{ "id": 581106, "title": "79th Place Solution: Using stable-diffusion-cpp-python", "content": "

First, I had compiled llama-cpp-python for a previous Kaggle competition, so I decided to use stable-diffusion-cpp-python in this competition.

\n

My plan was simple: use FLUX.1-schnell Q6_K with stable-diffusion-cpp-python to outperform SD-based or more heavily quantized FLUX.1-based solutions.

\n

For conversion to compliant SVG, I used VTracer and Scour. To optimize parameters within the time limit, I used binary search and SSIM-based comparison.

\n

However, the biggest problem was that the image generation was much slower than I expected — over 55 seconds on the first image! I tried to use the following patch for T4 x2 GPUs, but it didn’t work well:
\n#461 Ability to split model components to different backend devices

\n

The bug may come from the following issue so I attempted to avoid NaNs that occurred when running on CUDA, but it was too complicated, and I wasn't able to solve it.
\n#636 CLIP: Using CPU Backend despite configuring for CUDA

\n

Due to time constraints, I couldn’t really pay attention to optimizing VQA or aesthetic scores. The prompt I used was \"{original prompt}, simple toon vector image, monoline, icon style, limited color palette, aesthetic\" but it was just a heuristic, and I didn’t test it thoroughly. The score varied a lot with different seeds.

\n

Additionally, I used three nearly fixed-position OCR decoy and CLIP-GmP-ViT-L-14 — a fine-tuned version of the CLIP model — for stabilizing the score. They were also heuristic and not well tested, so I don't know how good they are.

\n

The Final Code: https://www.kaggle.com/code/toshimitsukimura/79th-place-solution-flux-1-schnell-vtracer-scour

", "comments": [], "comment_count": 0, "messages": [ { "id": 3211380, "postDate": "2025-05-28T11:03:53.520Z", "content": "

First, I had compiled llama-cpp-python for a previous Kaggle competition, so I decided to use stable-diffusion-cpp-python in this competition.

\n

My plan was simple: use FLUX.1-schnell Q6_K with stable-diffusion-cpp-python to outperform SD-based or more heavily quantized FLUX.1-based solutions.

\n

For conversion to compliant SVG, I used VTracer and Scour. To optimize parameters within the time limit, I used binary search and SSIM-based comparison.

\n

However, the biggest problem was that the image generation was much slower than I expected — over 55 seconds on the first image! I tried to use the following patch for T4 x2 GPUs, but it didn’t work well:
\n#461 Ability to split model components to different backend devices

\n

The bug may come from the following issue so I attempted to avoid NaNs that occurred when running on CUDA, but it was too complicated, and I wasn't able to solve it.
\n#636 CLIP: Using CPU Backend despite configuring for CUDA

\n

Due to time constraints, I couldn’t really pay attention to optimizing VQA or aesthetic scores. The prompt I used was \"{original prompt}, simple toon vector image, monoline, icon style, limited color palette, aesthetic\" but it was just a heuristic, and I didn’t test it thoroughly. The score varied a lot with different seeds.

\n

Additionally, I used three nearly fixed-position OCR decoy and CLIP-GmP-ViT-L-14 — a fine-tuned version of the CLIP model — for stabilizing the score. They were also heuristic and not well tested, so I don't know how good they are.

\n

The Final Code: https://www.kaggle.com/code/toshimitsukimura/79th-place-solution-flux-1-schnell-vtracer-scour

", "rawMarkdown": "First, I had compiled `llama-cpp-python` for a previous Kaggle competition, so I decided to use `stable-diffusion-cpp-python` in this competition.\n\nMy plan was simple: use `FLUX.1-schnell Q6_K` with `stable-diffusion-cpp-python` to outperform SD-based or more heavily quantized FLUX.1-based solutions.\n\nFor conversion to compliant SVG, I used VTracer and Scour. To optimize parameters within the time limit, I used binary search and SSIM-based comparison.\n\nHowever, the biggest problem was that the image generation was much slower than I expected — over 55 seconds on the first image! I tried to use the following patch for T4 x2 GPUs, but it didn’t work well: \n[#461 Ability to split model components to different backend devices](https://github.com/leejet/stable-diffusion.cpp/pull/461)\n\nThe bug may come from the following issue so I attempted to avoid NaNs that occurred when running on CUDA, but it was too complicated, and I wasn't able to solve it.\n[#636 CLIP: Using CPU Backend despite configuring for CUDA](https://github.com/leejet/stable-diffusion.cpp/issues/636)\n\nDue to time constraints, I couldn’t really pay attention to optimizing VQA or aesthetic scores. The prompt I used was *\"{original prompt}, simple toon vector image, monoline, icon style, limited color palette, aesthetic\"* but it was just a heuristic, and I didn’t test it thoroughly. The score varied a lot with different seeds.\n\nAdditionally, I used three nearly fixed-position OCR decoy and [CLIP-GmP-ViT-L-14](https://huggingface.co/zer0int/CLIP-GmP-ViT-L-14) — a fine-tuned version of the CLIP model — for stabilizing the score. They were also heuristic and not well tested, so I don't know how good they are.\n\nThe Final Code: https://www.kaggle.com/code/toshimitsukimura/79th-place-solution-flux-1-schnell-vtracer-scour", "votes": 4 } ], "complete": true, "source": "live", "fetched_at": "2026-10-20T09:10:10Z", "competition_id": 89659, "competition_slug": "drawing-with-llms", "team_id": 13746267, "solution_url": "/competitions/drawing-with-llms/writeups/toshimitsu-kimura-79th-place-solution-using-stable" }