Redvirex/eflic-icml2026-repro-results
EF-LIC ICML 2026 reproduction This folder contains an independent verification wrapper for paper 654, Efficient Learned Image Compression without Entropy Coding (arXiv 2605.23323, OpenReview sfroJn43YY). It never modifies the authors' CC BY-NC-ND source: the GPU script clones commit b45653d77c0e03bf184a1b2d7b0f577ebc2956c9 into a temporary directory and imports it unchanged. The published Trackio logbook, interactive poster, and claim-by-claim evidence are at… See the full description on the dataset page: https://huggingface.co/datasets/Redvirex/eflic-icml2026-repro-results.
EF-LIC ICML 2026 reproduction
This folder contains an independent verification wrapper for paper 654, Efficient Learned Image Compression without Entropy Coding (arXiv 2605.23323, OpenReview sfroJn43YY). It never modifies the authors' CC BY-NC-ND source: the GPU script clones commit b45653d77c0e03bf184a1b2d7b0f577ebc2956c9 into a temporary directory and imports it unchanged.
The published Trackio logbook, interactive poster, and claim-by-claim evidence are at https://huggingface.co/spaces/Redvirex/eflic-icml2026-repro-logbook.
Contents
smoke_mechanisms.py: deterministic NumPy-only toy smoke test for entropy shaping and invertible conditional affine normalization.run_official_repro.py: released-checkpoint A100 evaluation, VQ entropy instrumentation, affine-decorrelation diagnostic, and latency benchmark.paper_claims.json: exact source-reported values and baseline correction.outputs/: captured local outputs and Hugging Face Job submission diagnostics.figures/: raw-data-backed figures added to the Trackio logbook.
The official checkpoint and benchmark images are downloaded at runtime and are not redistributed. The output manifest records source URLs, upstream commit, file sizes, and SHA-256 checksums. This respects the upstream and dataset terms while keeping the run auditable.
Outcome
The released-checkpoint entropy result reproduces at the highest-rate setting on all 24 Kodak images: 3.4346% weighted entropy gap versus 3.4200% reported. At that same rate, the affine transform reduces the diagnostic correlation with predicted context by 60.7% and lowers same-codebook latent MSE by 14.1%. The exact BD-rates remain source-only because raw MS-ILLM curves and integration details are unavailable. A100 latency remains source-only because the available HF credentials could not schedule a Job. EF-LIC-s remains source-only because no lightweight code or checkpoint was released.
Local smoke test
python repro_eflic/smoke_mechanisms.py \
--output repro_eflic/outputs/smoke_results.jsonHugging Face A100 run
hf jobs uv run --flavor a100-large --timeout 1h --detach --secrets HF_TOKEN \
repro_eflic/run_official_repro.py \
--datasets all \
--trackio-space Redvirex/eflic-icml2026-metrics \
--results-repo Redvirex/eflic-icml2026-repro-resultsThe paper's exact BD-rates cannot be independently recomputed because neither the raw MS-ILLM curves nor the entropy-coded ablation checkpoints are released. The script therefore emits new EF-LIC R-D points but never substitutes them for a two-curve BD-rate calculation.
Licensing
This is a mixed-license bundle; see LICENSING.md. The original reproduction wrapper is MIT-licensed. poster/poster.html is a derivative of Posterly's AGPL-3.0 template and retains Posterly's license and notices. The authors' EF-LIC source, checkpoint, and benchmark images are not redistributed.
If GPU Job authorization or credit is unavailable, a clearly labelled reduced CPU run can exercise the real released checkpoint:
uv run repro_eflic/run_official_repro.py \
--allow-cpu --datasets kodak --max-images 1 --skip-perceptual \
--skip-batch-benchmark --warmups 0 \
--trackio-space "" --results-repo "" \
--output-dir repro_eflic/outputs/cpu_subset