus1
+  uncompressed_dir: /root/comma_video_compression_challenge/videos
+  video_names_file: /root/comma_video_compression_challenge/public_test_video_names.txt
+=== Evaluation results over 600 samples ===
+  Average PoseNet Distortion: 0.00003460
+  Average SegNet Distortion: 0.00057602
+  Submission file size: 178,981 bytes
+  Original uncompressed size: 37,545,489 bytes
+  Compression Rate: 0.00476704
+  Final score: 100*segnet_dist + √(10*posenet_dist) + 25*rate = 0.20
+```
+Exact score from rounded CPU report components: `0.1953791765`.
+
+Fast PyAV/CUDA scorer during tuning gave exact `0.194986956` with Pose `0.000033274` and Seg `0.000575697`.
+
+# does your submission require gpu for evaluation (inflation)?
+no
+
+# did you include the compression script? and want it to be merged?
+yes. `compress.sh` reproduces `archive.zip` by fetching the unchanged PR #100 release archive payload and verifying SHA256. Yes, please merge this submission if accepted.
+
+# additional comments
+Built on top of @BradyMeighan's `hnerv_lc_v2` PR #100, which is built on top of @EthanYangTW's `hnerv_muon_finetuned_from_pr95` PR #98 and @AaronLeslie138's `hnerv_muon` PR #95.
+
+Changes from PR #100:
+- retuned latent correction scale from `0.0100` to `0.0095`;
+- added a zero-byte decode-side nudge: frame 0 red channel `+1`.
+
+The archive payload is unchanged from PR #100; only inference-time code constants changed.

public_pr102_intake_20260505_auto/pr_metadata.json

@@ -0,0 +1,16 @@
+{
+  "additions": 367,
+  "author": "EthanYangTW",
+  "changed_files": 7,
+  "closed_at": "2026-05-04T20:08:31Z",
+  "created_at": "2026-05-04T11:54:32Z",
+  "deletions": 0,
+  "head_repo": "EthanYangTW/comma_video_compression_challenge",
+  "head_sha": "1e330ec5633539c48278ce3cc96d2b15ea7a9eac",
+  "leaderboard_name": "hnerv_lc_v2_scale095_rplus1",
+  "leaderboard_score": 0.195,
+  "merged_at": "2026-05-04T20:08:31Z",
+  "pr_number": 102,
+  "state": "closed",
+  "title": "hnerv_lc_v2_scale095_rplus1 submission (0.19538 CPU)"
+}

public_pr103_intake_20260505_auto/INTAKE_LOG.md

@@ -0,0 +1,8 @@
+# PR #103 intake — hnerv_lc_ac (score 0.195)
+Started UTC: 2026-05-05T22:44:43.528228+00:00
+
+✓ pr_metadata.json (head_sha=d20270750e603346056f4e867cddd4744272c9d8, author=rem2)
+✓ pr_body.md (1542 chars)
+✓ archive.zip exists (178223 bytes, sha256 31881b2d23d027e6...) — skipping
+
+Status: archive=exists_skipped source=skipped_exists

public_pr103_intake_20260505_auto/archive_provenance.json

@@ -0,0 +1,8 @@
+{
+  "candidates_tried": [],
+  "attempts": [],
+  "archive_sha256": "31881b2d23d027e6619f2d8df2fe35d4d207d08882ec673d6c1b7ff119f18c30",
+  "archive_size_bytes": 178223,
+  "status": "exists_skipped",
+  "fetched_at_utc": "2026-05-05T22:44:44.008072+00:00"
+}

public_pr103_intake_20260505_auto/pr_body.md

@@ -0,0 +1,29 @@
+# submission name:
+hnerv_lc_ac
+
+# upload zipped `archive.zip`
+[archive.zip](https://github.com/user-attachments/files/27355831/archive.zip)
+
+# report.txt
+```
+=== Evaluation results over 600 samples ===
+  Average PoseNet Distortion: 0.00003443
+  Average SegNet Distortion: 0.00057638
+  Submission file size: 178,223 bytes
+  Original uncompressed size: 37,545,489 bytes
+  Compression Rate: 0.00474686
+  Final score: 100*segnet_dist + √(10*posenet_dist) + 25*rate = 0.19
+```
+
+# does your submission require gpu for evaluation (inflation)?
+no
+
+# did you include the compression script? and want it to be merged?
+no
+
+# additional comments
+This submission is a derivative of #95, so credit goes to @AaronLeslie138. Added a few tricks to push down the score to 0.19487 locally
+
+Lossless byte-level repack of @BradyMeighan's `hnerv_lc_v2` (#100). Decoder weights, latents, and latent-correction sidecar are all his (built on @EthanYangTW's #98).
+
+The substantive change: arithmetic coding (constriction range coder) on the 8 largest weight tensors and the latent-hi byte stream. Brady's pipeline uses pure brotli; switching the densest payloads to AC with q8 (uint8) histograms beats brotli's symbol-level entropy by ~290 B. Rest of the bytes come from hardcoded section lengths inside `inflate.py` (no length prefixes inside the archive), adaptive `lgwin` search in brotli, a single-byte filename inside the zip, and merging all 9 AC streams into one constriction `RangeEncoder` to eliminate per-stream rounding overhead.

public_pr103_intake_20260505_auto/pr_metadata.json

@@ -0,0 +1,16 @@
+{
+  "additions": 241,
+  "author": "rem2",
+  "changed_files": 2,
+  "closed_at": "2026-05-04T17:51:27Z",
+  "created_at": "2026-05-04T11:55:56Z",
+  "deletions": 0,
+  "head_repo": "rem2/comma_video_compression_challenge",
+  "head_sha": "d20270750e603346056f4e867cddd4744272c9d8",
+  "leaderboard_name": "hnerv_lc_ac",
+  "leaderboard_score": 0.195,
+  "merged_at": null,
+  "pr_number": 103,
+  "state": "closed",
+  "title": "hnerv_lc_ac submission (0.19)"
+}

public_pr106_intake_20260505_auto/INTAKE_LOG.md

@@ -0,0 +1,8 @@
+# PR #106 intake — belt_and_suspenders (score 0.209)
+Started UTC: 2026-05-05T22:44:47.331850+00:00
+
+✓ pr_metadata.json (head_sha=339910eb2da2048796c8555cb964c5d8cb5c194e, author=valtterivalo)
+✓ pr_body.md (1051 chars)
+✓ archive.zip exists (186239 bytes, sha256 3fefbe5dfdd73817...) — skipping
+
+Status: archive=exists_skipped source=skipped_exists

