🔍 Read the full analysis: The Future Of Watercolour Painting: AI Training With TRL And OpenEnv on ThorstenMeyerAI.com
Get business pricing on monitors, keyboards and dev gear
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
A developer has successfully reproduced Surya Narreddi’s viral watercolour AI model using open-source infrastructure, releasing all datasets, training scripts, and models. This project tests reinforcement learning for aesthetic taste without fixed answers, emphasizing transparency and open access.
An independent engineer has completed and published a full, open reproduction of Surya Narreddi’s viral watercolour painting AI, utilizing TRL and OpenEnv on Hugging Face infrastructure. This reproduction includes datasets, training scripts, and trained models, making the project fully accessible to the community as detailed in the original analysis. The work aims to explore whether reinforcement learning can optimize models based on aesthetic preference rather than explicit correctness, marking a significant step in AI art development.
The project replicates Narreddi’s watercolour model, which gained viral attention in August after a video showcasing AI-generated watercolour paintings amassed over 1.5 million views. Unlike typical AI models trained for accuracy or pixel-perfect images, Narreddi’s approach involves training a language model to generate JavaScript code that produces watercolour effects using the p5.brush library. The reproduction, built with TRL and OpenEnv, trains a Qwen model with reinforcement learning against a composite reward that balances code correctness, style adherence, and human aesthetic preferences.
All artifacts, including datasets, environment scripts, and trained models, are openly available on Hugging Face. The project tests multiple reward mixes, combining a code compilation gate, code length, style scoring via a vision-language model, and a preference model trained on human choices. This approach is similar to recent efforts in training AI models for creative tasks, such as training a coding model to generate artistic watercolours. The training process involves a single command executed against Qwen/Qwen3.5-35B-A3B with LoRA, running 110 steps, with 240 episodes per step and eight generations per episode. The goal: to see if RL can effectively optimize for aesthetic taste without explicit correctness, a departure from traditional AI training focused on verifiable answers.
The author emphasizes that this open release provides the community with a transparent, replicable pipeline, addressing gaps in the original project, which did not include open artifacts or detailed datasets. For more on innovative AI art techniques, see the original analysis. The project also explores the lineage of AI art, from early generative tools like DeepDream to curated datasets used by artists like Anna Ridler. This work underscores a shift toward models that produce more human-like, imperfect, and expressive art, contrasting with the ‘perfect’ outputs of mainstream image generators.
Implications of Open-Source Watercolour AI Reproduction
This project demonstrates that reinforcement learning can be applied to aesthetic preferences, not just objective correctness, opening new avenues for AI art creation. By releasing all datasets, models, and scripts openly, it lowers barriers for researchers and artists to experiment with AI-driven art, fostering transparency and collaborative development.
The shift toward models that optimize for taste rather than accuracy could influence future AI art tools, making them more adaptable to human preferences and creative expression. It also highlights the potential for AI to assist artists in exploring new styles and techniques, broadening the scope of digital art practice. The project’s emphasis on interpretability—since the code-based output allows inspection and editing—further enhances its relevance for artists and developers interested in understanding AI decision-making processes.
As an affiliate, we earn on qualifying purchases.
Background on AI Art and Reinforcement Learning Approaches
The project builds on a history of AI art experimentation, from DeepDream (2015) to neural network portraits by Mario Klingemann, and datasets curated by artists like Anna Ridler. Early AI art often involved using neural networks as tools or debugging aids, but recent advances have seen models producing increasingly sophisticated images. Narreddi’s initial work focused on training models to generate floral images, using human-curated datasets and custom reward functions to guide artistic style.
Traditional reinforcement learning in AI typically relies on well-defined, verifiable rewards—such as solving math problems or passing tests. Narreddi’s approach diverges by using aesthetic preference as a reward, modeled through a combination of style scoring and human-like preference models. The original project gained attention through a viral video showing AI-generated watercolour paintings, which prompted further exploration into open, replicable methods for training such models. The current reproduction aims to make these methods accessible for broader experimentation, with full transparency of datasets and training procedures.
“This open reproduction aims to test whether reinforcement learning can optimize models based on aesthetic taste, rather than fixed correctness, opening new possibilities for AI art.”
— Thorsten Meyer, project author
digital watercolor brushes for drawing tablets
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Unresolved Questions About Model Performance and Fidelity
It remains unclear how closely the reproduction’s outputs match the quality and style of the original viral paintings. The project presents visual median outputs but does not provide a quantitative comparison or final evaluation of which reward mix yields the best results. Additionally, the full technical report from Narreddi, which could clarify the original model’s performance and training details, has not yet been published. The extent to which the open reproduction replicates the artistic expressiveness of the viral video is still under assessment.
AI art generation tools for watercolor effects
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Next Steps for Open-Source Watercolour AI Development
The next phase involves awaiting Narreddi’s full technical report, which promises detailed insights into the original model’s training and performance. The open reproduction team plans to experiment with different reward configurations, refine the training pipeline, and compare outputs against the original viral paintings. Further community engagement and collaborative projects are expected to explore how reinforcement learning can be applied to other artistic styles and mediums. Additionally, researchers and artists may adapt the pipeline to develop new AI tools that prioritize aesthetic preferences, fostering innovation in AI-assisted art creation.
As an affiliate, we earn on qualifying purchases.
Key Questions
What is the main goal of this open reproduction?
The primary aim is to demonstrate that reinforcement learning can optimize AI models based on aesthetic taste rather than fixed correctness, making AI art more expressive and human-like.
Are the datasets and models publicly available?
Yes, all datasets, training scripts, and trained models are openly released on Hugging Face, allowing others to replicate and build upon the work.
How does this differ from traditional AI image generation?
Unlike typical models that generate pixel-perfect images, this approach produces code-based outputs that can be inspected, edited, and guided by aesthetic preferences, emphasizing interpretability and artistic expression.
The reproduction aims to replicate the process and style, but it is not yet confirmed how closely the outputs match the original viral paintings in quality and expressiveness.
What are the implications for future AI art tools?
This work suggests that AI models can be trained to prioritize taste and style, potentially leading to more creative, adaptable, and artistically expressive AI-assisted tools.
Primary source: Hugging Face · via ThorstenMeyerAI.com
Halloween Picks
halloween
As an affiliate, we earn on qualifying purchases.
