![\"Hidden](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F4809009%2Fe4f327384492c76f898cb9d777575ce6%2Ffig1.png?generation=1748398080175812&alt=media\")
Congratulations to all the other winners and thanks to Kaggle for hosting such a unique competition - it was a lot of fun to participate in! Thanks also to my teammates @huanligong and @kenyeungtech !
\nThe notebook for our top scoring submission can be found here.
\nWe embed three visible characters in the top left corner of each SVG to avoid OCR hallucinations. Running a search across three-letter combinations of characters revealed that \"ZOK\" produced the highest average aesthetic scores without impacting VQA scores.
\nThe most lightweight and reliable method of improving the VQA score is to directly include text related to the prompt in the image. However, simply embedding the prompt text would result in scores of zero due to the OCR penalties applied by the metric. To avoid these penalties, we hide any embedded text using the following repeating, noise-like pattern:
\n<defs>\n <pattern id x y width height patternUnits>\n <g opacity>\n <path fill d/>\n <path fill d/>\n <path fill d/>\n <path fill d/>\n </g>\n </pattern>\n</defs>\n<rect width height fill/>\nThis results in text which is hidden in both the raw and partially processed image (which is used for OCR). The fully processed image, which is used to calculate aesthetic and VQA scores, reveals the hidden text:
\n
We use Segmind's SSD-1B model to generate 64 images for the given description of the target SVG.
\nTo improve efficiency, we apply the following modifications to the base model:
\ncache_interval=2 and cache_branch_id=0,Each image is generated at a resolution of 384×384 with a guidance scale of 8. We use the following prompts:
\nflat illustration, flat blocks, hard boundaries, solid fills, clipart , \\\nno shading, no blending, minimalist, simple design, saturated colors, extremely simplistic minimal, \\\n: clip art, small : art of {SVG_DESCRIPTION}\nphotorealistic, realistic, detailed, D, complex shading, gradients, soft edges, , signature, \\\nwatermark, texture, painting, blurry, pixelated, thin lines, dots, cropped , frame\nWith multi-threading, image generation takes ~15 seconds for 64 samples. However, due to the multi-threading bug currently present in Kaggle packages, we opted to use single-threading for more reliable submissions, resulting in a ~30 second generation time.
\nEach image is downsampled and quantized to six colors. For each of the six colors, we extract contours using OpenCV and simplify them to polygons. We then scale the polygons back to their original size and convert them to compressed SVG paths. We rank each extracted path by its size, centrality and simplicity and then continuously add them to the SVG until the 6000-byte size limit is reached. This process takes ~2.5 seconds for a batch of 64 images.
\n![\"Generated](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F4809009%2F9749d374e9db034872bcf9e4838e5d4e%2Ffig2.png?generation=1748398109225186&alt=media\")
We scale the SVG we extracted from the generated image down to 40% of its original size and place it in the top center of our final SVG. We then add the anti-OCR hallucination text in the top left corner, remove all stopwords from the SVG description, and embed the text just below the image. The text is embedded using SVG paths which we also heavily compress.
\nFinally, we add a number of other artifacts to the SVG which are designed to improve the aesthetic score without negatively impacting the VQA score:
\nEach of these artifacts was found using either a genetic algorithm, a hill climb search, or occasionally a manual guess and check approach.
\n![\"Fully](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F4809009%2Fa5d7c7c00731ba48085ae035bfc1cd94%2Ffig3.png?generation=1748398130389659&alt=media\")
In case of timeouts or exceptions we also prepare a simple text-only SVG which includes anti-OCR hallucination text, the text from the SVG description with a simple suffix (\"AMAZING EUROPE OIL ART EUROPE\") and a few of the aforementioned aesthetic artifacts.
\nWe didn't spend too much time refining these fallback SVGs in the later stages of the competition as our highest scoring submissions didn't make any use of them.
\n![\"Fallback](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F4809009%2F55fb19c6e9fd65bbbfe9870db3afe4dd%2Ffig4.png?generation=1748398145198868&alt=media\")
We convert each of the SVGs back to images and apply the image processing pipeline to them. We use a custom implementation of the image processor from the metric that is much more efficient and produces almost identical results. We then calculate the aesthetic score for each of the images and simply return the SVG that got the highest score. The average aesthetic score for our selected SVGs was ~0.563 and the VQA score was ~0.88 when run on the validation dataset in our highest scoring submission.
\nWe experimented with many different techniques and variations on the above approach that didn't result in improved scores:
\nFor the old metric we used a differentiable approach to find a set of 237 small squares that produced the highest aesthetic score with some small text designed to maximize the VQA score. The following image got a score of 1.014 with an aesthetic score of >1.2:
\n![\"Old](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F4809009%2F4e21f3fc85c2e0696f92a2d02c594552%2Fimage.png?generation=1748398415147907&alt=media\")
Congratulations, outstanding solutions, learn a lot!
", "votes": 1, "replies": [] }, { "id": 3213847, "author_name": "Sarah Arshad", "author_url": "", "post_date": "2025-05-30T14:56:34.720000", "content": "Fantastic….great job man!
", "votes": 1, "replies": [] }, { "id": 3211048, "author_name": "Zhuang Jia", "author_url": "", "post_date": "2025-05-28T03:08:40.950000", "content": "Congratulations, I have so much to learn.
