![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F5256933%2F5d143ee1a6e3da4d04cd99c52a51a880%2FScreenshot%202025-06-06%20200646.jpg?generation=1749211619603421&alt=media\")
First of all, we would like to thank the hosts for their continued effort to improve the metric to ensure that the top solutions aligns with the objectives of the competition!
\nTeam members:
\nCode: https://github.com/bogoconic1/16th-place-kaggle-drawing-with-llms
\n
In summary, we allocated 36 seconds for SVG generation and aesthetic scores computation, and the remaining 27 seconds for candidate selection.
\nWe tried lots of open-sourced SD models (such as Sana, SD 3.5) etc and stuck with SDXL-Flash as it was the fastest model which could generate \"nice-looking\" images that met our expectations.
\nUsing the default settings (1024x1024), SDXL-Flash took 8 seconds to generate an image. However, we improved the speed by up till 80% by generating images with lower resolution. For 512x512, we were able to generate 18 candidate images within 30 seconds, without compromising much quality. It was validated locally that 512x512 or 640x640 gave the best tradeoff between the speed and accuracy. On the LB, 704x704 did pretty well too but we didn’t select it due to the low validation score.
\nAdditionally, we used the prompt f'Cartoon style of \"{description}\".' as we noticed that the cover page of the DOODL paper showed an image that appeared to have a cartoon style after aesthetic optimization. When examining the SVG outputs, it was found that there is less quality loss in the conversion process, compared to a normal image.
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F5256933%2Fa8b2ac2422555b96ea8c5c3b6c9beaf0%2FIMG_8871.jpeg?generation=1748425985094002&alt=media\")
The bitmap_to_svg_layered technique shared publicly took around 4.5 seconds to run per image, and has significant information loss when compressed to 10k bytes.
\nWhen doing function profiling, we noticed that cv2's KMeans was the main culprit. Therefore, we sped it up using NVIDIA's cuML KMeans on GPU, with that, each image was converted to SVG in ~0.3 seconds (15x speedup).
\nAdditionally to squeeze in more elements into 10k bytes, we embedded the following techniques.
\n<polygon> into <path><g> tag)This allowed us to fit in 60 -> 168 elements (on average) for an improvement of 280% without any quality loss compared to the publicly shared function
\nThe OCR decoy by @richolson was used to mitigate hallucinations. We did not make any changes since a failure rate of 1/150 is deemed acceptable
\nTo quicken the aesthetic computation locally, we sped up apply_median_filter using cv2 and apply_fft_low_pass in ImageProcessor using NVIDIA CuPy (0.5 -> 0.3 s/image)
\n (metric.ImageProcessor):\n\n ():\n ().__init__(image, seed)\n\n ():\n \n \n img_array = np.asarray(.image)\n \n size % == :\n size += \n filtered = cv2.medianBlur(img_array, ksize=size)\n .image = Image.fromarray(filtered)\n \n\n ():\n x = cp.asarray(.image, dtype=cp.float32) \n f = cp.fft.fftshift(cp.fft.fft2(x, axes=(,)), axes=(,))\n\n rows, cols = x.shape[:]\n crow, ccol = rows // , cols // \n r = ((crow, ccol) * cutoff_frequency)\n\n y, xg = cp.ogrid[:rows, :cols]\n mask = ((y - crow) ** + (xg - ccol) ** ) <= r * r\n f *= mask[..., ] \n\n img_back = cp.fft.ifft2(cp.fft.ifftshift(f, axes=(,)), axes=(,)).real\n img_back = cp.clip(img_back, , ).astype(cp.uint8).get()\n .image = Image.fromarray(img_back)\n \nWe estimated the VQA score by asking one question: 'Does <image> portray \"{}\"? Answer yes or no.'
Given that
\n| Method | \nTime Taken/image | \n
|---|---|
| Aesthetic Score | \n0.3s | \n
| VQA + OCR Score | \n3s | \n
Instead of estimating VQA of all images, we employed a reranker approach to estimate which images are more likely to get a higher score on competition metric, which allowed us to generate more candidate SVGs within the time limit.
\nOur solution has been deployed and is accessible via this URL: https://geremieyeo--text-to-svg-generator-run.modal.run//
\nCode: https://github.com/bogoconic1/16th-place-kaggle-drawing-with-llms
\nThe application uses a 15-minute idle timeout. If no requests are made within 15 minutes, the container automatically shuts down to save resources. When you visit the URL after a period of inactivity, expect approximately 1 minute of waiting time for the container to restart and load.
\nThe rendering of the bitmap and SVG on the UI takes about 25 seconds after clicking \"Generate SVG\"
", "votes": 1, "replies": [] }, { "id": 3214328, "author_name": "BeaKaggle", "author_url": "", "post_date": "2025-05-31T09:34:02.483000", "content": "Hi @yeoyunsianggeremie Thank you for sharing. N,ce work. I was looking at your notebook https://www.kaggle.com/code/yeoyunsianggeremie/fork-of-16th-place-solution-sd-reranker-prox Question: can you make the PRIVATE inputs at your notebook public, so that we can try? Thanks
", "votes": 0, "replies": [ { "id": 3214821, "author_name": "Geremie Yeo", "author_url": "", "post_date": "2025-06-01T06:08:55.203000", "content": "done 😀
\nfeel free to use the original notebook here if the forked one has issues
First of all, we would like to thank the hosts for their continued effort to improve the metric to ensure that the top solutions aligns with the objectives of the competition!
