Browse › Computer Vision › Image Captioning › COCO Captions
COCO Captions Benchmark (Image Captioning)
Image Captioning is the task of describing the content of an image in words. This task lies at the intersection of computer vision and natural language processing. Most image captioning systems use an encoder-decoder framework, where an input image is encoded into an intermediate representation of the information in the image, and then decoded into a descriptive text sequence. The most popular benchmarks are nocaps and COCO, and models are typically evaluated according to a BLEU or CIDER metric.
The archive carries no text for this table; the description above is the archive's text for the task Image Captioning. archive 2025-07-28
Over time archive 2025-07-28
The chart needs JavaScript; the table below carries every value.
Direction inferred from the metric name, not from the archive: BLEU-4 (higher is better), ROUGE-L (higher is better), BLEU-1 (higher is better), BLEU-2 (higher is better), BLEU-3 (higher is better). Not inferred (points only, no best-so-far line): CIDER, METEOR, SPICE, CLIPScore. Points are placed at the row's paper date; 40 of 41 rows carry one.
Results archive 2025-07-28
Archive rows end at the archive snapshot, 2025-07-28: no result published after that date is in this table. Rank is the archive's row order at that snapshot; not re-ranked here. Metric values are the archive's strings. Column headers sort the table in your browser; each row keeps its archive rank.
| Paper | Code | Ran Syntology | Report | |||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| 1 | mPLUG | 46.5 | 155.1 | 32.0 | 26.0 | – | Paper | Code | 2022 | linked, not harvested | report | |||||
| 2 | OFA | 44.9 | 154.9 | 32.5 | 26.6 | – | Paper | Code | 2022 | 1 of 1 ran · 0 unverified | report | |||||
| 3 | GIT | 44.1 | 151.1 | 32.2 | 26.3 | – | Paper | Code | 2022 | 0 of 21 ran · 21 unverified | report | |||||
| 4 | BLIP-2 ViT-G OPT 2.7B (zero-shot) | 43.7 | 145.8 | – | Paper | Code | 2023 | 4 of 8 ran · 4 unverified | report | |||||||
| 5 | BLIP-2 ViT-G OPT 6.7B (zero-shot) | 43.5 | 145.2 | – | Paper | Code | 2023 | 4 of 8 ran · 4 unverified | report | |||||||
| 6 | ExpansionNet v2 (No VL pretraining) | 42.7 | 143.7 | 30.6 | 24.7 | 61.1 | 83.5 | – | Paper | Code | 2022 | linked, not harvested | report | |||
| 7 | LEMON | 42.6 | 145.5 | 31.4 | 25.5 | – | Paper | – | 2021 | no code linked | report | |||||
| 8 | BLIP-2 ViT-G FlanT5 XL (zero-shot) | 42.4 | 144.5 | – | Paper | Code | 2023 | 4 of 8 ran · 4 unverified | report | |||||||
| 9 | GRIT (No VL pretraining - base) | 42.4 | 144.2 | 30.6 | 24.3 | 60.7 | 84.2 | – | Paper | Code | 2022 | linked, not harvested | report | |||
| 10 | Prompt Tuning | 41.81 | 141.4 | 31.51 | 24.42 | – | Paper | Code | 2022 | 1 of 1 ran · 0 unverified | report | |||||
| 11 | Oscar | 41.7 | 140 | 30.6 | 24.5 | – | Paper | Code | 2020 | 11 of 23 ran · 12 unverified | report | |||||
| 12 | Xmodal-Ctx | 41.4 | 139.9 | 30.4 | 24.0 | 60.4 | 83.4 | – | Paper | Code | 2022 | linked, not harvested | report | |||
| 13 | Xmodal-Ctx + OSCAR | 41.3 | 142.2 | 24.9 | – | Paper | Code | 2022 | linked, not harvested | report | ||||||
| 14 | X-VLM (base) | 41.3 | 140.8 | – | Paper | Code | 2021 | 1 of 1 ran · 0 unverified | report | |||||||
| 15 | VinVL | 41.0 | 140.9 | 31.1 | 25.2 | – | Paper | Code | 2021 | 2 of 2 ran · 0 unverified | report | |||||
| 16 | CoCa | 40.9 | 143.6 | 33.9 | 24.7 | – | Paper | Code | 2022 | 9 of 17 ran · 8 unverified | report | |||||
| 17 | SimVLM | 40.6 | 143.3 | 33.4 | 25.4 | – | Paper | Code | 2021 | 18 of 37 ran · 19 unverified | report | |||||
| 18 | Prismer | 40.4 | 136.5 | 31.4 | 24.4 | – | Paper | Code | 2023 | linked, not harvested | report | |||||
| 19 | PTP-BLIP (14M) | 40.1 | 135.0 | 30.4 | 23.7 | – | Paper | Code | 2022 | 3 of 5 ran · 2 unverified | report | |||||
| 20 | L-Verse | 39.9 | 31.4 | 23.3 | 60.4 | – | Paper | Code | 2021 | linked, not harvested | report | |||||
| 21 | Xmodal-Ctx | 39.7 | 135.9 | 30.0 | 23.7 | 59.5 | 81.5 | – | Paper | Code | 2022 | linked, not harvested | report | |||