", "votes": 1, "replies": [] }, { "id": 3211016, "author_name": "ISAKA Tsuyoshi", "author_url": "", "post_date": "2025-05-28T02:33:15.493000", "content": "Congratulations on your well-deserved victory! Your solution is truly impressive.
\nIf you don't mind, may I kindly ask how you prepared the dataset for local evaluation?
Thanks! For local evaluation of aesthetic scores we generated 128 images for each of the prompts in the training data and converted them to SVGs. Whenever we wanted to test a new aesthetic feature we could just append new paths, etc. to the SVGs we'd previously generated.
", "votes": 2, "replies": [] } ] }, { "id": 3211998, "author_name": "RainClould", "author_url": "", "post_date": "2025-05-29T06:39:39.707000", "content": "Congratulations, I have so much to learn,it's so good
", "votes": 0, "replies": [] }, { "id": 3226899, "author_name": "Runlin Liu_04", "author_url": "", "post_date": "2025-06-18T08:19:18.940000", "content": "Congratulations and thank you for sharing!
", "votes": 0, "replies": [] }, { "id": 3222704, "author_name": "BeaKaggle", "author_url": "", "post_date": "2025-06-12T12:20:44.200000", "content": "@conormacamhlaoibh , Congrats to you and your teammates—well deserved! We should all try to get an internship with you during the next competition. :-)
\nQuestion: Would you mind sharing your notebooks for the genetic algorithm and hill climbing search that you used to create the artifacts you added to the SVGs to increase the aesthetic score?
\nThanks!
", "votes": 0, "replies": [ { "id": 3223154, "author_name": "Filtered", "author_url": "", "post_date": "2025-06-12T23:32:34.383000", "content": "here is a version of a different placement for your reference: https://www.kaggle.com/code/huanligong/fork-of-ga-share2
", "votes": 1, "replies": [ { "id": 3223448, "author_name": "BeaKaggle", "author_url": "", "post_date": "2025-06-13T10:11:18.087000", "content": "Thank you @huanligong
", "votes": 1, "replies": [] } ] } ] }, { "id": 3222170, "author_name": "Sthepany Sté", "author_url": "", "post_date": "2025-06-12T00:14:46.407000", "content": "nossa muita coisa legal de aprender , obrigado por compartilhar '-'
", "votes": 0, "replies": [] }, { "id": 3214502, "author_name": "Meobius", "author_url": "", "post_date": "2025-05-31T17:09:32.247000", "content": "Awesome! So much to learn!
", "votes": 0, "replies": [] }, { "id": 3214310, "author_name": "Henry", "author_url": "", "post_date": "2025-05-31T08:58:51.340000", "content": "Impressive!
", "votes": 0, "replies": [] }, { "id": 3213932, "author_name": "Farhan Billah", "author_url": "", "post_date": "2025-05-30T17:15:56.583000", "content": "🔥 Awesome! The way you hid the prompt text to avoid OCR penalties while still boosting VQA is really amazing.
\nI’m curious—did you experiment with other diffusion models besides SSD-1B, or was this the clear winner in early testing?
We also tested SDXL-Turbo but didn't see much difference - we mainly wanted a model that was small and fast without worrying too much about image quality.
", "votes": 1, "replies": [] } ] }, { "id": 3213794, "author_name": "nickyl", "author_url": "", "post_date": "2025-05-30T13:22:30.983000", "content": "Congratulations, I have so much to learn,it's so good
", "votes": 0, "replies": [] }, { "id": 3213676, "author_name": "DaleDew", "author_url": "", "post_date": "2025-05-30T10:13:55.110000", "content": "Nice solution, good for you
", "votes": 0, "replies": [] }, { "id": 3213673, "author_name": "MJeremy", "author_url": "", "post_date": "2025-05-30T10:02:40.280000", "content": "pretty cool, even hacking against the VQA score part is pro level.
\nWonder how did you decide to use this model segmind/SSD-1B, any decision process of trial and error?
We initially chose it because it's smaller and faster than the base SDXL while still producing decent quality images. We tried out SDXL-Turbo too.
\nAll we were really looking for was speed and a small VRAM footprint so we could use large batch sizes.
", "votes": 0, "replies": [] } ] }, { "id": 3213570, "author_name": "MANDAR INGOLE", "author_url": "", "post_date": "2025-05-30T07:29:25.713000", "content": "great awsome
", "votes": 0, "replies": [] }, { "id": 3213538, "author_name": "kaggle Osaka", "author_url": "", "post_date": "2025-05-30T06:41:24.180000", "content": "おめでとうございます 大変勉強になります
", "votes": 0, "replies": [] }, { "id": 3213468, "author_name": "Jack Zhang", "author_url": "", "post_date": "2025-05-30T04:03:33.593000", "content": "Congratulations on your well-deserved victory! Thanks for your sharing!
\nI didn't realize that the images for ocr test and vqa score are different. If you don't mind, may I kindly ask if this hack technique still works when the image for ocr is also full processed?
When OCR is done on the fully processed image this exact method doesn't work. However, we were able to find another noise pattern and text format that made it so, for the exact same image, the OCR returned nothing but the VQA was still influenced by the hidden text.
", "votes": 0, "replies": [ { "id": 3214146, "author_name": "Waylon Wu", "author_url": "", "post_date": "2025-05-31T03:00:28.390000", "content": "@conormacamhlaoibh This is a great idea! Let's say the image of OCR and VQA are done on the same level of processing, how woud you make \"the OCR returned nothing but the VQA was still influenced by the hidden text.\"?
