\n
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F6624777%2Fe01716185fc3f07178f828ab179a8b75%2Foriginal%20(3).png?generation=1748396425784049&alt=media\")
We ( @cnumber and @daiwakun ) would like to thank the hosts who continually worked hard to improve the metric, making this competition interesting and meaningful. We learned a lot about image generation, which we knew little before entering this competition.
\n[Final Submission]
\nDescription of the solution is below
Original Image generated by FLUX.1-schnell
\n
Original SVG image generated by vtracer
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F6624777%2F171f68fb81c4bae916de373743e063de%2Fsvg_before%20(2).png?generation=1748396445969721&alt=media\")
SVG image after iterative optimization
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F6624777%2F3d3210e5cf6c296f3b0f6e7455bda3a2%2Fsvg_after%20(2).png?generation=1748396464177345&alt=media\")
SVG image after iterative optimization, processed with ImageProcessor (final image used for VQA evaluation)
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F6624777%2F94cf6fe4ab3e816b1a2a2eb952c473d8%2Fafterproces.png?generation=1748396499775453&alt=media\")
We are excited to share our approach that secured 3rd place in the \"Drawing with LLMs\" competition. This competition challenged us to generate Scalable Vector Graphics (SVG) from text prompts, adhering to strict constraints on size and content, while maximizing visual fidelity and aesthetic appeal. Our solution is a multi-stage pipeline that combines the strengths of powerful generative models, vectorization tools, and differentiable graphics rendering. We’re proud to be able to present our solution that DOES NOT hack the VQA evaluator using text embedding!
\nOur core idea was to first generate a high-quality bitmap image that captures the essence of the prompt, convert it into an initial SVG, and then iteratively refine this SVG to improve its faithfulness to the prompt, its aesthetic quality, and its adherence to the competition's VQA metrics, all while respecting the output constraints.
\nHere's a breakdown of the stages:
\nBitmap Generation (Flux):
\nflux.1-schnell) to generate candidate bitmap images.'{description}' (the raw prompt)'Beautiful illustration of \"{description}\"''Simple svg illustration of \"{description}\", with colorful background, beautiful background.''Simple svg illustration of \"{description}\", with background of minimal flat vector illustration of \"colorful layered big mountains\"'Initial SVG Conversion (vtracer):
vtracer. We found vtracer to be effective in producing clean polygonal vector graphics.Candidate Selection & Pre-Optimization:
\ngoogle/siglip-so400m-patch14-384) to select the images most similar to the original text prompt; the same model used as the image encoder the vqa scorer PaliGemma2. We selected the top 2 candidates from the initial 8.iterations = 10) using our pydiffvg-based optimizer (details below).Main SVG Optimization (pydiffvg & Custom Losses):
pydiffvg.google/siglip2-base-patch16-224) embedding of the rendered SVG (after processing with a differentiable ImageProcessor detailed below) and the SigLIP embedding of the target png image.ImageProcessorTorch) and the target bitmap image (after processing with a reference ImageProcessor with the same random seed for cropping). This helps maintain structural and color fidelity to the initial high-quality bitmap`.iterations = 200.ImageProcessorTorch - Differentiable Preprocessing:
ImageProcessorTorch. This PyTorch module applies a sequence of image transformations similar to those described in the evaluation details.diff_jpeg), median filter, FFT low-pass filter, and bilateral filter.Final SVG Compression & Formatting:
\npydiffvg output / vtracer input that, after compression, would fit within the 10KB limit. If necessary, paths were aggressively pruned.<path d=\"M167 200h20l-20 30h20\"...).The most beautiful solution. I couldn't get diffvg to generate stable solutions in time locally, so I dropped this approach—glad to see you could make it work.
", "votes": 5, "replies": [] }, { "id": 3211248, "author_name": "Mehara Rothila", "author_url": "", "post_date": "2025-05-28T08:12:07.557000", "content": "Really fascinating approach! I'm curious about your ImageProcessorTorch implementation - could you share more details about how you made the JPEG compression, median filter, and bilateral filter operations differentiable? Did you use straight-through estimators for all the non-differentiable operations, or did you find differentiable approximations for some of them? Also, how significant was the performance impact of this differentiable preprocessing compared to using the actual evaluation pipeline?
