Speed vs. Quality in GenAI: I Built a Low Latency Voice-Controlled Image Generation Application. Here’s What I Learned. 

/ By

We have all been there. You type a prompt into an image generator, hit enter, and then… wait. And wait. 
This waiting led to the central question for my Master’s thesis: Do we have to choose between speed and quality? 

What if we could run a powerful visual AI system right on our home computer, control it instantly with just our voice, and keep all our data 100% private? 

I decided to build one to find out. 

The Build: A Local, Voice-Guided Pipeline 

I designed a voice-controlled AI pipeline capable of running on a single high-end consumer GPU. It operates through a distinct four-step process: 

  1. Transcription: OpenAI’s Whisper transcribes spoken commands to text. 
  2. Interpretation: A small Language Model analyzes the text to determine user intent. 
  3. Translation: A compact model translates Finnish commands into English to improve accuracy. 
  4. Generation: Diffusion models generate or edit images according to the user’s wishes. 
(The components of the pipeline work together to allow use to create and edit images and video.)

The Comparison: Local vs. Cloud 

With the system built, I tested my local pipeline against comparable commercial cloud-based services. In terms of raw latency, the local system was dramatically faster. 

  • Creating an Image: Local took 2.66 seconds vs. Cloud’s 11.0 seconds. 
  • Editing an Image: Local took 1.10 seconds vs. Cloud’s 13.5 seconds. 

The local system was nearly 10x faster at editing tasks. It felt instant, incurred zero API costs, and kept all data private. 

(End-to-end latency comparison between local and cloud pipelines across content creation tasks.)

The Unexpected Result: Users Preferred Quality 

To validate these findings, I ran a user experience study with 20 participants. They used both systems to perform creative tasks. The result was surprising: 18 out of 20 participants said they were willing to wait longer for the cloud output. 

While the local system won on speed, the cloud system won on overall preference. Users felt the cloud models followed complex instructions more accurately and produced higher-quality images. They preferred waiting 11 seconds for the “right” image rather than getting an “okay” image in 2 seconds. 

(User preferences across speed, instruction-following accuracy, image quality and overall preference, comparing local and cloud systems.)

Conclusion: The Future is Hybrid? 

My research suggests we do not have to choose strictly between local or cloud hosting. The ideal workflow for creative AI tools is likely a hybrid model: 

  • Use Local AI for brainstorming and rapid iteration where speed and privacy are paramount. 
  • Use Cloud AI for the final, high-definition render once the concept is finalized. 

Want to explore the details? 

🎥 Watch the Demo: https://www.youtube.com/watch?v=_C46f6m7Dxs (in Finnish) 

📄 Read the Thesis: https://trepo.tuni.fi/handle/10024/230809  

💻 Get the Code: https://github.com/Koodattu/voice-guided-imaging  

(Transparency Note: Consistent with our topic, this thesis and this blog post were produced with the support of AI tools for ideation, editing, structure, and visualization.) 

About the author

Juha Ala-Rantala

Research Assistant

Scroll to Top