| 22 | X-Transformer | 39.7 | 132.8 | 29.5 | 23.4 | 59.1 | 80.9 | 65.8 | 51.5 | – | Paper | Code | 2020 | linked, not harvested | report | |
| 23 | AoANet + VC | 39.5 | 29.3 | 59.3 | – | Paper | Code | 2020 | 1 of 4 ran · 3 unverified | report | ||||||
| 24 | Transformer_NSC | 39.4 | 129.6 | 28.9 | 22.8 | 58.7 | 80.7 | 65.6 | 51.3 | – | Paper | Code | 2020 | linked, not harvested | report | |
| 25 | Meshed-Memory Transformer | 39.1 | 131.2 | 29.2 | 22.6 | 58.6 | 80.8 | – | Paper | Code | 2019 | 4 of 8 ran · 4 unverified | report | |||
| 26 | CLIP Text Encoder (RL w/ CIDEr-reward) | 38.2 | 124.9 | 28.7 | 58.5 | – | Paper | Code | 2022 | linked, not harvested | report | |||||
| 27 | RefineCap (w/ REINFORCE) | 37.8 | 127.2 | 28.3 | 22.5 | 58.0 | 80.2 | 64.5 | 49.9 | – | Paper | – | 2021 | no code linked | report | |
| 28 | RDN | 37.3 | 125.2 | 28.1 | 57.4 | 80.2 | – | Paper | – | 2019 | no code linked | report | ||||
| 29 | SmallCapd=16, Large | 37.2 | 121.8 | 28.3 | 21.5 | – | Paper | Code | 2022 | linked, not harvested | report | |||||
| 30 | ClipCap (Transformer) | 33.53 | 113.08 | 27.45 | 21.05 | – | Paper | Code | 2021 | 3 of 6 ran · 3 unverified | report | |||||
| 31 | ClipCap (MLP + GPT2 tuning) | 32.15 | 108.35 | 27.1 | 20.12 | – | Paper | Code | 2021 | 3 of 6 ran · 3 unverified | report | |||||
| 32 | CapDec | 26.4 | 91.8 | 25.1 | – | Paper | Code | 2022 | 4 of 5 ran · 1 unverified | report | ||||||
| 33 | From Captions to Visual Concepts and Back | 25.7 | 23.6 | – | Paper | Code | 2014 | 0 of 6 ran · 6 unverified | report | |||||||
| 34 | VLKD (ViT-B/16) | 16.7 | 58.3 | 19.7 | 13.4 | – | Paper | – | 2021 | no code linked | report | |||||
| 35 | LaDiC (ours, 30 steps) | 0.382 | 126.2 | 29.5 | – | Paper | Code | 2024 | linked, not harvested | report | ||||||
| 36 | VALOR | 152.5 | 25.7 | ✓ | Paper | Code | 2023 | linked, not harvested | report | |||||||
| 37 | VAST | 149.0 | 27.0 | ✓ | Paper | Code | 2023 | 15 of 42 ran · 27 unverified | report | |||||||
| 38 | Virtex (ResNet-101) | 94 | 18.5 | – | Paper | Code | 2020 | 0 of 1 ran · 1 unverified | report | |||||||
| 39 | KOSMOS-1 (1.6B) (zero-shot) | 84.7 | 16.8 | – | – | – | not matched | report | ||||||||
| 40 | LaDiC | 22.4 | 58.7 | – | Paper | Code | 2024 | linked, not harvested | report | |||||||
| 41 | BLIP-FuseCap | 78.5 | – | Paper | Code | 2023 | linked, not harvested | report |
All 41 rows shown. 40 link to a paper page on this site; 2 are marked as using additional training data in the archive. No GitHub stars are tracked; "Code" is the first repository the archive lists for the row. The archive carries no row tags, review links or community-submitted rows for this table; none are shown. archive 2025-07-28
Syntology Ran reads "N of M ran · U unverified": of the M code samples Syntology harvested from repositories linked to that row's paper (joined by arXiv id), N executed on a synthesized input and the other U = M−N are unverified (harvested, no recorded run). It counts code from repositories linked to that row's paper, not this result: the row's number was not reproduced and nothing here is a correctness claim. The other cell texts mean no graph line for the row: "linked, not harvested" (the archive links code, Syntology has not harvested it), "no code linked" (no code link in the archive), "not matched" (the row's paper URL matched no paper on this site). 20 rows have a graph line, from 17 distinct papers; 17 rows (14 papers) have at least one sample that ran. Counting each paper once: Syntology ran 77 of 188 samples; 111 unverified. Separately, 32 of those 188 are pointer-only (licence): the site points at that code rather than redistributing it, a licence property recorded for ran and unverified samples alike; each cell's tooltip carries the row's own pointer-only count. Read from the graph 2026-09-24. Per-sample status is on the paper page.
Report a problem or propose a change · a person checks every report against the paper or source before anything changes; decisions are listed on /corrections