![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F17096593%2F6702f89386eee00cc5a472dfcc64fd33%2Fa.png?generation=1748401229132058&alt=media\")
Congratulations to the winners and all participants for their efforts!
\nI want to give a huge shoutout to @richolson for his unbelievably educational and helpful notebook and discussion posts. I personally learned a lot from his notebook and it was the starting point for my solution.
\nEDIT: I published my version of the notebook! Feel free to check it out!
\n'a' for OCR mitigationFrom previous discussions, we know that adding a letter to the iamge can reduce OCR hallucinations. I added the letter 'a' to the bottom right corner of the image.
svg_a = \"\"\"\"\"\"\n
In a public notebook, VTracer was used for SVG conversion but in that pipeline, the bitmap was resized to a constant smaller size to reduce the SVG size (10kB restriction). One improvement to this is to use binary search to find the maximum size to convert the bitmap to SVG without exceeding the 10kB limit.
\nVTracer's output SVG looks like this:
\n\nCompression techniques:
\ntranslate.\nThis allows bitmaps to have be resized about 50px bigger than without compression. It does not affect images that are relatively SVG-friendly but it helps with images that are not.
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F17096593%2Ff70b7938bcbd9296c77a613e80f14565%2Fcompresion_1.png?generation=1748401292468823&alt=media\")
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F17096593%2Ffec70c2c25ddfa192f4b64dd8473b1fb%2Fcompression_2.png?generation=1748401301655450&alt=media\")
I used FLUX.1 Schnell for image generation and nunchaku for quantization. I tried base SD models, finetuning SDXL Lora but found that FLUX.1 Schnell has better prompt adherence. Also, with quantization, FLUX.1 Schnell is faster.
\nFrom 10-12 images generated, I evaluated the images using the questions below.
\nquestion = [\n ,\n \n]\nchoices = [\n [, , , ],\n [, , ]\n]\nanswer = [\n ,\n \n]\nIn addition, I performed a linear regression on fidelity VQA scores using dummy VQA scores from an extended training set to predict the final VQA score.
\n$$VQA_{predicted} = w_1 * VQA_{dummy_1} + w_2 * VQA_{dummy_2} + b$$
\nEvaluating an image per question takes < 2 seconds while generating an image with FLUX.1 Schnell takes ~ 1 second. Hence, I found that generating more images and evaluating them with 1-2 questions is more efficient than generating fewer images and evaluating them with 3-4 questions.
After reading @cnumber's post, I regret not trying pydiffvg to further optimize the scores and not trying Siglip. I also had moments where I wished I had participated in a team as it was difficult to come up with ideas solo. Regardless, I am happy with the knowledge and experience I gained from this competition!
PS: In future competitions, I hope to join a team to bounce ideas off each other so feel free to reach out if you happen to see me in a future competition!
", "comments": [ { "id": 3211148, "author_name": "WOOSUNG YOON", "author_url": "", "post_date": "2025-05-28T05:52:58.967000", "content": "Congratulations!
\nActually, after I saw your Kurapika profile,
\nI ended up spending a lot of time changing True/False answers for a VQA model.
Hahaha good one!
\nI'm just a huge Kurapika fan :D
And congratulations to you too!
", "votes": 4, "replies": [] } ] }, { "id": 3211359, "author_name": "Olowe Daniel", "author_url": "", "post_date": "2025-05-28T10:42:14.117000", "content": "Congratulations! Seems like such a succinct yet ingenious approach. I hope I can be this skilled someday.
", "votes": 1, "replies": [] }, { "id": 3211411, "author_name": "Geremie Yeo", "author_url": "", "post_date": "2025-05-28T11:48:31.230000", "content": "Congrats! It seems like we did the same thing, prompting the SD model to generate a cartoon-style image 😅
", "votes": 2, "replies": [ { "id": 3211454, "author_name": "Howard（ホワルド）", "author_url": "", "post_date": "2025-05-28T13:02:38.097000", "content": "Congrats to you too! Yeah, funny coincidence 😆
", "votes": 0, "replies": [] } ] }, { "id": 3211172, "author_name": "XDUnlper", "author_url": "", "post_date": "2025-05-28T06:41:45.303000", "content": "Congratulations! Could you share your code? I am very curious about the part \"Optimized VTracer's bitmap -> SVG conversion with binary search and SVG compression\" which you mentioned in your post and expect to learn more from your code. I failed to think about this optimization part. 😭
", "votes": 2, "replies": [ { "id": 3211232, "author_name": "Howard（ホワルド）", "author_url": "", "post_date": "2025-05-28T08:02:51.380000", "content": "Thanks and congratulations to you too!