public_pr106_intake_20260505_auto/archive_provenance.json

@@ -0,0 +1,8 @@
+{
+  "candidates_tried": [],
+  "attempts": [],
+  "archive_sha256": "3fefbe5dfdd738179a55ca5c995ff8f63ec2755662d60684706f20d313913f58",
+  "archive_size_bytes": 186239,
+  "status": "exists_skipped",
+  "fetched_at_utc": "2026-05-05T22:44:47.797603+00:00"
+}

public_pr106_intake_20260505_auto/pr_body.md

@@ -0,0 +1,37 @@
+# submission name:
+belt_and_suspenders
+
+# upload zipped `archive.zip`
+[archive.zip](https://github.com/user-attachments/files/27356518/archive.zip)
+
+
+# report.txt
+```
+=== Evaluation config ===
+  batch_size: 16
+  device: cuda
+  num_threads: 2
+  prefetch_queue_depth: 4
+  report: submissions/belt_and_suspenders/report.txt
+  seed: 1234
+  submission_dir: submissions/belt_and_suspenders
+  uncompressed_dir: videos
+  video_names_file: public_test_video_names.txt
+=== Evaluation results over 600 samples ===
+  Average PoseNet Distortion: 0.00003351
+  Average SegNet Distortion: 0.00067142
+  Submission file size: 186,239 bytes
+  Original uncompressed size: 37,545,489 bytes
+  Compression Rate: 0.00496036
+  Final score: 100*segnet_dist + √(10*posenet_dist) + 25*rate = 0.20946
+```
+
+# does your submission require gpu for evaluation (inflation)?
+yes
+
+# did you include the compression script? and want it to be merged?
+yes
+
+# additional comments
+
+HNeRV decoder/architecture lineage: PR #95 (hnerv_muon, AaronLeslie138).

public_pr106_intake_20260505_auto/pr_metadata.json

@@ -0,0 +1,16 @@
+{
+  "additions": 1780,
+  "author": "valtterivalo",
+  "changed_files": 21,
+  "closed_at": "2026-05-04T20:09:54Z",
+  "created_at": "2026-05-04T11:57:01Z",
+  "deletions": 0,
+  "head_repo": "valtterivalo/comma_video_compression_challenge-1",
+  "head_sha": "339910eb2da2048796c8555cb964c5d8cb5c194e",
+  "leaderboard_name": "belt_and_suspenders",
+  "leaderboard_score": 0.209,
+  "merged_at": "2026-05-04T20:09:54Z",
+  "pr_number": 106,
+  "state": "closed",
+  "title": "belt_and_suspenders (0.20946)"
+}

public_pr107_intake_20260505_auto/INTAKE_LOG.md

@@ -0,0 +1,8 @@
+# PR #107 intake — apogee (score 0.229)
+Started UTC: 2026-05-05T22:44:48.501180+00:00
+
+✓ pr_metadata.json (head_sha=20c411b58c9270dfb6ce19742e657fefd1f54829, author=adpena)
+✓ pr_body.md (7358 chars)
+✓ archive.zip exists (178392 bytes, sha256 7ecb0df1c4627d55...) — skipping
+
+Status: archive=exists_skipped source=skipped_exists

public_pr107_intake_20260505_auto/archive_provenance.json

@@ -0,0 +1,8 @@
+{
+  "candidates_tried": [],
+  "attempts": [],
+  "archive_sha256": "7ecb0df1c4627d55d88e03eff3d890b7a7a5b047c62515acff20232cf29310eb",
+  "archive_size_bytes": 178392,
+  "status": "exists_skipped",
+  "fetched_at_utc": "2026-05-05T22:44:48.948849+00:00"
+}

