First of all, I’d like to express my gratitude to my teammates ( @kazukishida , @nozomiyamamoto ), Kaggle, and everyone who participated in this competition. Special thanks to @richolson for releasing such an excellent starter notebook!
\nThe three of us on the team have been friends for a long time, and after placing 4th on the public leaderboard in the Konwinski Prize competition (private results will be announced on June 12), we joined this competition right afterward.
\nSince the provided training dataset included only 15 samples, we needed to create our own dataset.
\nWe followed the steps below:
Based on our judgment, we believe the following list covers most relevant categories:
\ncategories: [] = [\n ,\n ,\n ,\n ,\n ,\n ,\n ,\n ,\n ,\n ,\n ,\n ,\n ,\n ,\n ,\n ,\n ,\n ,\n ,\n ,\n ,\n ,\n ,\n ,\n]\nTo ensure completeness and avoid category omissions, we used the LLM 20 Questions solution.
\nWhile we brainstormed many possible categories, we concluded that it was not necessary to cover every possible category exhaustively.
\nThe key was to design an algorithm that performs well on the distribution of the dataset.
So we built a 3-class classification model for the categories landscape, abstract, and fashion, and split our dataset into landscape-like, abstract-like, and fashion-like subsets.
\nWe then fine-tuned generation parameters for each subset.
\nOne of our final submissions also included an animals-like category.
We tried various models from the list shared by @kawchar85 and selected SDXL Turbo, a high-performance model with fast inference time.
\nFor the scheduler, we adopted HeunDiscreteScheduler, which performed best both in CV and on the leaderboard.
\nEach of the categories — landscape-like, abstract-like, fashion-like, and animal-like (or landscape/abstract/fashion in another final submission) — used the following parameter configurations:
\n| Category | \nImage Size | \nnum_colors | \nignore_area_threshold | \nguidance_scale | \n
|---|---|---|---|---|
| landscape-like | \n512 | \n12 or 9 | \n65 or 55 | \n5 or 20 | \n
| abstract-like | \n384 | \n10 or 9 | \n55 or 55 | \n5 or 20 | \n
| fashion-like | \n384 | \n8 or 9 | \n60 or 55 | \n5 or 20 | \n
| animal-like | \n384 | \n9 | \n65 | \n5 | \n
Common parameters across all categories:
\nnum_attempts_per_description: = \nnum_inference_steps: = \nprompt_prefix: = \nprompt_suffix: = \nnegative_prompt: = \nInitially, we experimented with adding various short text snippets to reduce hallucination, which showed some effect.
\nHowever, the OCR hallucination prevention function published by @richolson was excellent, so we adopted it.
To reduce the number of online evaluations, we only ran OCR evaluations on top candidate SVGs ranked by metrics, instead of evaluating all of them.
\nThis allowed us to maintain quality while significantly reducing evaluation calls.
Observation: In online evaluation, if an SVG was flagged for OCR hallucination, it typically was in the final evaluation as well.
\nHowever, there were rare cases where an SVG passed the online test but still failed in the final evaluation.
We tried to simulate VQA evaluation using AI, inspired by the provided dataset of questions, choices, and answers.
\nWe generated 417 questions and evaluated them to find the combinations with the highest correlation to the true VQA.
The final selected 3 questions were:
\nquestions.append()\nchoices_list.append()\nanswers.append()\n\nquestions.append()\nchoices_list.append()\nanswers.append()\n\nquestions.append()\nchoices_list.append()\nanswers.append()\nWe used the following weights for scoring: [0.50, 0.29, 0.21].
\nCongratulations, this is a very systematic plan. I've learned a lot from it.
", "votes": 2, "replies": [] } ], "comment_count": 1, "messages": [ { "id": 3210951, "postDate": "2025-05-28T00:34:09.113Z", "content": "First of all, I’d like to express my gratitude to my teammates ( @kazukishida , @nozomiyamamoto ), Kaggle, and everyone who participated in this competition. Special thanks to @richolson for releasing such an excellent starter notebook!
