![\"Sample](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F12376195%2Fb2db60069ff4dbdb76090bdb1f2a8fca%2FScreenshot%202025-05-28%20at%201.57.31PM.png?generation=1748419083311977&alt=media\")
\nI’d like to extend my sincere thanks to the organizers for their hard work and dedication in making this competition both engaging and meaningful. I also appreciate everyone who shared helpful ideas throughout—this community support made a big difference. My custom image-to-SVG pipeline was inspired by the work of @richolson , and I’m especially grateful for his contributions.
I'm a beginner in this field with limited experience and knowledge, so feedback and suggestions are always welcome!
\nTo generate high-quality evaluation questions, I created a dataset of approximately 10,000 structured examples using Gemini and Claude. Each set follows a strict 4-question TIFA format:
\nI fine-tuned SmolLM2-1.7B-Instruct on this dataset. The model performs well and significantly improved my competition score. In around 2% of cases, Q2 and Q3 were duplicates—likely due to minor inconsistencies in my training data or the limited capacity of the small model. To handle such edge cases, I added fallback logic to inject dummy questions when fewer than four unique ones are generated.
\nInference is extremely fast, averaging about 3 seconds per prompt.
\nI used SDXL Flash due to its speed-quality tradeoff. Prompts were constructed using a consistent template:
\n\n\nflat vector illustration of {description}, minimalist design, solid colors, flat color blocks
\n
Each prompt gets 4 attempts by default, extended dynamically using leftover time from previous predictions.
\nEach generated image is first resized from 768×768 to 384×384 to reduce SVG complexity. The image is then quantized to 16 colors using MiniBatchKMeans in LAB color space. After quantization, I apply a Disjoint Set Union (DSU) algorithm to merge adjacent regions with similar colors. The processed image is then passed to two SVG conversion pipelines:
\nThis method allocates a total point budget (e.g., 2000 points) and distributes it across contours based on an importance metric combining area, shape complexity, and proximity to the image center. For each contour, I perform a binary search to find an optimal epsilon value that simplifies the contour to the target number of points using cv2.approxPolyDP.
\nEach simplified shape is then encoded as either a polygon or a path, depending on which representation is shorter in bytes. This yields visually accurate SVGs while keeping file size under control.
\nAs an alternative, I also use VTracer to convert the image into layered SVG paths. After tracing, a binary search simplification pass is applied to the path elements using cv2.approxPolyDP to further reduce the file size and ensure the final SVG remains under the 10KB limit.
\nFor each generated image, I produce two SVGs—one using the custom polygon/path pipeline and the other using VTracer with binary search.
\nThe competition's scoring model often hallucinates the presence of text in SVGs, negatively impacting the OCR score and overall performance. To counter this, I inject a small \"L\" shaped watermark into a fixed position within each SVG.
\nThe color of this watermark is chosen dynamically—either black or white—based on the local brightness of the region to ensure it blends naturally with the image while still being detectable. This simple yet effective trick significantly reduces false positives in OCR evaluation.
\nTo select the best SVG, I use PaliGemma-3B, which is approximately 3× faster than the 10B version, with minimal loss in accuracy.
\nEach description is allocated a fixed time budget of 64 seconds. If a sample finishes early, the unused time is accumulated and carried forward.
\nThis saved time is then reused to allow additional image generation attempts for future descriptions, calculated as:
\nattempts = int(base_num_attempts + saved_time / per_attempt_time)
To guide submission selecting, I created a validation dataset of 100 descriptions covering 11 diverse categories (Nature, Abstract, Fashion, Animal, Household, Food, Urban, Technology, Transport, Sports/Recreation, Medical/Healthcare), all generated using Gemini.
\nYou can find the dataset here.
\nI ran multiple inference notebooks and tracked both CV and leaderboard performance:
\n| Notebook | \nPublic LB | \nCV Score | \nPrivate LB | \nFinal Submission | \n
|---|---|---|---|---|
| 1 | \n0.699 | \n0.724 | \n0.690 | \nNo | \n
| 2 | \n0.696 | \n0.730 | \n0.699 | \nNo | \n
| 3 | \n0.690 | \n0.752 | \n0.686 | \nYes | \n
| 4 | \n0.689 | \n0.763 | \n0.693 | \nYes | \n
Notebook 3 and Notebook 4 were selected for final submission based on strong cross-validation performance. Unfortunately, I missed selecting Notebook 2—which actually had a slightly higher private leaderboard score.