public_pr107_intake_20260505_auto/pr_body.md

@@ -0,0 +1,153 @@
+# submission name:
+apogee
+
+# upload zipped `archive.zip`
+https://github.com/adpena/comma_video_compression_challenge/releases/download/apogee-pr98-hnerv-adapter-20260504/archive.zip
+
+# report.txt
+```
+=== Evaluation config ===
+  batch_size: 16
+  device: cuda
+  num_threads: 2
+  prefetch_queue_depth: 4
+  seed: 1234
+  submission_dir: submissions/apogee
+=== Evaluation results over 600 samples ===
+  Average PoseNet Distortion: 0.00017394
+  Average SegNet Distortion: 0.00068841
+  Submission file size: 178,392 bytes
+  Original uncompressed size: 37,545,489 bytes
+  Compression Rate: 0.00475136
+  Final score: 100*segnet_dist + sqrt(10*posenet_dist) + 25*rate = 0.23
+
+Exact local CUDA/T4 custody:
+  score_recomputed_from_components: 0.22933111465960354
+  archive_sha256: 7ecb0df1c4627d55d88e03eff3d890b7a7a5b047c62515acff20232cf29310eb
+  archive_size_bytes: 178392
+  archive_member: 0.bin
+  archive_member_sha256: fce200db2fe087cc6a051945b3fda2c37f5bbb3e19b8f20a1aea7201db0c9f5f
+  n_samples: 600
+  eval_hardware: Tesla T4
+  runtime_tree_sha256: 0232154c17410621325ec1647e0f0723b3310d63b0d4bc4bf7bbb5e9aa2fccd0
+```
+
+# does your submission require gpu for evaluation (inflation)?
+yes
+
+# did you include the compression script? and want it to be merged?
+no
+
+# additional comments
+Apogee is a contest-faithful HNeRV submission with the exact archive bytes
+hosted as a release asset and the deterministic inflate runtime included in
+this PR.
+
+Public artifacts:
+- PR branch: https://github.com/adpena/comma_video_compression_challenge/tree/apogee-pr98-hnerv-adapter/submissions/apogee
+- Release asset: https://github.com/adpena/comma_video_compression_challenge/releases/tag/apogee-pr98-hnerv-adapter-20260504
+- Runtime entrypoint: `submissions/apogee/inflate.sh`
+- Archive manifest: `submissions/apogee/archive_manifest.json`
+
+The submission was validated locally through the canonical path:
+
+```
+archive.zip -> inflate.sh -> upstream/evaluate.py
+```
+
+We also ran a strict pre-submission compliance gate before opening this PR.
+That gate checked archive SHA/size, single-member ZIP integrity,
+local-header/central-directory consistency, runtime tree custody, report
+linkage, public hygiene, T4-equivalent CUDA auth eval, and exact
+component-score recomputation.
+
+Design summary:
+- HNeRV-style learned decoder and per-pair latent stream optimized for the
+  fixed contest video and task scorers.
+- Deterministic runtime wrapper with a closed one-video contract for `0.mkv`.
+- All score-affecting bytes are charged inside `archive.zip`; runtime source
+  is only the decoder/inflate program.
+- Score claims are tied to exact archive SHA-256 and runtime-tree SHA-256 rather than human logs.
+
+Methodology and meta-analysis:
+- The challenge objective was treated as a rate-distortion compiler problem:
+  representation, decoder, quantization, side information, and packing are
+  optimized against the exact SegNet/PoseNet/rate formula.
+- We tracked both positive and negative evidence. Failed archive variants,
+  wrapper bugs, CPU/MPS/proxy disagreements, and byte-only wins that harmed
+  PoseNet/SegNet were used as guardrails rather than score claims.
+- Public frontier submissions showed that the winning basin is not a generic
+  perceptual video codec; it is a compact machine-score representation for one
+  fixed video and two fixed neural evaluators. Apogee follows that principle
+  while keeping all score-affecting state inside the charged archive.
+- The archive is intentionally simple at the outer layer: one ZIP member,
+  deterministic byte identity, no hidden sidecars, no scorer patches, and a
+  runtime that fails closed on the expected one-video contest contract.
+
+Public source for this submitted runtime is in this PR. The archive asset is immutable at the release URL above.
+
+
+## Production and OSS deployment notes
+
+This PR intentionally keeps the contest runtime small and auditable rather than
+shipping the full research workspace. The production-relevant properties are:
+
+- deterministic single-payload archive: one `0.bin` member, fixed SHA-256, no
+  hidden sidecars or local paths;
+- explicit runtime custody: the submitted runtime tree is small enough to audit
+  file-by-file and was hashed during exact evaluation;
+- deployable inference shape: `inflate.sh` is the only entrypoint and delegates
+  to a compact Python/PyTorch decoder;
+- hardware fit: local exact CUDA/T4 auth eval completed in about one minute,
+  with inflate itself taking about 22.6 seconds, well inside the 30-minute
+  contest limit;
+- reproducible validation: score claims are tied to exact archive SHA, runtime
+  tree SHA, component distances, sample count, and T4 hardware evidence;
+- clean separation of concerns: the research system (`tac`) produced and
+  validated the archive, while this PR exposes only the minimal runtime needed
+  by comma's evaluator.
+
+A productionized version inside comma/openpilot should preserve this boundary:
+keep the trained representation and entropy-coded payload as charged data,
+keep the inflate path deterministic and side-effect-light, and move reusable
+codec components into a small package with typed payload contracts, CI replay
+fixtures, and byte-for-byte archive tests. The broader research tooling can be
+released separately after sanitizing provider state, local custody ledgers, and
+operator-specific paths.
+
+---
+
+## OSS extraction status (updated 2026-05-05)
+
+Two repos hold the work behind this submission:
+
+**`tac`** — *Task-Aware Compression* — the tight production library:
+https://github.com/adpena/tac (MIT, 214 source files, 415 tests, CI-passing
+on Python 3.11/3.12). This is the reusable codec primitives, predictor with
+refusal modes, distortion proxy, parallel-dispatch toolchain, and hardened
+preflight infrastructure used to produce and validate this submission.
+Importable via `pip install -e .` from a clone.
+
+**`comma-lab`** — *the comma video compression contest research workspace* —
+the comprehensive research record:
+https://github.com/adpena/comma-lab (private pending sanitization). This is
+the parent directory containing the full experimental ledger, dispatch
+wrappers, lane registry, byte-level deconstruction of all 54 scored public
+PRs, methodology writeup (`docs/paper/01_introduction.md` through
+`docs/paper/07_discussion.md`), and ~585 session-memory files capturing the
+research trajectory, findings, failures, and bug-class extinctions. The
+sanitization pass strips ~30GB of regenerable cache files (uncompressed
+`0.raw` ground-truth frame dumps, `gt_frames.pt`/`comp_frames.pt` training
+caches, `renderer_output_for_postfilter.pt`, decoded mask tensors, the
+`raft_flow.pt` optical flow cache) — all re-derivable from the source
+archive, none containing unique research signal. Public release expected
+within days.
+
+**Honest accounting:** this OSS extraction work was deprioritized during the
+all-night sub-deadline scramble for leaderboard position, and is now being
+done post-deadline. Future PRs to this challenge (if comma runs adjacent
+contests) will demonstrate the closed-loop parallel-dispatch toolchain that
+was missing during this contest's race window — see `docs/paper/07_discussion.md`
+§7.8 in `comma-lab` for the postmortem of how the May 4 race window was
+decided in 4h 8min after PR #95 (HNeRV root) published, with the silver
+medal landing in 241 lines / 2 files / under 4 hours.

public_pr107_intake_20260505_auto/pr_metadata.json

@@ -0,0 +1,16 @@
+{
+  "additions": 1665,
+  "author": "adpena",
+  "changed_files": 20,
+  "closed_at": "2026-05-04T17:51:52Z",
+  "created_at": "2026-05-04T12:10:42Z",
+  "deletions": 0,
+  "head_repo": "adpena/comma_video_compression_challenge",
+  "head_sha": "20c411b58c9270dfb6ce19742e657fefd1f54829",
+  "leaderboard_name": "apogee",
+  "leaderboard_score": 0.229,
+  "merged_at": null,
+  "pr_number": 107,
+  "state": "closed",
+  "title": "apogee submission (0.2293)"
+}

public_pr21_intake_20260505_auto/INTAKE_LOG.md

@@ -0,0 +1,8 @@
+# PR #21 intake — h265_g16_512x384_veryslow (score 2.553)
+Started UTC: 2026-05-05T22:45:08.467629+00:00
+
+✓ pr_metadata.json (head_sha=8edf47c18f5a35dd0553b2bce2e257f54f655a97, author=royjr)
+✓ pr_body.md (1311 chars)
+✓ archive.zip exists (1238597 bytes, sha256 9c563f054c59eb5d...) — skipping
+
+Status: archive=exists_skipped source=skipped_exists

public_pr21_intake_20260505_auto/archive_provenance.json

@@ -0,0 +1,8 @@
+{
+  "candidates_tried": [],
+  "attempts": [],
+  "archive_sha256": "9c563f054c59eb5de539a1c685c06e94894622ab87ba281a49b53a9cec910848",
+  "archive_size_bytes": 1238597,
+  "status": "exists_skipped",
+  "fetched_at_utc": "2026-05-05T22:45:08.845457+00:00"
+}

public_pr21_intake_20260505_auto/pr_body.md

@@ -0,0 +1,35 @@
+# submission name:
+h265_g16_512x384_veryslow
+
+# upload zipped `archive.zip`
+https://github.com/royjr/comma_video_compression_challenge/releases/download/h265_g16_512x384_veryslow/archive.zip
+
+# report.txt
+```
+=== Evaluation config ===
+  batch_size: 16
+  device: mps
+  num_threads: 2
+  prefetch_queue_depth: 4
+  report: submissions/my_submission_h265_g16_512x384_veryslow/report.txt
+  seed: 1234
+  submission_dir: submissions/my_submission_h265_g16_512x384_veryslow
+  uncompressed_dir: comma_video_compression_challenge/videos
+  video_names_file: comma_video_compression_challenge/public_test_video_names.txt
+=== Evaluation results over 600 samples ===
+  Average PoseNet Distortion: 0.12133539
+  Average SegNet Distortion: 0.00626843
+  Submission file size: 1,238,597 bytes
+  Original uncompressed size: 37,545,489 bytes
+  Compression Rate: 0.03298924
+  Final score: 100*segnet_dist + √(10*posenet_dist) + 25*rate = 2.55
+```
+
+# does your submission require gpu for evaluation (inflation)?
+no 
+
+# did you include the compression script? and want it to be merged?
+yes
+
+# additional comments
+H.265 (libx265) with veryslow preset, 512×384 lanczos downscale, CRF 27, GOP 16, bf=1. Decoder-side bicubic upscale with 5×5 avg-pool unsharp mask (alpha=3.0) to restore edge structure.  