\nTeam members:
\nCode: https://github.com/bogoconic1/16th-place-kaggle-drawing-with-llms
\n
In summary, we allocated 36 seconds for SVG generation and aesthetic scores computation, and the remaining 27 seconds for candidate selection.
\nWe tried lots of open-sourced SD models (such as Sana, SD 3.5) etc and stuck with SDXL-Flash as it was the fastest model which could generate \"nice-looking\" images that met our expectations.
\nUsing the default settings (1024x1024), SDXL-Flash took 8 seconds to generate an image. However, we improved the speed by up till 80% by generating images with lower resolution. For 512x512, we were able to generate 18 candidate images within 30 seconds, without compromising much quality. It was validated locally that 512x512 or 640x640 gave the best tradeoff between the speed and accuracy. On the LB, 704x704 did pretty well too but we didn’t select it due to the low validation score.
\nAdditionally, we used the prompt f'Cartoon style of \"{description}\".' as we noticed that the cover page of the DOODL paper showed an image that appeared to have a cartoon style after aesthetic optimization. When examining the SVG outputs, it was found that there is less quality loss in the conversion process, compared to a normal image.
The bitmap_to_svg_layered technique shared publicly took around 4.5 seconds to run per image, and has significant information loss when compressed to 10k bytes.
\nWhen doing function profiling, we noticed that cv2's KMeans was the main culprit. Therefore, we sped it up using NVIDIA's cuML KMeans on GPU, with that, each image was converted to SVG in ~0.3 seconds (15x speedup).
\nAdditionally to squeeze in more elements into 10k bytes, we embedded the following techniques.
\n<polygon> into <path><g> tag)This allowed us to fit in 60 -> 168 elements (on average) for an improvement of 280% without any quality loss compared to the publicly shared function
\nThe OCR decoy by @richolson was used to mitigate hallucinations. We did not make any changes since a failure rate of 1/150 is deemed acceptable
\nTo quicken the aesthetic computation locally, we sped up apply_median_filter using cv2 and apply_fft_low_pass in ImageProcessor using NVIDIA CuPy (0.5 -> 0.3 s/image)
\n (metric.ImageProcessor):\n\n ():\n ().__init__(image, seed)\n\n ():\n \n \n img_array = np.asarray(.image)\n \n size % == :\n size += \n filtered = cv2.medianBlur(img_array, ksize=size)\n .image = Image.fromarray(filtered)\n \n\n ():\n x = cp.asarray(.image, dtype=cp.float32) \n f = cp.fft.fftshift(cp.fft.fft2(x, axes=(,)), axes=(,))\n\n rows, cols = x.shape[:]\n crow, ccol = rows // , cols // \n r = ((crow, ccol) * cutoff_frequency)\n\n y, xg = cp.ogrid[:rows, :cols]\n mask = ((y - crow) ** + (xg - ccol) ** ) <= r * r\n f *= mask[..., ] \n\n img_back = cp.fft.ifft2(cp.fft.ifftshift(f, axes=(,)), axes=(,)).real\n img_back = cp.clip(img_back, , ).astype(cp.uint8).get()\n .image = Image.fromarray(img_back)\n \nWe estimated the VQA score by asking one question: 'Does <image> portray \"{}\"? Answer yes or no.'
Given that
\n| Method | \nTime Taken/image | \n
|---|---|
| Aesthetic Score | \n0.3s | \n
| VQA + OCR Score | \n3s | \n
Instead of estimating VQA of all images, we employed a reranker approach to estimate which images are more likely to get a higher score on competition metric, which allowed us to generate more candidate SVGs within the time limit.
\nOur solution has been deployed and is accessible via this URL: https://geremieyeo--text-to-svg-generator-run.modal.run//
\nCode: https://github.com/bogoconic1/16th-place-kaggle-drawing-with-llms
\nThe application uses a 15-minute idle timeout. If no requests are made within 15 minutes, the container automatically shuts down to save resources. When you visit the URL after a period of inactivity, expect approximately 1 minute of waiting time for the container to restart and load.
\nThe rendering of the bitmap and SVG on the UI takes about 25 seconds after clicking \"Generate SVG\"
", "rawMarkdown": "Update\n--------------------\n\nOur solution has been deployed and is accessible via this URL: https://geremieyeo--text-to-svg-generator-run.modal.run//\n\nCode: https://github.com/bogoconic1/16th-place-kaggle-drawing-with-llms\n\nNote\n--------------------\n\nThe application uses a 15-minute idle timeout. If no requests are made within 15 minutes, the container automatically shuts down to save resources. When you visit the URL after a period of inactivity, expect approximately 1 minute of waiting time for the container to restart and load.\n\nThe rendering of the bitmap and SVG on the UI takes about 25 seconds after clicking \"Generate SVG\"", "votes": 1 }, { "id": 3214328, "postDate": "2025-05-31T09:34:02.483Z", "content": "Hi @yeoyunsianggeremie Thank you for sharing. N,ce work. I was looking at your notebook https://www.kaggle.com/code/yeoyunsianggeremie/fork-of-16th-place-solution-sd-reranker-prox Question: can you make the PRIVATE inputs at your notebook public, so that we can try? Thanks
", "rawMarkdown": "Hi @yeoyunsianggeremie Thank you for sharing. N,ce work. I was looking at your notebook https://www.kaggle.com/code/yeoyunsianggeremie/fork-of-16th-place-solution-sd-reranker-prox Question: can you make the PRIVATE inputs at your notebook public, so that we can try? Thanks", "replies": [ { "id": 3214821, "postDate": "2025-06-01T06:08:55.203Z", "content": "done 😀
\nfeel free to use the original notebook here if the forked one has issues