The size of the model demanded extensive NVIDIA GPU resources, which was challenging, but the real hurdle fal faced was how to consume those resources—mainly, when training and experimenting with their models, how could they access what they needed, when they needed it. Initially, fal trained version 0.1 of AuraFlow using a onemonth contract with a cloud provider. While this allowed them to start development, it took weeks to negotiate a short-term contract that worked for them, and those were cycles fal didn’t want to spend every time they trained a new version of the model.

This prompted fal to seek a more agile infrastructure that can scale up for experimentation and training within seconds, then instantly scale down when not in use. They needed a system that provided on-demand resources, without tying them down to long-term agreements or paying for idle GPUs. As the fal’s Founding Engineer—Batuhan Taskaya—put it, “the first question we had was how the hell are we going to train this in terms of the GPUs we need, where are we going to find the GPUs?” With future iterations in mind, including versions 0.2 and 0.3, they sought asystem capable of adapting resource levels in real-time.

Multi-node, Multimodal, Multipurpose

Lambda’s 1-Click Cluster (1CC) provided the answer for the next stages of AuraFlow development, offering fal a self-serve, on-demand cluster solution to rapidly experiment without the upfront commitments. Taskaya immediately realized the potential of 1CC for future iterations of AuraFlow, saying, “I saw 1CC... and I was like, this is an amazing idea... Lambda has [a model where you pay only for what you need without us needing to spend time working out a short-term agreement—it was all basically self-serve.]” The shift to Lambda for AuraFlow 0.2 and 0.3 development allowed them to expedite the experimentation process, iterate faster, and significantly reduce setup time.

Accelerated Processes and Technical Impact

By using Lambda’s 1CC, fal was able to act quickly, releasing AuraFlow versions 0.2 and 0.3 in weeks rather than months. The ability to run multiple experiments simultaneously was critical for the advancement of AuraFlow as it allowed for faster iterations, optimal use of NVIDIA GPU resources, comprehensive testing of diverse approaches, and more adaptability as they learned and tuned the model.

Taskaya commented, “We are a very agile startup, it's a very, very competitive market and we want to iterate fast, right? If I have a cool idea, I need like 22 GPU assets for a week, I don't want to wait four days to go into a sales meeting and then another four days [to finalize a contract and get going] … I think that’s the main benefit I see from my 1CC.”

Business Agility and Impact

This flexibility allowed them to continually experiment with new ideas, test hypotheses, and make quick pivots as necessary by leveraging only the resources they needed and when they needed them. This agility helped amplify their standing as a leader in the open-source image generation community and opened new doors as they became a thought leader for the space.

Building on the success of AuraFlow and the recognition of fal, they have expanded their offerings by collaborating with customers to create custom generative AI models tailored to specific needs. These projects have opened new revenue streams, as businesses seek out fal’s expertise to develop cutting-edge AI solutions, positioning them as a go-to partner for generative AI innovation.