public_pr21_intake_20260505_auto/pr_metadata.json

@@ -0,0 +1,16 @@
+{
+  "additions": 75,
+  "author": "royjr",
+  "changed_files": 4,
+  "closed_at": "2026-04-04T23:57:37Z",
+  "created_at": "2026-04-04T19:33:02Z",
+  "deletions": 0,
+  "head_repo": "royjr/comma_video_compression_challenge",
+  "head_sha": "8edf47c18f5a35dd0553b2bce2e257f54f655a97",
+  "leaderboard_name": "h265_g16_512x384_veryslow",
+  "leaderboard_score": 2.553,
+  "merged_at": "2026-04-04T23:57:37Z",
+  "pr_number": 21,
+  "state": "closed",
+  "title": "h265_g16_512x384_veryslow (score: 2.55)"
+}

public_pr24_intake_20260505_auto/INTAKE_LOG.md

@@ -0,0 +1,10 @@
+# PR #24 intake — svtav1_cheetah (score 2.052)
+Started UTC: 2026-05-05T22:45:05.314192+00:00
+
+✓ pr_metadata.json (head_sha=cbc0ac3970fd9d3a3baadc3597f858008e06c88f, author=ylevental)
+✓ pr_body.md (1107 chars)
+  body URL: https://github.com/ylevental/comma_video_compression_challenge/releases/download/svtav1-cheetah-v1/archive.zip
+✗ all 1 candidates failed
+    https://github.com/ylevental/comma_video_compression_challenge/releases/download/svtav1-cheetah-v1/archive.zip: HTTPError: HTTP Error 404: Not Found
+
+Status: archive=no_candidate_succeeded source=skipped_no_head

public_pr24_intake_20260505_auto/archive_provenance.json

@@ -0,0 +1,16 @@
+{
+  "candidates_tried": [
+    "https://github.com/ylevental/comma_video_compression_challenge/releases/download/svtav1-cheetah-v1/archive.zip"
+  ],
+  "attempts": [
+    {
+      "url": "https://github.com/ylevental/comma_video_compression_challenge/releases/download/svtav1-cheetah-v1/archive.zip",
+      "success": false,
+      "result": "HTTPError: HTTP Error 404: Not Found"
+    }
+  ],
+  "archive_sha256": null,
+  "archive_size_bytes": null,
+  "status": "no_candidate_succeeded",
+  "fetched_at_utc": "2026-05-05T22:45:05.849123+00:00"
+}

public_pr24_intake_20260505_auto/pr_body.md

@@ -0,0 +1,25 @@
+# submission name:
+svtav1_cheetah
+
+# upload zipped `archive.zip`
+https://github.com/ylevental/comma_video_compression_challenge/releases/download/svtav1-cheetah-v1/archive.zip
+
+# report.txt
+=== Evaluation results over 600 samples ===
+  Average PoseNet Distortion: 0.08254363
+  Average SegNet Distortion: 0.00513639
+  Submission file size: 945,882 bytes
+  Original uncompressed size: 37,545,489 bytes
+  Compression Rate: 0.02519296
+  Final score: 100*segnet_dist + √(10*posenet_dist) + 25*rate = 2.05
+
+# does your submission require gpu for evaluation (inflation)?
+no
+
+# did you include the compression script? and want it to be merged?
+yes
+
+# additional comments
+After last time, I won't be looking for an interview :p.  But I still have claim on the lowest score as of this writing.  I used Claude Opus 4.6 Extended for my AI.
+
+SVT-AV1 preset 0, CRF 33, 45% lanczos downscale (528×392), film-grain=22, keyint=180, scd=0. Decoder-side bicubic upscale to 1164×874 with 9×9 Gaussian unsharp mask (strength 0.85). Same approach as svtav1_45pct_unsharp but evaluated with SVT-AV1 v2.3.0.

public_pr24_intake_20260505_auto/pr_metadata.json

@@ -0,0 +1,16 @@
+{
+  "additions": 127,
+  "author": "ylevental",
+  "changed_files": 3,
+  "closed_at": "2026-04-04T23:59:01Z",
+  "created_at": "2026-04-04T22:05:20Z",
+  "deletions": 0,
+  "head_repo": null,
+  "head_sha": "cbc0ac3970fd9d3a3baadc3597f858008e06c88f",
+  "leaderboard_name": "svtav1_cheetah",
+  "leaderboard_score": 2.052,
+  "merged_at": "2026-04-04T23:59:01Z",
+  "pr_number": 24,
+  "state": "closed",
+  "title": "submission: svtav1_cheetah (score 2.05)"
+}

public_pr37_intake_20260505_auto/INTAKE_LOG.md

@@ -0,0 +1,8 @@
+# PR #37 intake — svtav1_spline_fg22 (score 2.158)
+Started UTC: 2026-05-05T22:45:07.619170+00:00
+
+✓ pr_metadata.json (head_sha=2da739e98fdd9b108988e1c4a661f27ffc3575e8, author=IamTemmy)
+✓ pr_body.md (1172 chars)
+✓ archive.zip exists (930191 bytes, sha256 197eebc78a5bbff3...) — skipping
+
+Status: archive=exists_skipped source=skipped_exists

public_pr37_intake_20260505_auto/archive_provenance.json

@@ -0,0 +1,8 @@
+{
+  "candidates_tried": [],
+  "attempts": [],
+  "archive_sha256": "197eebc78a5bbff368f30d528afce4abb26d681e59723a0839ee7f51750ec898",
+  "archive_size_bytes": 930191,
+  "status": "exists_skipped",
+  "fetched_at_utc": "2026-05-05T22:45:08.068655+00:00"
+}