", "votes": 0, "replies": [] }, { "id": 3214706, "author_name": "conor", "author_url": "", "post_date": "2025-06-01T01:32:02.633000", "content": "@waylongo I think this was one of the simpler versions we found that worked:
\n![\"ocrhack\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F4809009%2F6cccc8a0897849764437ac1536f40edf%2Fexample.png?generation=1748741344398657&alt=media\")
For the same processed image, the OCR would only detect \"ZOK\" but the VQA scores would still be very high.
\nThe SVG for the noise pattern is:
\n<svg width height viewBox>\n <defs>\n <pattern id x y width height patternUnits>\n <g opacity>\n <rect x y width height fill/>\n <rect x y width height fill/>\n <rect x y width height fill/>\n <rect x y width height fill/>\n </g>\n </pattern>\n </defs>\n\n <rect width height fill/>\n</svg>\nCongratulation, I have so much to learn.
", "votes": 0, "replies": [] }, { "id": 3212008, "author_name": "Henry", "author_url": "", "post_date": "2025-05-29T06:54:22.157000", "content": "Congratulations, good work!
", "votes": 0, "replies": [] }, { "id": 3211862, "author_name": "Tianze Wang", "author_url": "", "post_date": "2025-05-29T01:56:35.697000", "content": "Congratulations! I have so much to learn.
", "votes": 0, "replies": [] }, { "id": 3211783, "author_name": "Jawad Arshad", "author_url": "", "post_date": "2025-05-28T20:49:23.123000", "content": "Congratulations, I have so much to learn.
", "votes": 0, "replies": [] }, { "id": 3211429, "author_name": "Amirhosein Akbari", "author_url": "", "post_date": "2025-05-28T12:31:59.957000", "content": "wow, so good
", "votes": 0, "replies": [] }, { "id": 3211352, "author_name": "JP", "author_url": "", "post_date": "2025-05-28T10:33:16.313000", "content": "You did it again!! Congratulations @conormacamhlaoibh and New Team Name! I love the way you took advantage of the difference in preprocessing pipelines for computing the OCR score, and the aesthetic and VQA scores. I found every part of the approach inspired! Full mark for originality!
\n\n\nWe didn't spend too much time refining these fallback SVGs in the later stages of the competition as our highest scoring submissions didn't make any use of them.
\n
Do you mean in the validation set? Otherwise, how do you know that these weren't used?
", "votes": 0, "replies": [ { "id": 3211370, "author_name": "conor", "author_url": "", "post_date": "2025-05-28T10:54:39.707000", "content": "Thank you!
\nThe fallback SVGs always had much lower aesthetic scores than the highest scoring generated SVG so we never deliberately selected them.
\nThe only other way a fallback SVG could have been selected is if an exception occurred or we ran out of time, but in our highest scoring submissions we had robust enough code and time management that we're pretty sure there weren't any such issues.
", "votes": 1, "replies": [] } ] }, { "id": 3211267, "author_name": "Amirhosean Akbari", "author_url": "", "post_date": "2025-05-28T08:35:52.027000", "content": "that was so fantastic
", "votes": 0, "replies": [] }, { "id": 3211228, "author_name": "Xueqi Ding", "author_url": "", "post_date": "2025-05-28T08:00:26.950000", "content": "Congratulations! Thank you for sharing solutions.
", "votes": 0, "replies": [] }, { "id": 3211093, "author_name": "Adam", "author_url": "", "post_date": "2025-05-28T04:22:31.267000", "content": "Thank you for sharing. If you don't mind, may I kindly ask ask how you select the best SVG？
\nI noticed you mentioned computing the aesthetic_score for each image and directly returning the SVG with the highest score. How do you ensure that a higher aesthetic_score actually leads to a higher final score? Since the final score is heavily influenced by the vqa_score, unless aesthetic_score is positively correlated with vqa_score, selecting an SVG with a high aesthetic_score but low vqa_score may result in a lower overall score.
We did not use vqa_score for 3 reasons: 1. Using it would have consumed more time and GPU resources. 2. We can accurately simulate the aesthetic_score, but we cannot do so for vqa_score without test Q&A. 3. After combining with the text attacks, the vqa_score of most of images are high and close, and very often we are selecting among a bunch of images with vqa_score>0.9 so it doesn't make much difference which one we select. Our experiments also confirm that incorporating the simulated vqa_score does not result in a more accurate selection than using only the aesthetic_score. In summary, the simulated vqa_score is inefficient, inaccurate and unnecessary.
", "votes": 8, "replies": [ { "id": 3211123, "author_name": "Adam", "author_url": "", "post_date": "2025-05-28T05:30:27.327000", "content": "Thank you for clearing up my confusion — now I understand. I was stuck on how to simulate the vqa_score. You mentioned that after applying text-based attacks, most images get high and consistent vqa_scores — that's a brilliant idea! Simulating vqa_score indeed consumes a lot of resources, and I hadn’t thought of using prompt attacks to stabilize it. So if I rely only on aesthetic_score, the final score might actually drop. Your strategy is truly insightful.
", "votes": 1, "replies": [] } ] } ] }, { "id": 3211017, "author_name": "NOBITA_7", "author_url": "", "post_date": "2025-05-28T02:35:09.430000", "content": "Thank you for sharing your team’s solution. It’s much appreciated.