", "votes": 1, "replies": [ { "id": 3211315, "author_name": "c-number", "author_url": "", "post_date": "2025-05-28T09:50:42.103000", "content": "They are implemented in the ImageProcessorTorch class in our submission.
\nWe made sure that each process was in consistency with the original implementation, making the difference minimal.
\nAt maximum, we only had an AES score difference of around 0.15.
Thank you for sharing your team’s solution. Impressive idea!
", "votes": 1, "replies": [] }, { "id": 3211042, "author_name": "Ochir Dorzhiev", "author_url": "", "post_date": "2025-05-28T03:01:23.833000", "content": "Huge congratulations on 3rd place — truly impressive ideas! I was wondering if you could estimate how much the Main SVG Optimization step contributed to the score?
", "votes": 1, "replies": [ { "id": 3211319, "author_name": "c-number", "author_url": "", "post_date": "2025-05-28T09:51:29.457000", "content": "Without the main svg optimization, we had an aes score less than 0.5, so the optimization really boosted our score.
", "votes": 0, "replies": [] } ] }, { "id": 3211021, "author_name": "NOBITA_7", "author_url": "", "post_date": "2025-05-28T02:36:55.193000", "content": "Thank you for sharing your team’s solution. Congratulations!
", "votes": 1, "replies": [] }, { "id": 3211018, "author_name": "guo dashuai", "author_url": "", "post_date": "2025-05-28T02:35:12.843000", "content": "Congratulations to you! I'm impressed that you didn't use OCR-based attacks. Some elements in our attacks couldn't be adapted to DiffVG, and we also found that DiffVG tends not to converge when the complexity of the trained elements becomes too high (this was observed in Metric 2). Did you encounter a similar issue?
", "votes": 1, "replies": [ { "id": 3211033, "author_name": "c-number", "author_url": "", "post_date": "2025-05-28T02:52:16.033000", "content": "Congrats to your non-drawing solution! We didn't expect paligemma2 to be able to extract information from embedded texts when it couldn't detect them with OCR.
\n\n\nthe complexity of the trained elements becomes too high
\n
Could you explain this in more detail? Does it mean that there are many objects defined in the svg file?
", "votes": 4, "replies": [ { "id": 3211041, "author_name": "guo dashuai", "author_url": "", "post_date": "2025-05-28T02:57:56.077000", "content": "Too many elements refer to the fact that I perform backpropagation on the parameters of first-, second-, and third-order Bézier curves in the path (of course with some constraints) in metric2, but I notice the loss fluctuates up and down. However, if I fix the shape and only backpropagate on the color, the aesthetic score in metric2 can exceed 1.1 points.
", "votes": 0, "replies": [] }, { "id": 3211044, "author_name": "c-number", "author_url": "", "post_date": "2025-05-28T03:03:39.743000", "content": "We decided to join this compeititon after the metric was fixed as it is now, so we haven't had time to play around with metric 2.
\nAs for the current metric, we didn't use Bézier curves but only straight lines in the path, because we didn't want to make things too complicated, and also because we didn't think it would matter much after smoothing the image with ImageProcessor.apply().
\nMaybe that was the reason why our optimization worked.
Congratulations! It's great to see such a cool solution achieve high scores.
", "votes": 1, "replies": [ { "id": 3211034, "author_name": "c-number", "author_url": "", "post_date": "2025-05-28T02:53:12.570000", "content": "Congrats to you too! It's impressive that a text-based attack and an image genration solution can be merged to result in such as high scoring solution.
", "votes": 1, "replies": [ { "id": 3211047, "author_name": "Filtered", "author_url": "", "post_date": "2025-05-28T03:07:25.603000", "content": "I believe your images combined with text-based attacks will get higher scores, and that may be the final solution to this game
", "votes": 1, "replies": [] } ] } ] }, { "id": 3211012, "author_name": "Khoi Nguyen", "author_url": "", "post_date": "2025-05-28T02:28:23.117000", "content": "This is actually a really nice and general solution to such a niche problem, congrats.