public_pr37_intake_20260505_auto/pr_body.md

@@ -0,0 +1,31 @@
+# submission name:
+
+svtav1_spline_fg22
+
+# upload zipped `archive.zip`
+
+https://github.com/IamTemmy/comma_video_compression_challenge/releases/download/v0.1/archive.zip
+
+# report.txt
+
+=== Evaluation results over 600 samples ===
+  Average PoseNet Distortion: 0.08102395
+  Average SegNet Distortion: 0.00638626
+  Submission file size: 930,191 bytes
+  Original uncompressed size: 37,545,489 bytes
+  Compression Rate: 0.02477504
+  Final score: 100*segnet_dist + √(10*posenet_dist) + 25*rate = 2.16
+
+# does your submission require gpu for evaluation (inflation)?
+
+no
+
+# did you include the compression script? and want it to be merged?
+
+yes
+
+# additional comments
+
+SVT-AV1 with 45% spline downscale, CRF 32, preset 0, GOP 180, film-grain=22 (denoise+synth).
+
+The key insight: ffmpeg's `spline` downscale algorithm preserves PoseNet-relevant temporal features better than `lanczos` at this scale factor. Combined with film-grain synthesis (which strips sensor noise before encoding for cleaner inter-frame consistency that PoseNet rewards), this achieves PoseNet distortion of 0.081 — meaningfully lower than equivalent lanczos-based configurations.

public_pr37_intake_20260505_auto/pr_metadata.json

@@ -0,0 +1,16 @@
+{
+  "additions": 109,
+  "author": "IamTemmy",
+  "changed_files": 4,
+  "closed_at": "2026-04-08T03:10:44Z",
+  "created_at": "2026-04-07T10:04:43Z",
+  "deletions": 0,
+  "head_repo": "IamTemmy/comma_video_compression_challenge",
+  "head_sha": "2da739e98fdd9b108988e1c4a661f27ffc3575e8",
+  "leaderboard_name": "svtav1_spline_fg22",
+  "leaderboard_score": 2.158,
+  "merged_at": "2026-04-08T03:10:44Z",
+  "pr_number": 37,
+  "state": "closed",
+  "title": "svtav1_spline_fg22 submission (score: 2.16)"
+}

public_pr39_intake_20260505_auto/INTAKE_LOG.md

@@ -0,0 +1,8 @@
+# PR #39 intake — baseline_fast (score 4.39)
+Started UTC: 2026-05-05T22:45:09.723627+00:00
+
+✓ pr_metadata.json (head_sha=aa67a3aa6d0b8223720e163e03302e7a00d6fdfd, author=dwallener)
+✓ pr_body.md (3486 chars)
+✓ archive.zip exists (1025938 bytes, sha256 30b26c0a2dc2fafb...) — skipping
+
+Status: archive=exists_skipped source=skipped_exists

public_pr39_intake_20260505_auto/archive_provenance.json

@@ -0,0 +1,8 @@
+{
+  "candidates_tried": [],
+  "attempts": [],
+  "archive_sha256": "30b26c0a2dc2fafb171ccbb52b5651d65979c58a2ef27e44b097c30cf166835a",
+  "archive_size_bytes": 1025938,
+  "status": "exists_skipped",
+  "fetched_at_utc": "2026-05-05T22:45:10.148776+00:00"
+}

public_pr39_intake_20260505_auto/pr_body.md

@@ -0,0 +1,71 @@
+# submission name:
+<!-- the directory name of your submission -->
+<!-- please make sure it matches exactly -->
+damir_bearclaw_003
+
+# upload zipped `archive.zip`
+<!-- do not check it in the code (it's already ignored in .gitignore) -->
+<!-- you can use the upload file feature (drag and drop), make sure curl -L works -->
+https://github.com/user-attachments/files/26582273/archive.zip
+
+# report.txt
+<!-- do not check it in the code (it's already ignored in .gitignore) -->
+<!-- copy the report.txt content here -->
+```
+=== Evaluation config ===
+  batch_size: 16
+  device: cpu
+  num_threads: 2
+  prefetch_queue_depth: 4
+  report: submissions/damir_bearclaw_003/report.txt
+  seed: 1234
+  submission_dir: submissions/damir_bearclaw_003
+  uncompressed_dir: /home/runner/work/comma_video_compression_challenge/comma_video_compression_challenge/videos
+  video_names_file: /home/runner/work/comma_video_compression_challenge/comma_video_compression_challenge/public_test_video_names.txt
+=== Evaluation results over 600 samples ===
+  Average PoseNet Distortion: 1.18874884
+  Average SegNet Distortion: 0.00954937
+  Submission file size: 1,025,938 bytes
+  Original uncompressed size: 37,545,489 bytes
+  Compression Rate: 0.02732520
+  Final score: 100*segnet_dist + √(10*posenet_dist) + 25*rate = 5.09
+```
+
+# does your submission require gpu for evaluation (inflation)?
+no
+<!-- this only applies to inflation, you can use gpu for compression but not require gpu for evaluation -->
+
+
+## Summary
+
+This PR adds `damir_bearclaw_003`, a single-stream representation-oriented submission.
+
+Instead of preserving the whole frame uniformly, it:
+
+- stores only the middle 50% of the frame
+- applies boundary-preserving smoothing to that middle band
+- reconstructs synthetic top and bottom filler bands during inflate
+
+This does not beat our previous best leaderboard score. That is intentional. The goal here is to probe the point where proxy-fidelity improvements become less practically meaningful than size reduction.
+
+## Why this direction
+
+Our later experiments suggested that the fixed scoring models are rewarding increasingly fine proxy exactness, even when the remaining differences appear small in practical terms.
+
+For SegNet, most semantic burden sits in the central driving corridor, while the upper and lower filler regions are strongly position-biased:
+
+![SegNet class map](https://raw.githubusercontent.com/dwallener/comma_video_compression_challenge/codex/damir_bearclaw_003/submissions/damir_bearclaw_003/images/segnet_classes_frame_0171_mask.png)
+
+For PoseNet, occlusion analysis suggested strong sensitivity to the central road corridor and near-field anchor regions, but also diminishing practical value from chasing ever smaller scalar deviations:
+
+![PoseNet occlusion example](https://raw.githubusercontent.com/dwallener/comma_video_compression_challenge/codex/damir_bearclaw_003/submissions/damir_bearclaw_003/images/pose0_occlusion_rank_04_pair_0252.png)
+
+## Takeaway
+
+This submission is best read as a limit probe:
+
+- preserve the task-relevant middle-region structure
+- accept small proxy-error increases once they are practically tolerable
+- then push aggressively on stored size
+
+More broadly, this suggests that once PoseNet and SegNet are within an acceptable tolerance band, the rubric should increasingly emphasize compression rate rather than continuing to reward uniform proxy-fidelity gains equally.