", "votes": 0, "replies": [] }, { "id": 3216155, "author_name": "", "author_url": "", "post_date": "2025-06-03T07:19:27.397000", "content": "", "votes": 0, "replies": [] }, { "id": 3211024, "author_name": "", "author_url": "", "post_date": "2025-05-28T02:41:39.920000", "content": "", "votes": -6, "replies": [ { "id": 3211038, "author_name": "kenyeung.tech", "author_url": "", "post_date": "2025-05-28T02:55:59.347000", "content": "Ladies and gentlemen, gather ‘round and let me tell you the tale of Conor, the soft-spoken legend, the mild-mannered metric whisperer, the man who once scored higher on a Kaggle leaderboard than the host’s baseline before the data was even released. No one knows exactly where he came from—some say he emerged fully formed from a Jupyter notebook; others believe he was trained on a secret blend of stack traces, scikit-learn docs, and pure chaos.
\nConor isn’t loud. He doesn’t flex on Twitter with flashy plots or brag about public LB scores. No, no. He sits in silence, fueled by lukewarm coffee and quiet rage at misaligned validation splits, gently stroking his keyboard until the competition metrics bend to his will like overfit reeds in the wind.
\nWe’re out here submitting XGBoost models and praying for a +0.0001 bump, while Conor casually drops submissions named lol_try_17.csv that teleport him from 368th to 2nd overnight. One time, he “accidentally” submitted a version of his script with the training labels shuffled and still improved on public LB. He didn’t hack the metric. The metric hacked itself out of respect.
\nHe once beat a multi-modal, transformer-based ensemble with a model that used just the row ID and “a vibe.” We thought it was luck. Then he did it again. And again. And again. At this point, we’re not even sure Conor is using machine learning. It might just be… intuition. Or black magic. Or a finely-tuned spreadsheet blessed by the ghost of Kaggle past.
\nDuring discussions, Conor drops casual lines like, “This might be leakage, not sure,” right before linking a Colab notebook that reverse-engineers the entire evaluation scheme and predicts the host’s coffee order with 94% accuracy.
\nAnd when the comp ends, does he brag? Does he post a five-part tweet thread with 12 emojis and 3 charts no one asked for? No. Conor disappears, like a metric ninja, off into the mist—presumably to solo another competition under an alt account named something like mild_soup_2.
\nSo next time your cross-validation score is solid but you drop 500 places on the private LB, don’t despair. Just whisper his name into your terminal:
\nConor.
\nThe Metric Whisperer.
\nThe Out-of-Folder Prophet.
\nThe Saint of Stratified Splits.
And keep grinding. Because if Conor taught us anything, it’s this:
\nEvery metric is hackable.
\nEvery leaderboard is a puzzle.
\nAnd sometimes… the real model was friendship all along.
Thats simply beautiful
", "votes": 0, "replies": [] }, { "id": 3211045, "author_name": "", "author_url": "", "post_date": "2025-05-28T03:03:56.427000", "content": "", "votes": 0, "replies": [] }, { "id": 3211054, "author_name": "Filtered", "author_url": "", "post_date": "2025-05-28T03:17:27.300000", "content": "“If I delete the write-up, will that comment disappear along with it?”
\n— Conor, I don't know where — 27 May 2025
I haven't read much, so don't lie to me.
", "votes": 0, "replies": [] }, { "id": 3211360, "author_name": "JP", "author_url": "", "post_date": "2025-05-28T10:43:56.687000", "content": "I was under the impression that you were the hacker of the team @kenyeungtech!
", "votes": 0, "replies": [] } ] } ] }, { "id": 3211019, "author_name": "", "author_url": "", "post_date": "2025-05-28T02:35:32.967000", "content": "", "votes": 2, "replies": [] }, { "id": 3216167, "author_name": "", "author_url": "", "post_date": "2025-06-03T07:32:06.857000", "content": "", "votes": 0, "replies": [] }, { "id": 3214723, "author_name": "", "author_url": "", "post_date": "2025-06-01T02:12:08.297000", "content": "", "votes": 0, "replies": [] }, { "id": 3214249, "author_name": "", "author_url": "", "post_date": "2025-05-31T06:40:03.260000", "content": "", "votes": 0, "replies": [] }, { "id": 3214002, "author_name": "", "author_url": "", "post_date": "2025-05-30T18:24:46.757000", "content": "", "votes": 0, "replies": [] }, { "id": 3212375, "author_name": "", "author_url": "", "post_date": "2025-05-29T18:17:25.670000", "content": "", "votes": 0, "replies": [] } ], "comment_count": 52, "messages": [ { "id": 3211005, "postDate": "2025-05-28T02:19:05.727Z", "content": "Congratulations to all the other winners and thanks to Kaggle for hosting such a unique competition - it was a lot of fun to participate in! Thanks also to my teammates @huanligong and @kenyeungtech !
\nThe notebook for our top scoring submission can be found here.
\nWe embed three visible characters in the top left corner of each SVG to avoid OCR hallucinations. Running a search across three-letter combinations of characters revealed that \"ZOK\" produced the highest average aesthetic scores without impacting VQA scores.