", "votes": 1, "replies": [ { "id": 3211036, "author_name": "c-number", "author_url": "", "post_date": "2025-05-28T02:54:22.340000", "content": "Thanks! We are quite happy to be able to secure our ranking without using any model-specific attacks!
", "votes": 0, "replies": [] } ] }, { "id": 3210992, "author_name": "Tâm Anh Nguyễn Duy", "author_url": "", "post_date": "2025-05-28T02:02:01.387000", "content": "Can I ask what your average aesthetic score was, and how you improved that score in the competition? Thank you.
", "votes": 1, "replies": [ { "id": 3211000, "author_name": "c-number", "author_url": "", "post_date": "2025-05-28T02:10:49.197000", "content": "Our VQA/AES scores were 0.81/0.64.
\nThanks for asking! (after reading your comment, we added them in the title)
this is what i call honest smart solution with lot of hard work behind, cheating should never be considered smart.. congratulations! just a short question: what was the average time for prediction?
", "votes": 2, "replies": [ { "id": 3211311, "author_name": "c-number", "author_url": "", "post_date": "2025-05-28T09:47:24.410000", "content": "We used around 63 seconds per image, which is just below the maximum time budget
", "votes": 2, "replies": [] } ] }, { "id": 3211030, "author_name": "Casualty", "author_url": "", "post_date": "2025-05-28T02:47:20.147000", "content": "Your solution is highly versatile. After reviewing your work, everything clicked into place. My sincere congratulations to you!
", "votes": 2, "replies": [] }, { "id": 3210990, "author_name": "WOOSUNG YOON", "author_url": "", "post_date": "2025-05-28T01:59:28.130000", "content": "I was truly impressed by your work. 😄
\nYou managed to create an excellent solution using tools like FLUX.1-schnell, diffvg, and diffjpeg — tools I had personally thought would be difficult to run.
Your approach has given me new insight and has been a great source of inspiration.
\nThank you for the opportunity to learn so much from your work.
I asked cursor to write a Differentiable Preprocessing to mimic the ImageProcessor, but didn't work😅
Thare are libraries that can handle pytorch-based bilateral filtering and jpeg compression.
\nOther processes can easily be implemented using vanilla pytorch.
This is actually a really nice and general solution to such a niche problem, congrats.
", "votes": -1, "replies": [] }, { "id": 3212039, "author_name": "Sarah Arshad", "author_url": "", "post_date": "2025-05-29T08:24:34.270000", "content": "A huge thank uh to the organizers for creating such a thoughtful challenge!
", "votes": -1, "replies": [] }, { "id": 3217806, "author_name": "BeaKaggle", "author_url": "", "post_date": "2025-06-05T12:19:01.910000", "content": "Very nice work @cnumber Congrats!
\nTo be able to run your code, can you make private inputs like kagglehub.model_download(\"cnumber/nunchaku/pyTorch/default\") public? Thank you!
I took a different tack by leveraging a lightweight GAN to generate preliminary vector‐style sketches, then used a custom CLIP‐guided evolutionary optimizer to iteratively mutate and select SVG primitives under the 10 KB limit. Instead of pydiffvg, I implemented a pure PyTorch renderer with differentiable anti‐aliasing, which cut GPU memory usage by 40 %. I found that focusing on Bézier stitch simplifications—merging overlapping curves dynamically—boosted my AES score by 0.07. Finally, I used a small ensemble of CLIP‐ranked candidates before a final pass of SVG pruning, which kept fidelity high while staying within size constraints.
", "votes": 0, "replies": [] }, { "id": 3214160, "author_name": "Waylon Wu", "author_url": "", "post_date": "2025-05-31T03:47:49.917000", "content": "Congrats! That's a really smart solution. Would you make the pydiffvg package in your notebook public? I believe most of us cannot get it work xD.
", "votes": 0, "replies": [ { "id": 3215470, "author_name": "c-number", "author_url": "", "post_date": "2025-06-02T09:31:24.243000", "content": "Hope you like it!
\nhttps://www.kaggle.com/models/cnumber/pydiffvg/
Wonderful solution, especially because it follows the policy and
\n DOES NOT hack the VQA evaluator using text embedding!
amazing.