public_pr39_intake_20260505_auto/pr_metadata.json

@@ -0,0 +1,16 @@
+{
+  "additions": 263,
+  "author": "dwallener",
+  "changed_files": 8,
+  "closed_at": "2026-04-09T02:40:02Z",
+  "created_at": "2026-04-07T15:55:58Z",
+  "deletions": 0,
+  "head_repo": "dwallener/comma_video_compression_challenge",
+  "head_sha": "aa67a3aa6d0b8223720e163e03302e7a00d6fdfd",
+  "leaderboard_name": "baseline_fast",
+  "leaderboard_score": 4.39,
+  "merged_at": "2026-04-09T02:40:02Z",
+  "pr_number": 39,
+  "state": "closed",
+  "title": "Add middle-band single-stream submission (damir_bearclaw_003)"
+}

public_pr39_intake_20260505_auto/source/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2026 comma.ai
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

public_pr39_intake_20260505_auto/source/evaluate.sh

@@ -0,0 +1,76 @@
+#!/usr/bin/env bash
+set -euo pipefail
+
+HERE="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+
+SUBMISSION_DIR="${HERE}/submissions/baseline"
+VIDEO_NAMES_FILE="${HERE}/public_test_video_names.txt"
+DEVICE="cpu"
+
+while [[ $# -gt 0 ]]; do
+  case "$1" in
+    --submission-dir|--submission_dir)
+      SUBMISSION_DIR="${2%/}"; shift 2 ;;
+    --video-names-file|--video_names_file)
+      VIDEO_NAMES_FILE="$2"; shift 2 ;;
+    --device)
+      DEVICE="$2"; shift 2 ;;
+    *)
+      echo "Unknown arg: $1" >&2
+      echo "Usage: $0 [--submission-dir <dir>] [--video-names-file <file>] [--device <cpu|cuda|mps>]" >&2
+      exit 2 ;;
+  esac
+done
+
+ARCHIVE_ZIP="${SUBMISSION_DIR}/archive.zip"
+ARCHIVE_DIR="${SUBMISSION_DIR}/archive"
+INFLATED_DIR="${SUBMISSION_DIR}/inflated"
+
+INFLATE_SH="${SUBMISSION_DIR}/inflate.sh"
+
+if [ ! -f "$ARCHIVE_ZIP" ]; then
+  echo "ERROR: ${ARCHIVE_ZIP} not found" >&2
+  exit 1
+fi
+
+if [ ! -f "$INFLATE_SH" ]; then
+  echo "ERROR: ${INFLATE_SH} not found" >&2
+  exit 1
+fi
+
+# unzip
+rm -rf "$ARCHIVE_DIR"
+mkdir -p "$ARCHIVE_DIR"
+unzip -o "$ARCHIVE_ZIP" -d "$ARCHIVE_DIR"
+
+# inflate
+bash "${SUBMISSION_DIR}/inflate.sh" "$ARCHIVE_DIR" "$INFLATED_DIR" "$VIDEO_NAMES_FILE"
+
+# assert all videos have been inflated
+MISSING=0
+while IFS= read -r line; do
+  [ -z "$line" ] && continue
+  BASE="${line%.*}"
+  RAW_PATH="${INFLATED_DIR}/${BASE}.raw"
+  if [ ! -f "$RAW_PATH" ]; then
+    echo "ERROR: missing inflated file: ${RAW_PATH}" >&2
+    MISSING=$((MISSING + 1))
+  fi
+done < "$VIDEO_NAMES_FILE"
+
+if [ "$MISSING" -gt 0 ]; then
+  echo "ERROR: ${MISSING} video(s) not inflated" >&2
+  exit 1
+fi
+
+echo "All videos inflated to ${INFLATED_DIR}"
+
+# evaluate
+python "$HERE/evaluate.py" \
+  --submission-dir "$SUBMISSION_DIR" \
+  --uncompressed-dir "$HERE/videos" \
+  --report "$SUBMISSION_DIR/report.txt" \
+  --video-names-file "$VIDEO_NAMES_FILE" \
+  --device "$DEVICE"
+
+echo "Evaluation complete. Report saved to ${SUBMISSION_DIR}/report.txt"

public_pr39_intake_20260505_auto/source/public_test_video_names.txt

@@ -0,0 +1 @@
+0.mkv

public_pr39_intake_20260505_auto/source/pyproject.toml

@@ -0,0 +1,81 @@
+[project]
+name = "comma-video-compression-challenge"
+version = "0.1.0"
+requires-python = "~=3.11"
+dependencies = [
+  "numpy",
+  "einops",
+  "timm",
+  "safetensors",
+  "segmentation-models-pytorch",
+  "tqdm",
+  "pillow",
+  "av",
+]
+
+[dependency-groups]
+cpu = ["torch", "torchvision"]
+cu126 = ["torch", "torchvision", "nvidia-dali-cuda120"]
+cu128 = ["torch", "torchvision", "nvidia-dali-cuda120"]
+cu130 = ["torch", "torchvision", "nvidia-dali-cuda130"]
+mps = ["torch", "torchvision"]
+
+[tool.uv]
+conflicts = [
+  [{ group = "cpu" }, { group = "cu126" }],
+  [{ group = "cpu" }, { group = "cu128" }],
+  [{ group = "cpu" }, { group = "cu130" }],
+  [{ group = "cpu" }, { group = "mps" }],
+  [{ group = "cu126" }, { group = "cu128" }],
+  [{ group = "cu126" }, { group = "cu130" }],
+  [{ group = "cu126" }, { group = "mps" }],
+  [{ group = "cu128" }, { group = "cu130" }],
+  [{ group = "cu128" }, { group = "mps" }],
+  [{ group = "cu130" }, { group = "mps" }],
+]
+
+[tool.uv.sources]
+torch = [
+  { index = "pytorch-cpu", group = "cpu" },
+  { index = "pytorch-cu126", group = "cu126" },
+  { index = "pytorch-cu128", group = "cu128" },
+  { index = "pytorch-cu130", group = "cu130" },
+]
+torchvision = [
+  { index = "pytorch-cpu", group = "cpu" },
+  { index = "pytorch-cu126", group = "cu126" },
+  { index = "pytorch-cu128", group = "cu128" },
+  { index = "pytorch-cu130", group = "cu130" },
+]
+nvidia-dali-cuda120 = [
+  { index = "nvidia", group = "cu126" },
+  { index = "nvidia", group = "cu128" },
+]
+nvidia-dali-cuda130 = [
+  { index = "nvidia", group = "cu130" },
+]
+
+[[tool.uv.index]]
+name = "pytorch-cpu"
+url = "https://download.pytorch.org/whl/cpu/"
+explicit = true
+
+[[tool.uv.index]]
+name = "pytorch-cu126"
+url = "https://download.pytorch.org/whl/cu126/"
+explicit = true
+
+[[tool.uv.index]]
+name = "pytorch-cu128"
+url = "https://download.pytorch.org/whl/cu128/"
+explicit = true
+
+[[tool.uv.index]]
+name = "pytorch-cu130"
+url = "https://download.pytorch.org/whl/cu130/"
+explicit = true
+
+[[tool.uv.index]]
+name = "nvidia"
+url = "https://pypi.nvidia.com"
+explicit = true

