# 45th Place Solution

Acknowledgements and Reflections on the Competition
First of all, I would like to express my
sincere gratitude to Kaggle
for organizing this competition, which provided an excellent opportunity to expand my knowledge and skills.
I would also like to thank Kawsar Hossain
[@kawchar85](https://www.kaggle.com/kawchar85)
for his insightful post  (
[https://www.kaggle.com/competitions/drawing-with-llms/discussion/576181#3207144](url)
), which served as a significant source of inspiration for my work.
Competition Dynamics
In the final week of the competition, the dynamics on the leaderboard
intensified considerably
, leading me to suspect that a novel state-of-the-art (SOTA) text-to-image model might have been discovered by some participants.
Consequently, I conducted a thorough review of recent publications on
arXiv
,
Hugging Face
, and
OpenReview
, yet found no substantial breakthroughs.
This prompted me to re-evaluate my strategy and focus on
optimizing my existing solution
.
![Figure 1](assets/c7ba656ed5a92fd93f6e9147.jpg)
Key Experimental Findings
Below is a summary of key experimental findings from the competition:
1.
SDXL Base / Turbo and SDXL Hyper
User-friendly
and require minimal parameter tuning.
Facilitate a
lower barrier to entry
.
I adopted this approach during the mid-stage, which initially achieved a
high ranking
but declined towards the end.
2.
SANA 1.5
Performs well with
straightforward parameter settings
, yielding relatively high scores.
I hypothesize that
more advanced hyperparameter optimization
could further improve performance; however, time constraints precluded additional exploration.
3.
DMD2 (4-step UNet generation)
One of my final choices.
By utilizing the
demo code provided in the Model Card
, I achieved a high score with minimal effort.
Simple parameter adjustments
resulted in an approximate
0.02 improvement
.
Further fine-tuning could yield even better results.
Demonstrated considerable stability
.
4.
SDXL Flash
Also among my final submissions.
Achieved a score of
0.67 on the public leaderboard
using the demo code.
Basic parameter tuning
allowed for further improvements.
Limitation:
Instability—identical configurations often produced significantly different results across multiple submissions.
Summary
Throughout the competition, I experimented with a
variety of models and strategies
. While some models offered
ease of use and robust baseline performance
, others required more careful tuning to achieve optimal results.
Stability and reproducibility
emerged as important considerations, particularly in the final evaluation stage.