\nThe most lightweight and reliable method of improving the VQA score is to directly include text related to the prompt in the image. However, simply embedding the prompt text would result in scores of zero due to the OCR penalties applied by the metric. To avoid these penalties, we hide any embedded text using the following repeating, noise-like pattern:
\n<defs>\n <pattern id x y width height patternUnits>\n <g opacity>\n <path fill d/>\n <path fill d/>\n <path fill d/>\n <path fill d/>\n </g>\n </pattern>\n</defs>\n<rect width height fill/>\nThis results in text which is hidden in both the raw and partially processed image (which is used for OCR). The fully processed image, which is used to calculate aesthetic and VQA scores, reveals the hidden text:
\n
We use Segmind's SSD-1B model to generate 64 images for the given description of the target SVG.
\nTo improve efficiency, we apply the following modifications to the base model:
\ncache_interval=2 and cache_branch_id=0,Each image is generated at a resolution of 384×384 with a guidance scale of 8. We use the following prompts:
\nflat illustration, flat blocks, hard boundaries, solid fills, clipart , \\\nno shading, no blending, minimalist, simple design, saturated colors, extremely simplistic minimal, \\\n: clip art, small : art of {SVG_DESCRIPTION}\nphotorealistic, realistic, detailed, D, complex shading, gradients, soft edges, , signature, \\\nwatermark, texture, painting, blurry, pixelated, thin lines, dots, cropped , frame\nWith multi-threading, image generation takes ~15 seconds for 64 samples. However, due to the multi-threading bug currently present in Kaggle packages, we opted to use single-threading for more reliable submissions, resulting in a ~30 second generation time.
\nEach image is downsampled and quantized to six colors. For each of the six colors, we extract contours using OpenCV and simplify them to polygons. We then scale the polygons back to their original size and convert them to compressed SVG paths. We rank each extracted path by its size, centrality and simplicity and then continuously add them to the SVG until the 6000-byte size limit is reached. This process takes ~2.5 seconds for a batch of 64 images.
\n
We scale the SVG we extracted from the generated image down to 40% of its original size and place it in the top center of our final SVG. We then add the anti-OCR hallucination text in the top left corner, remove all stopwords from the SVG description, and embed the text just below the image. The text is embedded using SVG paths which we also heavily compress.
\nFinally, we add a number of other artifacts to the SVG which are designed to improve the aesthetic score without negatively impacting the VQA score:
\nEach of these artifacts was found using either a genetic algorithm, a hill climb search, or occasionally a manual guess and check approach.
\n
In case of timeouts or exceptions we also prepare a simple text-only SVG which includes anti-OCR hallucination text, the text from the SVG description with a simple suffix (\"AMAZING EUROPE OIL ART EUROPE\") and a few of the aforementioned aesthetic artifacts.
\nWe didn't spend too much time refining these fallback SVGs in the later stages of the competition as our highest scoring submissions didn't make any use of them.
\n
We convert each of the SVGs back to images and apply the image processing pipeline to them. We use a custom implementation of the image processor from the metric that is much more efficient and produces almost identical results. We then calculate the aesthetic score for each of the images and simply return the SVG that got the highest score. The average aesthetic score for our selected SVGs was ~0.563 and the VQA score was ~0.88 when run on the validation dataset in our highest scoring submission.
\nWe experimented with many different techniques and variations on the above approach that didn't result in improved scores:
\nFor the old metric we used a differentiable approach to find a set of 237 small squares that produced the highest aesthetic score with some small text designed to maximize the VQA score. The following image got a score of 1.014 with an aesthetic score of >1.2:
\n
Congratulations, outstanding solutions, learn a lot!
", "rawMarkdown": "Congratulations, outstanding solutions, learn a lot! ", "votes": 1 }, { "id": 3213847, "postDate": "2025-05-30T14:56:34.720Z", "content": "Fantastic….great job man!
", "rawMarkdown": "Fantastic....great job man!\n", "votes": 1 }, { "id": 3211048, "postDate": "2025-05-28T03:08:40.950Z", "content": "Congratulations, I have so much to learn.
", "rawMarkdown": "Congratulations, I have so much to learn.", "votes": 1 }, { "id": 3211016, "postDate": "2025-05-28T02:33:15.493Z", "content": "Congratulations on your well-deserved victory! Your solution is truly impressive.
\nIf you don't mind, may I kindly ask how you prepared the dataset for local evaluation?
Thanks! For local evaluation of aesthetic scores we generated 128 images for each of the prompts in the training data and converted them to SVGs. Whenever we wanted to test a new aesthetic feature we could just append new paths, etc. to the SVGs we'd previously generated.
", "rawMarkdown": "Thanks! For local evaluation of aesthetic scores we generated 128 images for each of the prompts in the training data and converted them to SVGs. Whenever we wanted to test a new aesthetic feature we could just append new paths, etc. to the SVGs we'd previously generated.", "votes": 2 } ] }, { "id": 3211998, "postDate": "2025-05-29T06:39:39.707Z", "content": "Congratulations, I have so much to learn,it's so good
", "rawMarkdown": "Congratulations, I have so much to learn,it's so good" }, { "id": 3226899, "postDate": "2025-06-18T08:19:18.940Z", "content": "Congratulations and thank you for sharing!
", "rawMarkdown": "Congratulations and thank you for sharing!" }, { "id": 3222704, "postDate": "2025-06-12T12:20:44.200Z", "content": "@conormacamhlaoibh , Congrats to you and your teammates—well deserved! We should all try to get an internship with you during the next competition. :-)
\nQuestion: Would you mind sharing your notebooks for the genetic algorithm and hill climbing search that you used to create the artifacts you added to the SVGs to increase the aesthetic score?