public_pr44_intake_20260505_auto/INTAKE_LOG.md

@@ -0,0 +1,8 @@
+# PR #44 intake — v4_qp_aq2_roi (score 2.02)
+Started UTC: 2026-05-05T22:45:04.359182+00:00
+
+✓ pr_metadata.json (head_sha=6c932b27a0a9760fc6480542e79a3b4f58c1e9e6, author=mil1200)
+✓ pr_body.md (1808 chars)
+✓ archive.zip exists (836886 bytes, sha256 f4fd7bc547f75a00...) — skipping
+
+Status: archive=exists_skipped source=skipped_exists

public_pr44_intake_20260505_auto/archive_provenance.json

@@ -0,0 +1,8 @@
+{
+  "candidates_tried": [],
+  "attempts": [],
+  "archive_sha256": "f4fd7bc547f75a006b8c58e15aaa8752559cba57ec05d0df8bfe9b334c9e14a2",
+  "archive_size_bytes": 836886,
+  "status": "exists_skipped",
+  "fetched_at_utc": "2026-05-05T22:45:04.835637+00:00"
+}

public_pr44_intake_20260505_auto/pr_body.md

@@ -0,0 +1,31 @@
+# submission name:
+v4_qp_aq2_roi
+
+# upload zipped archive.zip
+https://github.com/mil1200/comma_video_compression_challenge/releases/download/v4_qp_aq2_roi/archive.zip
+
+# report.txt
+=== Evaluation results over 600 samples ===
+  Average PoseNet Distortion: 0.08013804
+  Average SegNet Distortion: 0.00567680
+  Submission file size: 836,886 bytes
+  Original uncompressed size: 37,545,489 bytes
+  Compression Rate: 0.02228992
+  Final score: 100*segnet_dist + √(10*posenet_dist) + 25*rate = 2.02
+
+# does your submission require gpu for evaluation (inflation)?
+no
+
+# did you include the compression script? and want it to be merged?
+yes
+
+# additional comments
+This submission uses SegNet-guided per-block quality allocation via SVT-AV1's ROI map feature during encoding. The key idea: not all 64x64 blocks in the encoded frame are equally important for the downstream neural networks.
+
+During compression, SegNet is run on the original video to produce class predictions. For each 64x64 encoder block, the script checks whether it contains class boundaries (multiple semantic classes) or uniform regions. Blocks at class boundaries, especially near road edges, get a QP offset of -5 (more bits), while pure sky blocks get +5 (fewer bits). These offsets work alongside SVT-AV1's built-in adaptive quantization (aq-mode 2) rather than replacing it.
+
+This is combined with spatial ROI preprocessing (YUV-domain luma denoise + chroma collapse outside the driving corridor) and standard SVT-AV1 encoding at CRF 34, film-grain 22.
+
+The encoder used is upstream SVT-AV1 v4.0.1 built from source (https://gitlab.com/AOMediaCodec/SVT-AV1) with its native --roi-map-file support. The inflate script is a plain bicubic upscale with no neural network or special processing.
+
+Tested on local eval with AVVideoDataset (no DALI).

public_pr44_intake_20260505_auto/pr_metadata.json

@@ -0,0 +1,16 @@
+{
+  "additions": 384,
+  "author": "mil1200",
+  "changed_files": 6,
+  "closed_at": "2026-04-09T02:43:15Z",
+  "created_at": "2026-04-08T05:34:58Z",
+  "deletions": 0,
+  "head_repo": "mil1200/comma_video_compression_challenge",
+  "head_sha": "6c932b27a0a9760fc6480542e79a3b4f58c1e9e6",
+  "leaderboard_name": "v4_qp_aq2_roi",
+  "leaderboard_score": 2.02,
+  "merged_at": "2026-04-09T02:43:15Z",
+  "pr_number": 44,
+  "state": "closed",
+  "title": "v4_qp_aq2_roi submission"
+}

public_pr52_intake_20260505_auto/INTAKE_LOG.md

@@ -0,0 +1,8 @@
+# PR #52 intake — av1_crf31_bicubic (score 2.033)
+Started UTC: 2026-05-05T22:45:04.836039+00:00
+
+✓ pr_metadata.json (head_sha=ffbdf2647a05d241324a99cac1e2237b92d8cddc, author=m2kanodi)
+✓ pr_body.md (2661 chars)
+✓ archive.zip exists (1002966 bytes, sha256 5d65e258bee994ee...) — skipping
+
+Status: archive=exists_skipped source=skipped_exists

public_pr52_intake_20260505_auto/archive_provenance.json

@@ -0,0 +1,8 @@
+{
+  "candidates_tried": [],
+  "attempts": [],
+  "archive_sha256": "5d65e258bee994ee4304f103d690b2c6b12413a5aa84bd6c85d8a30862e21674",
+  "archive_size_bytes": 1002966,
+  "status": "exists_skipped",
+  "fetched_at_utc": "2026-05-05T22:45:05.313748+00:00"
+}