\nI've published my notebook and the models used. Feel free to try it out yourself!
", "votes": 2, "replies": [ { "id": 3211416, "author_name": "Xueqi Ding", "author_url": "", "post_date": "2025-05-28T12:04:03.850000", "content": "Thank you so much ❤️
", "votes": 2, "replies": [] } ] } ] }, { "id": 3211051, "author_name": "c-number", "author_url": "", "post_date": "2025-05-28T03:12:56.233000", "content": "Congratulations!
\nI'm surprised to see that our solutions are quite similar; vtracer with binary search, flux.1 schnell etc. .
\nLooking forward to competing with you in another competition!
Congratulations to you too!
\nI thought the same when I read your team's solution. To be fair, I do think tweaks such as binary search, using flux.1 schnell are straightforward, so I'm expecting many top solutions to do the same
\nSee you in future competitions!
Impressive work! Your method is both sharp and efficient. I aspire to reach this level of expertise someday.
", "votes": 0, "replies": [] } ], "comment_count": 11, "messages": [ { "id": 3211046, "postDate": "2025-05-28T03:04:30.927Z", "content": "Congratulations to the winners and all participants for their efforts!
\nI want to give a huge shoutout to @richolson for his unbelievably educational and helpful notebook and discussion posts. I personally learned a lot from his notebook and it was the starting point for my solution.
\nEDIT: I published my version of the notebook! Feel free to check it out!
\n'a' for OCR mitigationFrom previous discussions, we know that adding a letter to the iamge can reduce OCR hallucinations. I added the letter 'a' to the bottom right corner of the image.
svg_a = \"\"\"\"\"\"\n
In a public notebook, VTracer was used for SVG conversion but in that pipeline, the bitmap was resized to a constant smaller size to reduce the SVG size (10kB restriction). One improvement to this is to use binary search to find the maximum size to convert the bitmap to SVG without exceeding the 10kB limit.
\nVTracer's output SVG looks like this:
\n\nCompression techniques:
\ntranslate.\nThis allows bitmaps to have be resized about 50px bigger than without compression. It does not affect images that are relatively SVG-friendly but it helps with images that are not.
\n
I used FLUX.1 Schnell for image generation and nunchaku for quantization. I tried base SD models, finetuning SDXL Lora but found that FLUX.1 Schnell has better prompt adherence. Also, with quantization, FLUX.1 Schnell is faster.
\nFrom 10-12 images generated, I evaluated the images using the questions below.
\nquestion = [\n ,\n \n]\nchoices = [\n [, , , ],\n [, , ]\n]\nanswer = [\n ,\n \n]\nIn addition, I performed a linear regression on fidelity VQA scores using dummy VQA scores from an extended training set to predict the final VQA score.
\n$$VQA_{predicted} = w_1 * VQA_{dummy_1} + w_2 * VQA_{dummy_2} + b$$
\nEvaluating an image per question takes < 2 seconds while generating an image with FLUX.1 Schnell takes ~ 1 second. Hence, I found that generating more images and evaluating them with 1-2 questions is more efficient than generating fewer images and evaluating them with 3-4 questions.
After reading @cnumber's post, I regret not trying pydiffvg to further optimize the scores and not trying Siglip. I also had moments where I wished I had participated in a team as it was difficult to come up with ideas solo. Regardless, I am happy with the knowledge and experience I gained from this competition!
PS: In future competitions, I hope to join a team to bounce ideas off each other so feel free to reach out if you happen to see me in a future competition!