\nThe three of us on the team have been friends for a long time, and after placing 4th on the public leaderboard in the Konwinski Prize competition (private results will be announced on June 12), we joined this competition right afterward.
\nSince the provided training dataset included only 15 samples, we needed to create our own dataset.
\nWe followed the steps below:
Based on our judgment, we believe the following list covers most relevant categories:
\ncategories: [] = [\n ,\n ,\n ,\n ,\n ,\n ,\n ,\n ,\n ,\n ,\n ,\n ,\n ,\n ,\n ,\n ,\n ,\n ,\n ,\n ,\n ,\n ,\n ,\n ,\n]\nTo ensure completeness and avoid category omissions, we used the LLM 20 Questions solution.
\nWhile we brainstormed many possible categories, we concluded that it was not necessary to cover every possible category exhaustively.
\nThe key was to design an algorithm that performs well on the distribution of the dataset.
So we built a 3-class classification model for the categories landscape, abstract, and fashion, and split our dataset into landscape-like, abstract-like, and fashion-like subsets.
\nWe then fine-tuned generation parameters for each subset.
\nOne of our final submissions also included an animals-like category.
We tried various models from the list shared by @kawchar85 and selected SDXL Turbo, a high-performance model with fast inference time.
\nFor the scheduler, we adopted HeunDiscreteScheduler, which performed best both in CV and on the leaderboard.
\nEach of the categories — landscape-like, abstract-like, fashion-like, and animal-like (or landscape/abstract/fashion in another final submission) — used the following parameter configurations:
\n| Category | \nImage Size | \nnum_colors | \nignore_area_threshold | \nguidance_scale | \n
|---|---|---|---|---|
| landscape-like | \n512 | \n12 or 9 | \n65 or 55 | \n5 or 20 | \n
| abstract-like | \n384 | \n10 or 9 | \n55 or 55 | \n5 or 20 | \n
| fashion-like | \n384 | \n8 or 9 | \n60 or 55 | \n5 or 20 | \n
| animal-like | \n384 | \n9 | \n65 | \n5 | \n
Common parameters across all categories:
\nnum_attempts_per_description: = \nnum_inference_steps: = \nprompt_prefix: = \nprompt_suffix: = \nnegative_prompt: = \nInitially, we experimented with adding various short text snippets to reduce hallucination, which showed some effect.
\nHowever, the OCR hallucination prevention function published by @richolson was excellent, so we adopted it.
To reduce the number of online evaluations, we only ran OCR evaluations on top candidate SVGs ranked by metrics, instead of evaluating all of them.
\nThis allowed us to maintain quality while significantly reducing evaluation calls.
Observation: In online evaluation, if an SVG was flagged for OCR hallucination, it typically was in the final evaluation as well.
\nHowever, there were rare cases where an SVG passed the online test but still failed in the final evaluation.
We tried to simulate VQA evaluation using AI, inspired by the provided dataset of questions, choices, and answers.
\nWe generated 417 questions and evaluated them to find the combinations with the highest correlation to the true VQA.
The final selected 3 questions were:
\nquestions.append()\nchoices_list.append()\nanswers.append()\n\nquestions.append()\nchoices_list.append()\nanswers.append()\n\nquestions.append()\nchoices_list.append()\nanswers.append()\nWe used the following weights for scoring: [0.50, 0.29, 0.21].
\nCongratulations, this is a very systematic plan. I've learned a lot from it.
", "rawMarkdown": "Congratulations, this is a very systematic plan. I've learned a lot from it.", "votes": 2 } ], "complete": true, "source": "live", "fetched_at": "2026-10-20T09:07:41Z", "competition_id": 89659, "competition_slug": "drawing-with-llms", "team_id": 13419140, "solution_url": "/competitions/drawing-with-llms/writeups/graphninjas-135th-place-solution" }