\nImageProcessorTorch part, what's the input to this, is it generated svg code in text?ImageProcessorTorch module impelemented in our submission notebook. It's just a differentiable version of the original ImageProcessor.Thanks for the clarification! I had a look at your ImageProcessorTorch module — nicely done making it differentiable! That’s a smart move to keep the pipeline end-to-end trainable.
\nAlso, I really like your approach of balancing aesthetics and fidelity by combining MSE loss (for image similarity) and the SigLIP score (for alignment with the description). It’s a thoughtful way to preserve content while improving stylization. Great work!
", "votes": 0, "replies": [] } ] } ] }, { "id": 3212222, "author_name": "Marpini Pavan Srikar", "author_url": "", "post_date": "2025-05-29T14:12:49.243000", "content": "Thank you for sharing your teams idea and a beautiful solution.
", "votes": 0, "replies": [] }, { "id": 3212115, "author_name": "Daniel Gallas", "author_url": "", "post_date": "2025-05-29T10:20:33.340000", "content": "This is my favourite solution from the ones I read so far. Congratulations. This is inspirational stuff
", "votes": 0, "replies": [] }, { "id": 3211398, "author_name": "JP", "author_url": "", "post_date": "2025-05-28T11:33:49.923000", "content": "Congratulations @cnumber and @daiwakun! I'm still trying to wrap my head around how you could manage to run all this without exceeding the time limit! Very impressive!
", "votes": 0, "replies": [] }, { "id": 3210991, "author_name": "william.wu", "author_url": "", "post_date": "2025-05-28T02:00:50.900000", "content": "Congratulations on your impressive solution! Thank you for sharing it. I'm curious: how did you prevent diffvg from over-optimizing the colors, which can reduce their impact after running through ImageProcessor? I also experimented with diffvg, but I gave up when I noticed that the improvements in score reverted after applying ImageProcessor.
sorry, just found you already included that trick. I missed this part:
\nImageProcessorTorch - Differentiable Preprocessing:\n\n bridge the gap between our optimization process the competition\nCrucially, these operations were implemented be (mostly) differentiable use straight-through estimators necessary, allowing gradients flow back the SVG parameters.\nTransformations included: random crop & resize, JPEG compression (simulated via diff_jpeg), median filter, FFT low-pass filter, bilateral filter.\nThe rendered SVG was passed through this processor before loss calculation. The target bitmap the MSE loss was also processed the random crop & resize part this ( a synchronized random seed) ensure fair comparison.\nYou can also have a look at our ImageProcessorTorch class implemented in our solution!
", "votes": 1, "replies": [] } ] }, { "id": 3218302, "author_name": "", "author_url": "", "post_date": "2025-06-06T04:22:06.173000", "content": "", "votes": 0, "replies": [] }, { "id": 3213846, "author_name": "Meobius", "author_url": "", "post_date": "2025-05-30T14:51:22.177000", "content": "Amazing！Thanks for sharing！
", "votes": 0, "replies": [] } ], "comment_count": 43, "messages": [ { "id": 3210978, "postDate": "2025-05-28T01:46:34.013Z", "content": "We ( @cnumber and @daiwakun ) would like to thank the hosts who continually worked hard to improve the metric, making this competition interesting and meaningful. We learned a lot about image generation, which we knew little before entering this competition.
\n[Final Submission]
\nDescription of the solution is below
Original Image generated by FLUX.1-schnell
\n
Original SVG image generated by vtracer
\n
SVG image after iterative optimization
\n
SVG image after iterative optimization, processed with ImageProcessor (final image used for VQA evaluation)
\n
We are excited to share our approach that secured 3rd place in the \"Drawing with LLMs\" competition. This competition challenged us to generate Scalable Vector Graphics (SVG) from text prompts, adhering to strict constraints on size and content, while maximizing visual fidelity and aesthetic appeal. Our solution is a multi-stage pipeline that combines the strengths of powerful generative models, vectorization tools, and differentiable graphics rendering. We’re proud to be able to present our solution that DOES NOT hack the VQA evaluator using text embedding!
\nOur core idea was to first generate a high-quality bitmap image that captures the essence of the prompt, convert it into an initial SVG, and then iteratively refine this SVG to improve its faithfulness to the prompt, its aesthetic quality, and its adherence to the competition's VQA metrics, all while respecting the output constraints.