public_pr52_intake_20260505_auto/pr_body.md

@@ -0,0 +1,58 @@
+# submission name:
+av1_crf31_bicubic
+
+# upload zipped `archive.zip`
+[archive.zip](https://github.com/m2kanodi/comma_video_compression_challenge/releases/download/av1_crf31_bicubic/archive.zip)
+
+# report.txt
+```
+=== Evaluation results over 600 samples ===
+  Average PoseNet Distortion: 0.07131552
+  Average SegNet Distortion: 0.00520704
+  Submission file size: 1,002,966 bytes
+  Original uncompressed size: 37,545,489 bytes
+  Compression Rate: 0.02671336
+  Final score: 100*segnet_dist + √(10*posenet_dist) + 25*rate = 2.03
+```
+
+# does your submission require gpu for evaluation (inflation)?
+no
+
+# did you include the compression script? and want it to be merged?
+yes
+
+# additional comments
+SVT-AV1 preset 0 at CRF 31, with corridor-aware preprocessing and a
+sharpening pass at inflate.
+
+Pipeline:
+
+- **Preprocess (corridor ROI denoise).** PoseNet and SegNet both downsample
+  to 512x384 and only care about the lower-center driving corridor.
+  Everything outside (sky, buildings, hood, passing trees) is high-entropy
+  content the encoder would otherwise spend bits on for no scoring benefit.
+  The preprocess step draws a per-segment corridor polygon, hand-tuned on
+  the four 300-frame chunks of 0.mkv since the mount drifts, feathers the
+  mask by 24px, and on the outside region runs a YUV-space luma gaussian
+  blur (strength 2.5) plus medium chroma average-pooling. Inside the
+  corridor the frame is untouched. The output is written as an FFV1
+  lossless intermediate so the encoder only sees the denoised frames.
+- **Encode.** 45% Lanczos downscale, then SVT-AV1 preset 0, CRF 31,
+  film-grain 22, keyint 180, scd=0. The score has a sqrt on the PoseNet
+  term and a linear 25x weight on rate, so spending a few extra bits at
+  CRF 31 (vs a tighter CRF) gives PoseNet headroom that more than pays
+  for the rate hit at this operating point. Going lower than 31 flipped
+  the balance the other way in my sweeps.
+- **Inflate (bicubic + unsharp).** Decode, bicubic upscale back to the
+  original resolution, then a 9-tap binomial unsharp mask at strength 0.27.
+  Bicubic was a small but consistent PoseNet improvement over lanczos on
+  this content; SegNet was unchanged. The unsharp recovers the
+  high-frequency edges SegNet relies on without visible ringing on lane
+  markings.
+- **CPU-only inflation.** No GPU required. The inflate script uses CUDA or
+  MPS if available, otherwise falls back to CPU with identical output.
+
+Final local score: **2.03** (posenet 0.0713, segnet 0.00521, rate 0.0267).
+Evaluated on an M-series Mac (MPS). On the same box baseline_fast measured
+4.43 vs the README's 4.39, so the CI number should land at or just under
+2.03.

public_pr52_intake_20260505_auto/pr_metadata.json

@@ -0,0 +1,16 @@
+{
+  "additions": 254,
+  "author": "m2kanodi",
+  "changed_files": 5,
+  "closed_at": "2026-04-13T03:09:16Z",
+  "created_at": "2026-04-11T05:58:31Z",
+  "deletions": 0,
+  "head_repo": "m2kanodi/comma_video_compression_challenge",
+  "head_sha": "ffbdf2647a05d241324a99cac1e2237b92d8cddc",
+  "leaderboard_name": "av1_crf31_bicubic",
+  "leaderboard_score": 2.033,
+  "merged_at": "2026-04-13T03:09:15Z",
+  "pr_number": 52,
+  "state": "closed",
+  "title": "av1_crf31_bicubic"
+}

public_pr53_intake_20260505_auto/INTAKE_LOG.md

@@ -0,0 +1,8 @@
+# PR #53 intake — mask2mask (score 0.602)
+Started UTC: 2026-05-05T22:44:58.355539+00:00
+
+✓ pr_metadata.json (head_sha=ef78345815e5a31548dea65f30cf2942b36d0ddf, author=Quantizr)
+✓ pr_body.md (1960 chars)
+✓ archive.zip exists (386192 bytes, sha256 08f3f17921cc6ee7...) — skipping
+
+Status: archive=exists_skipped source=skipped_exists

public_pr53_intake_20260505_auto/archive_provenance.json

@@ -0,0 +1,8 @@
+{
+  "candidates_tried": [],
+  "attempts": [],
+  "archive_sha256": "08f3f17921cc6ee7bf1dd3da7293127d7332cce4f9686beb2371cfaa428438df",
+  "archive_size_bytes": 386192,
+  "status": "exists_skipped",
+  "fetched_at_utc": "2026-05-05T22:44:58.828424+00:00"
+}

public_pr53_intake_20260505_auto/pr_body.md

@@ -0,0 +1,38 @@
+# submission name:
+mask2mask
+
+# upload zipped  `archive.zip`
+[archive.zip](https://github.com/user-attachments/files/26644494/archive.zip)
+
+# report.txt
+```
+=== Evaluation config ===
+  batch_size: 16
+  device: cuda
+  num_threads: 2
+  prefetch_queue_depth: 4
+  report: submissions/mask2mask/report.txt
+  seed: 1234
+  submission_dir: submissions/mask2mask
+  uncompressed_dir: /mnt/d/comma_submission/comma_video_compression_challenge/videos
+  video_names_file: /mnt/d/comma_submission/comma_video_compression_challenge/public_test_video_names.txt
+=== Evaluation results over 600 samples ===
+  Average PoseNet Distortion: 0.00065614
+  Average SegNet Distortion: 0.00264182
+  Submission file size: 386,192 bytes
+  Original uncompressed size: 37,545,489 bytes
+  Compression Rate: 0.01028598
+  Final score: 100*segnet_dist + √(10*posenet_dist) + 25*rate = 0.60
+
+```
+
+# does your submission require gpu for evaluation (inflation)?
+yes
+
+# did you include the compression script? and want it to be merged?
+no
+
+# additional comments
+In theory it should still run on CPU in <30 min. The submission name is kind of true but doesn't really encompass what the model does. At least its not a ffmpeg settings tweak.
+
+On an unrelated note, I think the problem with challenges that also have open source submissions is that people are incentivized to not submit until the last minute to prevent other people from tweaking some tiny thing and surpassing their score. I kind of wanted to see some cool submissions, but I'm assuming the reason there's none yet is because they're saving all of their competitive submissions. I think <0.50 is pretty easily possible as a slightly different architecture gets a score at least 10% better, so there's definitely a lot of room to go. I minimally obfuscated the exact model architecture so people can't trivially spend more compute and clog up the leaderboards with exact clones to try to get prizes.

public_pr53_intake_20260505_auto/pr_metadata.json

@@ -0,0 +1,16 @@
+{
+  "additions": 174,
+  "author": "Quantizr",
+  "changed_files": 3,
+  "closed_at": "2026-04-13T16:32:26Z",
+  "created_at": "2026-04-11T08:44:37Z",
+  "deletions": 0,
+  "head_repo": "Quantizr/comma_video_compression_challenge",
+  "head_sha": "ef78345815e5a31548dea65f30cf2942b36d0ddf",
+  "leaderboard_name": "mask2mask",
+  "leaderboard_score": 0.602,
+  "merged_at": null,
+  "pr_number": 53,
+  "state": "closed",
+  "title": "mask2mask (0.60)"
+}

public_pr56_intake_20260505_auto/INTAKE_LOG.md

@@ -0,0 +1,8 @@
+# PR #56 intake — selfcomp (score 0.382)
+Started UTC: 2026-05-05T22:44:57.919511+00:00
+
+✓ pr_metadata.json (head_sha=2b2d76de6f5aa34c76352c2cc02b03ed44a03a26, author=szabolcs-cs)
+✓ pr_body.md (1972 chars)
+✓ archive.zip exists (279036 bytes, sha256 241da6aa0a82fb01...) — skipping
+
+Status: archive=exists_skipped source=skipped_exists