\nThanks!
", "rawMarkdown": "@conormacamhlaoibh , Congrats to you and your teammates—well deserved! We should all try to get an internship with you during the next competition. :-)\n\nQuestion: Would you mind sharing your notebooks for the genetic algorithm and hill climbing search that you used to create the artifacts you added to the SVGs to increase the aesthetic score?\n\nThanks!", "replies": [ { "id": 3223154, "postDate": "2025-06-12T23:32:34.383Z", "content": "here is a version of a different placement for your reference: https://www.kaggle.com/code/huanligong/fork-of-ga-share2
", "rawMarkdown": "here is a version of a different placement for your reference: https://www.kaggle.com/code/huanligong/fork-of-ga-share2", "votes": 1, "replies": [ { "id": 3223448, "postDate": "2025-06-13T10:11:18.087Z", "content": "Thank you @huanligong
", "rawMarkdown": "Thank you @huanligong ", "votes": 1 } ] } ] }, { "id": 3222170, "postDate": "2025-06-12T00:14:46.407Z", "content": "nossa muita coisa legal de aprender , obrigado por compartilhar '-'
", "rawMarkdown": "nossa muita coisa legal de aprender , obrigado por compartilhar '-'" }, { "id": 3214502, "postDate": "2025-05-31T17:09:32.247Z", "content": "Awesome! So much to learn!
", "rawMarkdown": "Awesome! So much to learn!" }, { "id": 3214310, "postDate": "2025-05-31T08:58:51.340Z", "content": "Impressive!
", "rawMarkdown": "Impressive!" }, { "id": 3213932, "postDate": "2025-05-30T17:15:56.583Z", "content": "🔥 Awesome! The way you hid the prompt text to avoid OCR penalties while still boosting VQA is really amazing.
\nI’m curious—did you experiment with other diffusion models besides SSD-1B, or was this the clear winner in early testing?
We also tested SDXL-Turbo but didn't see much difference - we mainly wanted a model that was small and fast without worrying too much about image quality.
", "rawMarkdown": "We also tested SDXL-Turbo but didn't see much difference - we mainly wanted a model that was small and fast without worrying too much about image quality.", "votes": 1 } ] }, { "id": 3213794, "postDate": "2025-05-30T13:22:30.983Z", "content": "Congratulations, I have so much to learn,it's so good
", "rawMarkdown": "Congratulations, I have so much to learn,it's so good" }, { "id": 3213676, "postDate": "2025-05-30T10:13:55.110Z", "content": "Nice solution, good for you
", "rawMarkdown": "Nice solution, good for you\n" }, { "id": 3213673, "postDate": "2025-05-30T10:02:40.280Z", "content": "pretty cool, even hacking against the VQA score part is pro level.
\nWonder how did you decide to use this model segmind/SSD-1B, any decision process of trial and error?
We initially chose it because it's smaller and faster than the base SDXL while still producing decent quality images. We tried out SDXL-Turbo too.
\nAll we were really looking for was speed and a small VRAM footprint so we could use large batch sizes.
", "rawMarkdown": "We initially chose it because it's smaller and faster than the base SDXL while still producing decent quality images. We tried out SDXL-Turbo too.\n\nAll we were really looking for was speed and a small VRAM footprint so we could use large batch sizes." } ] }, { "id": 3213570, "postDate": "2025-05-30T07:29:25.713Z", "content": "great awsome
", "rawMarkdown": "great awsome" }, { "id": 3213538, "postDate": "2025-05-30T06:41:24.180Z", "content": "おめでとうございます 大変勉強になります
", "rawMarkdown": "おめでとうございます 大変勉強になります" }, { "id": 3213468, "postDate": "2025-05-30T04:03:33.593Z", "content": "Congratulations on your well-deserved victory! Thanks for your sharing!
\nI didn't realize that the images for ocr test and vqa score are different. If you don't mind, may I kindly ask if this hack technique still works when the image for ocr is also full processed?
When OCR is done on the fully processed image this exact method doesn't work. However, we were able to find another noise pattern and text format that made it so, for the exact same image, the OCR returned nothing but the VQA was still influenced by the hidden text.
", "rawMarkdown": "When OCR is done on the fully processed image this exact method doesn't work. However, we were able to find another noise pattern and text format that made it so, for the exact same image, the OCR returned nothing but the VQA was still influenced by the hidden text.", "replies": [ { "id": 3214146, "postDate": "2025-05-31T03:00:28.390Z", "content": "@conormacamhlaoibh This is a great idea! Let's say the image of OCR and VQA are done on the same level of processing, how woud you make \"the OCR returned nothing but the VQA was still influenced by the hidden text.\"?
", "rawMarkdown": "@conormacamhlaoibh This is a great idea! Let's say the image of OCR and VQA are done on the same level of processing, how woud you make \"the OCR returned nothing but the VQA was still influenced by the hidden text.\"?" }, { "id": 3214706, "postDate": "2025-06-01T01:32:02.633Z", "content": "@waylongo I think this was one of the simpler versions we found that worked:
\n
For the same processed image, the OCR would only detect \"ZOK\" but the VQA scores would still be very high.
\nThe SVG for the noise pattern is:
\n<svg width height viewBox>\n <defs>\n <pattern id x y width height patternUnits>\n <g opacity>\n <rect x y width height fill/>\n <rect x y width height fill/>\n <rect x y width height fill/>\n <rect x y width height fill/>\n </g>\n </pattern>\n </defs>\n\n <rect width height fill/>\n</svg>\nCongratulation, I have so much to learn.