\nHere's a breakdown of the stages:
\nBitmap Generation (Flux):
\nflux.1-schnell) to generate candidate bitmap images.'{description}' (the raw prompt)'Beautiful illustration of \"{description}\"''Simple svg illustration of \"{description}\", with colorful background, beautiful background.''Simple svg illustration of \"{description}\", with background of minimal flat vector illustration of \"colorful layered big mountains\"'Initial SVG Conversion (vtracer):
vtracer. We found vtracer to be effective in producing clean polygonal vector graphics.Candidate Selection & Pre-Optimization:
\ngoogle/siglip-so400m-patch14-384) to select the images most similar to the original text prompt; the same model used as the image encoder the vqa scorer PaliGemma2. We selected the top 2 candidates from the initial 8.iterations = 10) using our pydiffvg-based optimizer (details below).Main SVG Optimization (pydiffvg & Custom Losses):
pydiffvg.google/siglip2-base-patch16-224) embedding of the rendered SVG (after processing with a differentiable ImageProcessor detailed below) and the SigLIP embedding of the target png image.ImageProcessorTorch) and the target bitmap image (after processing with a reference ImageProcessor with the same random seed for cropping). This helps maintain structural and color fidelity to the initial high-quality bitmap`.iterations = 200.ImageProcessorTorch - Differentiable Preprocessing:
ImageProcessorTorch. This PyTorch module applies a sequence of image transformations similar to those described in the evaluation details.diff_jpeg), median filter, FFT low-pass filter, and bilateral filter.Final SVG Compression & Formatting:
\npydiffvg output / vtracer input that, after compression, would fit within the 10KB limit. If necessary, paths were aggressively pruned.<path d=\"M167 200h20l-20 30h20\"...).Really fascinating approach! I'm curious about your ImageProcessorTorch implementation - could you share more details about how you made the JPEG compression, median filter, and bilateral filter operations differentiable? Did you use straight-through estimators for all the non-differentiable operations, or did you find differentiable approximations for some of them? Also, how significant was the performance impact of this differentiable preprocessing compared to using the actual evaluation pipeline?
", "rawMarkdown": "Really fascinating approach! I'm curious about your ImageProcessorTorch implementation - could you share more details about how you made the JPEG compression, median filter, and bilateral filter operations differentiable? Did you use straight-through estimators for all the non-differentiable operations, or did you find differentiable approximations for some of them? Also, how significant was the performance impact of this differentiable preprocessing compared to using the actual evaluation pipeline?", "votes": 1, "replies": [ { "id": 3211315, "postDate": "2025-05-28T09:50:42.103Z", "content": "They are implemented in the ImageProcessorTorch class in our submission.
\nWe made sure that each process was in consistency with the original implementation, making the difference minimal.
\nAt maximum, we only had an AES score difference of around 0.15.
Thank you for sharing your team’s solution. Impressive idea!
", "rawMarkdown": "Thank you for sharing your team’s solution. Impressive idea!", "votes": 1 }, { "id": 3211042, "postDate": "2025-05-28T03:01:23.833Z", "content": "Huge congratulations on 3rd place — truly impressive ideas! I was wondering if you could estimate how much the Main SVG Optimization step contributed to the score?
", "rawMarkdown": "Huge congratulations on 3rd place — truly impressive ideas! I was wondering if you could estimate how much the Main SVG Optimization step contributed to the score?", "votes": 1, "replies": [ { "id": 3211319, "postDate": "2025-05-28T09:51:29.457Z", "content": "Without the main svg optimization, we had an aes score less than 0.5, so the optimization really boosted our score.
", "rawMarkdown": "Without the main svg optimization, we had an aes score less than 0.5, so the optimization really boosted our score." } ] }, { "id": 3211021, "postDate": "2025-05-28T02:36:55.193Z", "content": "Thank you for sharing your team’s solution. Congratulations!