", "rawMarkdown": "Congratulations to the winners and all participants for their efforts!\n\nI want to give a huge shoutout to @richolson for his unbelievably educational and helpful notebook and discussion posts. I personally learned a lot from [his notebook](https://www.kaggle.com/code/richolson/stable-diffusion-svg-scoring-metric) and it was the starting point for my solution.\n\nEDIT: I published my version of the [notebook](https://www.kaggle.com/code/howarudo/13th-place-solution-flux-svg-scoring)! Feel free to check it out!\n\n# Summary\n- `'a'` for OCR mitigation\n- Optimized VTracer's bitmap -> SVG conversion with binary search and SVG compression\n- Used FLUX.1 Schnell and [nunchaku quantization](https://github.com/mit-han-lab/nunchaku)\n- VQA with dummy question and answer\n\n# OCR mitigation\nFrom previous discussions, we know that adding a letter to the iamge can reduce OCR hallucinations. I added the letter `'a'` to the bottom right corner of the image.\n\n```svg\nsvg_a = \"\"\"Congratulations!
\nActually, after I saw your Kurapika profile,
\nI ended up spending a lot of time changing True/False answers for a VQA model.
Hahaha good one!
\nI'm just a huge Kurapika fan :D
And congratulations to you too!
", "rawMarkdown": "Hahaha good one!\nI'm just a huge Kurapika fan :D\n\nAnd congratulations to you too!", "votes": 4 } ] }, { "id": 3211359, "postDate": "2025-05-28T10:42:14.117Z", "content": "Congratulations! Seems like such a succinct yet ingenious approach. I hope I can be this skilled someday.
", "rawMarkdown": "Congratulations! Seems like such a succinct yet ingenious approach. I hope I can be this skilled someday.", "votes": 1 }, { "id": 3211411, "postDate": "2025-05-28T11:48:31.230Z", "content": "Congrats! It seems like we did the same thing, prompting the SD model to generate a cartoon-style image 😅
", "rawMarkdown": "Congrats! It seems like we did the same thing, prompting the SD model to generate a cartoon-style image 😅", "votes": 2, "replies": [ { "id": 3211454, "postDate": "2025-05-28T13:02:38.097Z", "content": "Congrats to you too! Yeah, funny coincidence 😆
", "rawMarkdown": "Congrats to you too! Yeah, funny coincidence 😆" } ] }, { "id": 3211172, "postDate": "2025-05-28T06:41:45.303Z", "content": "Congratulations! Could you share your code? I am very curious about the part \"Optimized VTracer's bitmap -> SVG conversion with binary search and SVG compression\" which you mentioned in your post and expect to learn more from your code. I failed to think about this optimization part. 😭
", "rawMarkdown": "Congratulations! Could you share your code? I am very curious about the part \"Optimized VTracer's bitmap -> SVG conversion with binary search and SVG compression\" which you mentioned in your post and expect to learn more from your code. I failed to think about this optimization part. 😭", "votes": 2, "replies": [ { "id": 3211232, "postDate": "2025-05-28T08:02:51.380Z", "content": "Thanks and congratulations to you too!
\nI've published my notebook and the models used. Feel free to try it out yourself!
", "rawMarkdown": "Thanks and congratulations to you too!\n\nI've published my [notebook](https://www.kaggle.com/code/howarudo/13th-place-solution-flux-svg-scoring) and the models used. Feel free to try it out yourself!", "votes": 2, "replies": [ { "id": 3211416, "postDate": "2025-05-28T12:04:03.850Z", "content": "Thank you so much ❤️
", "rawMarkdown": "Thank you so much ❤️ ", "votes": 2 } ] } ] }, { "id": 3211051, "postDate": "2025-05-28T03:12:56.233Z", "content": "Congratulations!
\nI'm surprised to see that our solutions are quite similar; vtracer with binary search, flux.1 schnell etc. .
\nLooking forward to competing with you in another competition!
Congratulations to you too!
\nI thought the same when I read your team's solution. To be fair, I do think tweaks such as binary search, using flux.1 schnell are straightforward, so I'm expecting many top solutions to do the same
\nSee you in future competitions!
Impressive work! Your method is both sharp and efficient. I aspire to reach this level of expertise someday.
", "rawMarkdown": "Impressive work! Your method is both sharp and efficient. I aspire to reach this level of expertise someday." } ], "complete": true, "source": "live", "fetched_at": "2026-10-20T09:08:42Z", "competition_id": 89659, "competition_slug": "drawing-with-llms", "team_id": 13608330, "solution_url": "/competitions/drawing-with-llms/writeups/howard-13th-place-solution-a-kaggle-beginner-s-att" }