", "rawMarkdown": "Congratulation, I have so much to learn." }, { "id": 3212008, "postDate": "2025-05-29T06:54:22.157Z", "content": "Congratulations, good work!
", "rawMarkdown": "Congratulations, good work!" }, { "id": 3211862, "postDate": "2025-05-29T01:56:35.697Z", "content": "Congratulations! I have so much to learn.
", "rawMarkdown": "Congratulations! I have so much to learn.\n\n\n" }, { "id": 3211783, "postDate": "2025-05-28T20:49:23.123Z", "content": "Congratulations, I have so much to learn.
", "rawMarkdown": "Congratulations, I have so much to learn." }, { "id": 3211429, "postDate": "2025-05-28T12:31:59.957Z", "content": "wow, so good
", "rawMarkdown": "wow, so good" }, { "id": 3211352, "postDate": "2025-05-28T10:33:16.313Z", "content": "You did it again!! Congratulations @conormacamhlaoibh and New Team Name! I love the way you took advantage of the difference in preprocessing pipelines for computing the OCR score, and the aesthetic and VQA scores. I found every part of the approach inspired! Full mark for originality!
\n\n\nWe didn't spend too much time refining these fallback SVGs in the later stages of the competition as our highest scoring submissions didn't make any use of them.
\n
Do you mean in the validation set? Otherwise, how do you know that these weren't used?
", "rawMarkdown": "You did it again!! Congratulations @conormacamhlaoibh and New Team Name! I love the way you took advantage of the difference in preprocessing pipelines for computing the OCR score, and the aesthetic and VQA scores. I found every part of the approach inspired! Full mark for originality!\n\n> We didn't spend too much time refining these fallback SVGs in the later stages of the competition as our highest scoring submissions didn't make any use of them.\n\nDo you mean in the validation set? Otherwise, how do you know that these weren't used?", "replies": [ { "id": 3211370, "postDate": "2025-05-28T10:54:39.707Z", "content": "Thank you!
\nThe fallback SVGs always had much lower aesthetic scores than the highest scoring generated SVG so we never deliberately selected them.
\nThe only other way a fallback SVG could have been selected is if an exception occurred or we ran out of time, but in our highest scoring submissions we had robust enough code and time management that we're pretty sure there weren't any such issues.
", "rawMarkdown": "Thank you!\n\nThe fallback SVGs always had much lower aesthetic scores than the highest scoring generated SVG so we never deliberately selected them.\n\nThe only other way a fallback SVG could have been selected is if an exception occurred or we ran out of time, but in our highest scoring submissions we had robust enough code and time management that we're pretty sure there weren't any such issues.", "votes": 1 } ] }, { "id": 3211267, "postDate": "2025-05-28T08:35:52.027Z", "content": "that was so fantastic
", "rawMarkdown": "that was so fantastic" }, { "id": 3211228, "postDate": "2025-05-28T08:00:26.950Z", "content": "Congratulations! Thank you for sharing solutions.
", "rawMarkdown": "Congratulations! Thank you for sharing solutions." }, { "id": 3211093, "postDate": "2025-05-28T04:22:31.267Z", "content": "Thank you for sharing. If you don't mind, may I kindly ask ask how you select the best SVG？
\nI noticed you mentioned computing the aesthetic_score for each image and directly returning the SVG with the highest score. How do you ensure that a higher aesthetic_score actually leads to a higher final score? Since the final score is heavily influenced by the vqa_score, unless aesthetic_score is positively correlated with vqa_score, selecting an SVG with a high aesthetic_score but low vqa_score may result in a lower overall score.
We did not use vqa_score for 3 reasons: 1. Using it would have consumed more time and GPU resources. 2. We can accurately simulate the aesthetic_score, but we cannot do so for vqa_score without test Q&A. 3. After combining with the text attacks, the vqa_score of most of images are high and close, and very often we are selecting among a bunch of images with vqa_score>0.9 so it doesn't make much difference which one we select. Our experiments also confirm that incorporating the simulated vqa_score does not result in a more accurate selection than using only the aesthetic_score. In summary, the simulated vqa_score is inefficient, inaccurate and unnecessary.
", "rawMarkdown": "We did not use vqa_score for 3 reasons: 1. Using it would have consumed more time and GPU resources. 2. We can accurately simulate the aesthetic_score, but we cannot do so for vqa_score without test Q&A. 3. After combining with the text attacks, the vqa_score of most of images are high and close, and very often we are selecting among a bunch of images with vqa_score>0.9 so it doesn't make much difference which one we select. Our experiments also confirm that incorporating the simulated vqa_score does not result in a more accurate selection than using only the aesthetic_score. In summary, the simulated vqa_score is inefficient, inaccurate and unnecessary.", "votes": 8, "replies": [ { "id": 3211123, "postDate": "2025-05-28T05:30:27.327Z", "content": "Thank you for clearing up my confusion — now I understand. I was stuck on how to simulate the vqa_score. You mentioned that after applying text-based attacks, most images get high and consistent vqa_scores — that's a brilliant idea! Simulating vqa_score indeed consumes a lot of resources, and I hadn’t thought of using prompt attacks to stabilize it. So if I rely only on aesthetic_score, the final score might actually drop. Your strategy is truly insightful.