", "rawMarkdown": "Thank you for sharing your team’s solution. Congratulations!", "votes": 1 }, { "id": 3211018, "postDate": "2025-05-28T02:35:12.843Z", "content": "Congratulations to you! I'm impressed that you didn't use OCR-based attacks. Some elements in our attacks couldn't be adapted to DiffVG, and we also found that DiffVG tends not to converge when the complexity of the trained elements becomes too high (this was observed in Metric 2). Did you encounter a similar issue?
", "rawMarkdown": "Congratulations to you! I'm impressed that you didn't use OCR-based attacks. Some elements in our attacks couldn't be adapted to DiffVG, and we also found that DiffVG tends not to converge when the complexity of the trained elements becomes too high (this was observed in Metric 2). Did you encounter a similar issue?\n\n", "votes": 1, "replies": [ { "id": 3211033, "postDate": "2025-05-28T02:52:16.033Z", "content": "Congrats to your non-drawing solution! We didn't expect paligemma2 to be able to extract information from embedded texts when it couldn't detect them with OCR.
\n\n\nthe complexity of the trained elements becomes too high
\n
Could you explain this in more detail? Does it mean that there are many objects defined in the svg file?
", "rawMarkdown": "Congrats to your non-drawing solution! We didn't expect paligemma2 to be able to extract information from embedded texts when it couldn't detect them with OCR.\n\n> the complexity of the trained elements becomes too high\n\nCould you explain this in more detail? Does it mean that there are many objects defined in the svg file?", "votes": 4, "replies": [ { "id": 3211041, "postDate": "2025-05-28T02:57:56.077Z", "content": "Too many elements refer to the fact that I perform backpropagation on the parameters of first-, second-, and third-order Bézier curves in the path (of course with some constraints) in metric2, but I notice the loss fluctuates up and down. However, if I fix the shape and only backpropagate on the color, the aesthetic score in metric2 can exceed 1.1 points.
", "rawMarkdown": "Too many elements refer to the fact that I perform backpropagation on the parameters of first-, second-, and third-order Bézier curves in the path (of course with some constraints) in metric2, but I notice the loss fluctuates up and down. However, if I fix the shape and only backpropagate on the color, the aesthetic score in metric2 can exceed 1.1 points." }, { "id": 3211044, "postDate": "2025-05-28T03:03:39.743Z", "content": "We decided to join this compeititon after the metric was fixed as it is now, so we haven't had time to play around with metric 2.
\nAs for the current metric, we didn't use Bézier curves but only straight lines in the path, because we didn't want to make things too complicated, and also because we didn't think it would matter much after smoothing the image with ImageProcessor.apply().
\nMaybe that was the reason why our optimization worked.
Congratulations! It's great to see such a cool solution achieve high scores.
", "rawMarkdown": "Congratulations! It's great to see such a cool solution achieve high scores.", "votes": 1, "replies": [ { "id": 3211034, "postDate": "2025-05-28T02:53:12.570Z", "content": "Congrats to you too! It's impressive that a text-based attack and an image genration solution can be merged to result in such as high scoring solution.
", "rawMarkdown": "Congrats to you too! It's impressive that a text-based attack and an image genration solution can be merged to result in such as high scoring solution.", "votes": 1, "replies": [ { "id": 3211047, "postDate": "2025-05-28T03:07:25.603Z", "content": "I believe your images combined with text-based attacks will get higher scores, and that may be the final solution to this game
", "rawMarkdown": "I believe your images combined with text-based attacks will get higher scores, and that may be the final solution to this game", "votes": 1 } ] } ] }, { "id": 3211012, "postDate": "2025-05-28T02:28:23.117Z", "content": "This is actually a really nice and general solution to such a niche problem, congrats.
", "rawMarkdown": "This is actually a really nice and general solution to such a niche problem, congrats.", "votes": 1, "replies": [ { "id": 3211036, "postDate": "2025-05-28T02:54:22.340Z", "content": "Thanks! We are quite happy to be able to secure our ranking without using any model-specific attacks!
", "rawMarkdown": "Thanks! We are quite happy to be able to secure our ranking without using any model-specific attacks!" } ] }, { "id": 3210992, "postDate": "2025-05-28T02:02:01.387Z", "content": "Can I ask what your average aesthetic score was, and how you improved that score in the competition? Thank you.