\nLooking back, there are a few key areas where I could have improved:
\n
\nI’d like to extend my sincere thanks to the organizers for their hard work and dedication in making this competition both engaging and meaningful. I also appreciate everyone who shared helpful ideas throughout—this community support made a big difference. My custom image-to-SVG pipeline was inspired by the work of @richolson , and I’m especially grateful for his contributions.
I'm a beginner in this field with limited experience and knowledge, so feedback and suggestions are always welcome!
\nTo generate high-quality evaluation questions, I created a dataset of approximately 10,000 structured examples using Gemini and Claude. Each set follows a strict 4-question TIFA format:
\nI fine-tuned SmolLM2-1.7B-Instruct on this dataset. The model performs well and significantly improved my competition score. In around 2% of cases, Q2 and Q3 were duplicates—likely due to minor inconsistencies in my training data or the limited capacity of the small model. To handle such edge cases, I added fallback logic to inject dummy questions when fewer than four unique ones are generated.
\nInference is extremely fast, averaging about 3 seconds per prompt.
\nI used SDXL Flash due to its speed-quality tradeoff. Prompts were constructed using a consistent template:
\n\n\nflat vector illustration of {description}, minimalist design, solid colors, flat color blocks
\n
Each prompt gets 4 attempts by default, extended dynamically using leftover time from previous predictions.
\nEach generated image is first resized from 768×768 to 384×384 to reduce SVG complexity. The image is then quantized to 16 colors using MiniBatchKMeans in LAB color space. After quantization, I apply a Disjoint Set Union (DSU) algorithm to merge adjacent regions with similar colors. The processed image is then passed to two SVG conversion pipelines:
\nThis method allocates a total point budget (e.g., 2000 points) and distributes it across contours based on an importance metric combining area, shape complexity, and proximity to the image center. For each contour, I perform a binary search to find an optimal epsilon value that simplifies the contour to the target number of points using cv2.approxPolyDP.
\nEach simplified shape is then encoded as either a polygon or a path, depending on which representation is shorter in bytes. This yields visually accurate SVGs while keeping file size under control.
\nAs an alternative, I also use VTracer to convert the image into layered SVG paths. After tracing, a binary search simplification pass is applied to the path elements using cv2.approxPolyDP to further reduce the file size and ensure the final SVG remains under the 10KB limit.
\nFor each generated image, I produce two SVGs—one using the custom polygon/path pipeline and the other using VTracer with binary search.
\nThe competition's scoring model often hallucinates the presence of text in SVGs, negatively impacting the OCR score and overall performance. To counter this, I inject a small \"L\" shaped watermark into a fixed position within each SVG.
\nThe color of this watermark is chosen dynamically—either black or white—based on the local brightness of the region to ensure it blends naturally with the image while still being detectable. This simple yet effective trick significantly reduces false positives in OCR evaluation.
\nTo select the best SVG, I use PaliGemma-3B, which is approximately 3× faster than the 10B version, with minimal loss in accuracy.
\nEach description is allocated a fixed time budget of 64 seconds. If a sample finishes early, the unused time is accumulated and carried forward.
\nThis saved time is then reused to allow additional image generation attempts for future descriptions, calculated as:
\nattempts = int(base_num_attempts + saved_time / per_attempt_time)
To guide submission selecting, I created a validation dataset of 100 descriptions covering 11 diverse categories (Nature, Abstract, Fashion, Animal, Household, Food, Urban, Technology, Transport, Sports/Recreation, Medical/Healthcare), all generated using Gemini.
\nYou can find the dataset here.
\nI ran multiple inference notebooks and tracked both CV and leaderboard performance:
\n| Notebook | \nPublic LB | \nCV Score | \nPrivate LB | \nFinal Submission | \n
|---|---|---|---|---|
| 1 | \n0.699 | \n0.724 | \n0.690 | \nNo | \n
| 2 | \n0.696 | \n0.730 | \n0.699 | \nNo | \n
| 3 | \n0.690 | \n0.752 | \n0.686 | \nYes | \n
| 4 | \n0.689 | \n0.763 | \n0.693 | \nYes | \n
Notebook 3 and Notebook 4 were selected for final submission based on strong cross-validation performance. Unfortunately, I missed selecting Notebook 2—which actually had a slightly higher private leaderboard score.
\nLooking back, there are a few key areas where I could have improved:
\n