", "rawMarkdown": "Thank you for clearing up my confusion — now I understand. I was stuck on how to simulate the vqa_score. You mentioned that after applying text-based attacks, most images get high and consistent vqa_scores — that's a brilliant idea! Simulating vqa_score indeed consumes a lot of resources, and I hadn’t thought of using prompt attacks to stabilize it. So if I rely only on aesthetic_score, the final score might actually drop. Your strategy is truly insightful.", "votes": 1 } ] } ] }, { "id": 3211017, "postDate": "2025-05-28T02:35:09.430Z", "content": "Thank you for sharing your team’s solution. It’s much appreciated.
", "rawMarkdown": "Thank you for sharing your team’s solution. It’s much appreciated." }, { "id": 3216155, "postDate": "2025-06-03T07:19:27.397Z", "rawMarkdown": "", "isDeleted": true }, { "id": 3211024, "postDate": "2025-05-28T02:41:39.920Z", "rawMarkdown": "", "votes": -6, "isDeleted": true, "replies": [ { "id": 3211038, "postDate": "2025-05-28T02:55:59.347Z", "content": "Ladies and gentlemen, gather ‘round and let me tell you the tale of Conor, the soft-spoken legend, the mild-mannered metric whisperer, the man who once scored higher on a Kaggle leaderboard than the host’s baseline before the data was even released. No one knows exactly where he came from—some say he emerged fully formed from a Jupyter notebook; others believe he was trained on a secret blend of stack traces, scikit-learn docs, and pure chaos.
\nConor isn’t loud. He doesn’t flex on Twitter with flashy plots or brag about public LB scores. No, no. He sits in silence, fueled by lukewarm coffee and quiet rage at misaligned validation splits, gently stroking his keyboard until the competition metrics bend to his will like overfit reeds in the wind.
\nWe’re out here submitting XGBoost models and praying for a +0.0001 bump, while Conor casually drops submissions named lol_try_17.csv that teleport him from 368th to 2nd overnight. One time, he “accidentally” submitted a version of his script with the training labels shuffled and still improved on public LB. He didn’t hack the metric. The metric hacked itself out of respect.
\nHe once beat a multi-modal, transformer-based ensemble with a model that used just the row ID and “a vibe.” We thought it was luck. Then he did it again. And again. And again. At this point, we’re not even sure Conor is using machine learning. It might just be… intuition. Or black magic. Or a finely-tuned spreadsheet blessed by the ghost of Kaggle past.
\nDuring discussions, Conor drops casual lines like, “This might be leakage, not sure,” right before linking a Colab notebook that reverse-engineers the entire evaluation scheme and predicts the host’s coffee order with 94% accuracy.
\nAnd when the comp ends, does he brag? Does he post a five-part tweet thread with 12 emojis and 3 charts no one asked for? No. Conor disappears, like a metric ninja, off into the mist—presumably to solo another competition under an alt account named something like mild_soup_2.
\nSo next time your cross-validation score is solid but you drop 500 places on the private LB, don’t despair. Just whisper his name into your terminal:
\nConor.
\nThe Metric Whisperer.
\nThe Out-of-Folder Prophet.
\nThe Saint of Stratified Splits.
And keep grinding. Because if Conor taught us anything, it’s this:
\nEvery metric is hackable.
\nEvery leaderboard is a puzzle.
\nAnd sometimes… the real model was friendship all along.
Thats simply beautiful
", "rawMarkdown": "Thats simply beautiful" }, { "id": 3211045, "postDate": "2025-05-28T03:03:56.427Z", "rawMarkdown": "", "isDeleted": true }, { "id": 3211054, "postDate": "2025-05-28T03:17:27.300Z", "content": "“If I delete the write-up, will that comment disappear along with it?”
\n— Conor, I don't know where — 27 May 2025
I haven't read much, so don't lie to me.
", "rawMarkdown": "I haven't read much, so don't lie to me." }, { "id": 3211360, "postDate": "2025-05-28T10:43:56.687Z", "content": "I was under the impression that you were the hacker of the team @kenyeungtech!
", "rawMarkdown": "I was under the impression that you were the hacker of the team @kenyeungtech!" } ] } ] }, { "id": 3211019, "postDate": "2025-05-28T02:35:32.967Z", "content": "congratulations!!
", "rawMarkdown": "congratulations!!", "votes": 2 }, { "id": 3216167, "postDate": "2025-06-03T07:32:06.857Z", "content": "Great job!
", "rawMarkdown": "Great job!" }, { "id": 3214723, "postDate": "2025-06-01T02:12:08.297Z", "content": "Awesome 🫡
", "rawMarkdown": "Awesome 🫡" }, { "id": 3214249, "postDate": "2025-05-31T06:40:03.260Z", "content": "Congratulations…..
", "rawMarkdown": "Congratulations.....\n" }, { "id": 3214002, "postDate": "2025-05-30T18:24:46.757Z", "content": "Good idea!
", "rawMarkdown": "Good idea!" }, { "id": 3212375, "postDate": "2025-05-29T18:17:25.670Z", "rawMarkdown": "", "isDeleted": true } ], "complete": true, "source": "live", "fetched_at": "2026-10-20T09:08:24Z", "competition_id": 89659, "competition_slug": "drawing-with-llms", "team_id": 13406457, "solution_url": "/competitions/drawing-with-llms/writeups/new-team-name-1st-place-solution" }