", "rawMarkdown": "Can I ask what your average aesthetic score was, and how you improved that score in the competition? Thank you.", "votes": 1, "replies": [ { "id": 3211000, "postDate": "2025-05-28T02:10:49.197Z", "content": "Our VQA/AES scores were 0.81/0.64.
\nThanks for asking! (after reading your comment, we added them in the title)
this is what i call honest smart solution with lot of hard work behind, cheating should never be considered smart.. congratulations! just a short question: what was the average time for prediction?
", "rawMarkdown": "this is what i call honest smart solution with lot of hard work behind, cheating should never be considered smart.. congratulations! just a short question: what was the average time for prediction?", "votes": 2, "replies": [ { "id": 3211311, "postDate": "2025-05-28T09:47:24.410Z", "content": "We used around 63 seconds per image, which is just below the maximum time budget
", "rawMarkdown": "We used around 63 seconds per image, which is just below the maximum time budget", "votes": 2 } ] }, { "id": 3211030, "postDate": "2025-05-28T02:47:20.147Z", "content": "Your solution is highly versatile. After reviewing your work, everything clicked into place. My sincere congratulations to you!
", "rawMarkdown": "Your solution is highly versatile. After reviewing your work, everything clicked into place. My sincere congratulations to you!", "votes": 2 }, { "id": 3210990, "postDate": "2025-05-28T01:59:28.130Z", "content": "I was truly impressed by your work. 😄
\nYou managed to create an excellent solution using tools like FLUX.1-schnell, diffvg, and diffjpeg — tools I had personally thought would be difficult to run.
Your approach has given me new insight and has been a great source of inspiration.
\nThank you for the opportunity to learn so much from your work.
I asked cursor to write a Differentiable Preprocessing to mimic the ImageProcessor, but didn't work😅
Thare are libraries that can handle pytorch-based bilateral filtering and jpeg compression.
\nOther processes can easily be implemented using vanilla pytorch.
This is actually a really nice and general solution to such a niche problem, congrats.
", "rawMarkdown": "This is actually a really nice and general solution to such a niche problem, congrats.", "votes": -1 }, { "id": 3212039, "postDate": "2025-05-29T08:24:34.270Z", "content": "A huge thank uh to the organizers for creating such a thoughtful challenge!
", "rawMarkdown": "A huge thank uh to the organizers for creating such a thoughtful challenge! ", "votes": -1 }, { "id": 3217806, "postDate": "2025-06-05T12:19:01.910Z", "content": "Very nice work @cnumber Congrats!
\nTo be able to run your code, can you make private inputs like kagglehub.model_download(\"cnumber/nunchaku/pyTorch/default\") public? Thank you!
I took a different tack by leveraging a lightweight GAN to generate preliminary vector‐style sketches, then used a custom CLIP‐guided evolutionary optimizer to iteratively mutate and select SVG primitives under the 10 KB limit. Instead of pydiffvg, I implemented a pure PyTorch renderer with differentiable anti‐aliasing, which cut GPU memory usage by 40 %. I found that focusing on Bézier stitch simplifications—merging overlapping curves dynamically—boosted my AES score by 0.07. Finally, I used a small ensemble of CLIP‐ranked candidates before a final pass of SVG pruning, which kept fidelity high while staying within size constraints.
", "rawMarkdown": "I took a different tack by leveraging a lightweight GAN to generate preliminary vector‐style sketches, then used a custom CLIP‐guided evolutionary optimizer to iteratively mutate and select SVG primitives under the 10 KB limit. Instead of pydiffvg, I implemented a pure PyTorch renderer with differentiable anti‐aliasing, which cut GPU memory usage by 40 %. I found that focusing on Bézier stitch simplifications—merging overlapping curves dynamically—boosted my AES score by 0.07. Finally, I used a small ensemble of CLIP‐ranked candidates before a final pass of SVG pruning, which kept fidelity high while staying within size constraints.\n" }, { "id": 3214160, "postDate": "2025-05-31T03:47:49.917Z", "content": "Congrats! That's a really smart solution. Would you make the pydiffvg package in your notebook public? I believe most of us cannot get it work xD.
", "rawMarkdown": "Congrats! That's a really smart solution. Would you make the pydiffvg package in your notebook public? I believe most of us cannot get it work xD.", "replies": [ { "id": 3215470, "postDate": "2025-06-02T09:31:24.243Z", "content": "Hope you like it!
\nhttps://www.kaggle.com/models/cnumber/pydiffvg/
Wonderful solution, especially because it follows the policy and
\n DOES NOT hack the VQA evaluator using text embedding!
amazing.
\nImageProcessorTorch part, what's the input to this, is it generated svg code in text?ImageProcessorTorch module impelemented in our submission notebook. It's just a differentiable version of the original ImageProcessor.Thanks for the clarification! I had a look at your ImageProcessorTorch module — nicely done making it differentiable! That’s a smart move to keep the pipeline end-to-end trainable.
\nAlso, I really like your approach of balancing aesthetics and fidelity by combining MSE loss (for image similarity) and the SigLIP score (for alignment with the description). It’s a thoughtful way to preserve content while improving stylization. Great work!
", "rawMarkdown": "Thanks for the clarification! I had a look at your ImageProcessorTorch module — nicely done making it differentiable! That’s a smart move to keep the pipeline end-to-end trainable.\n\nAlso, I really like your approach of balancing aesthetics and fidelity by combining MSE loss (for image similarity) and the SigLIP score (for alignment with the description). It’s a thoughtful way to preserve content while improving stylization. Great work! \n\n" } ] } ] }, { "id": 3212222, "postDate": "2025-05-29T14:12:49.243Z", "content": "Thank you for sharing your teams idea and a beautiful solution.
", "rawMarkdown": "Thank you for sharing your teams idea and a beautiful solution." }, { "id": 3212115, "postDate": "2025-05-29T10:20:33.340Z", "content": "This is my favourite solution from the ones I read so far. Congratulations. This is inspirational stuff
", "rawMarkdown": "This is my favourite solution from the ones I read so far. Congratulations. This is inspirational stuff" }, { "id": 3211398, "postDate": "2025-05-28T11:33:49.923Z", "content": "Congratulations @cnumber and @daiwakun! I'm still trying to wrap my head around how you could manage to run all this without exceeding the time limit! Very impressive!
", "rawMarkdown": "Congratulations @cnumber and @daiwakun! I'm still trying to wrap my head around how you could manage to run all this without exceeding the time limit! Very impressive!" }, { "id": 3210991, "postDate": "2025-05-28T02:00:50.900Z", "content": "Congratulations on your impressive solution! Thank you for sharing it. I'm curious: how did you prevent diffvg from over-optimizing the colors, which can reduce their impact after running through ImageProcessor? I also experimented with diffvg, but I gave up when I noticed that the improvements in score reverted after applying ImageProcessor.
sorry, just found you already included that trick. I missed this part:
\nImageProcessorTorch - Differentiable Preprocessing:\n\n bridge the gap between our optimization process the competition\nCrucially, these operations were implemented be (mostly) differentiable use straight-through estimators necessary, allowing gradients flow back the SVG parameters.\nTransformations included: random crop & resize, JPEG compression (simulated via diff_jpeg), median filter, FFT low-pass filter, bilateral filter.\nThe rendered SVG was passed through this processor before loss calculation. The target bitmap the MSE loss was also processed the random crop & resize part this ( a synchronized random seed) ensure fair comparison.\nYou can also have a look at our ImageProcessorTorch class implemented in our solution!
", "rawMarkdown": "You can also have a look at our ImageProcessorTorch class implemented in our solution!", "votes": 1 } ] }, { "id": 3218302, "postDate": "2025-06-06T04:22:06.173Z", "rawMarkdown": "", "isDeleted": true }, { "id": 3213846, "postDate": "2025-05-30T14:51:22.177Z", "content": "Amazing！Thanks for sharing！
", "rawMarkdown": "Amazing！Thanks for sharing！" } ], "complete": true, "source": "live", "fetched_at": "2026-10-20T09:08:00Z", "competition_id": 89659, "competition_slug": "drawing-with-llms", "team_id": 13402127, "solution_url": "/competitions/drawing-with-llms/writeups/cp-gms-3rd-place-solution-vqa-aes-0-81-0-64-diffus" }