Gen AI in Action: Real-World Applications of Image Generation Models

This talk was delivered in PyCon Poland 2025 (https://pl.pycon.org/2025/en/agenda/) on August 30, 2025.
Abstract: Explore AI-generated art with this talk on image generation models like diffusion models and GANs. Watch live demos on virtual cameras, fine-tuning for personalized images, inpainting, and ControlNet. Learn practical applications, Python integration, and ethical considerations in the innovative world of AI-powered creativity.

1. Confidential and Proprietary. Do not distribute without Couchbase consent. © Couchbase 2025. All rights reserved.
Real-World Applications of
Image Generation Models
PyCon Poland 2025
Nithish Raghunandanan, Developer Advocate

2. Confidential and Proprietary. Do not distribute without Couchbase consent. © Couchbase 2025. All rights reserved. 2
About Me
●Background in Data/ML Engineering
●Developer Advocate @ Couchbase
●Building Integrations with Developer Ecosystem
●PyData Munich


@nithishr

3. 3

4. 4
This was 6 years ago ??????

5. 5
GAN Style Transfer
https://thisxdoesnotexist.com/

6. 6
Do you remember Prisma?

7. 7
Where are we
now?

8. 8
Lots of Generative AI Models for Image Generation

9. 9
Interesting Applications

10. 10
Interesting Applications

11. 11
Interesting Applications
Alternative: https://interiorai.com/

12. 12
How do Models
Work?

13. 13
Stable Diffusion Models
●Text to Image Generation
●Training
○Noising: Add random noise to
images
○Denoising: Learn to remove
noise
○Guided by Text Embeddings
○Latent Space Representation

14. 14
Stable Diffusion Models
Inference
●Start with Random Noise
●Iteratively, reduce noise with
guidance from text
embeddings
●Image matches text input over
time

15. 15
Let Us Build

16. 16
Virtual Camera
●Location
●Weather
●Surroundings from
OpenStreetMap
●Inspired by Paragraphica

17. 17
Fine Tuning Image Models
●Generate Personalized Images
●Few Example Images (5-20)
●Low-Rank Adaptation(LoRA)

18. 18
Inpainting
●Replace parts of Image
●https://huggingface.co/spaces/
ameerazam08/FLUX.1-dev-Inp
ainting-Model-Beta-GPU

19. 19
ControlNet
●Replace parts of Image
●Parts based on existing
content
●https://huggingface.co/spaces/
hysts/ControlNet-v1-1

20. 20
Many More Applications
●Outpainting
●Upscaling
●Restore Old Images
●3D Characters
●Run Models Locally
○ComfyUI
○DiffusionBee

21. 21
Key Takeaways

22. 22
Observations on Generating Realistic Images
●Look at examples of prompts & images from the community
○https://prompthero.com/
●Good idea to use an LLM to refine the prompts
○https://ai.gock.net/flux
●More steps leads to better results, especially with Flux.1 models
●Running locally needs a lot of RAM
●Cloud is quite cheap
●With smaller models
○Text is still problematic
○Some features like hands

23. 23
Ethics

24. Confidential and Proprietary. Do not distribute without Couchbase consent. © Couchbase 2025. All rights reserved. 24
●How can we Identify AI generated content?
●How can we combat Deep Fakes?
●How can the Models get properly Licensed content?
●Reduce Bias in Training data


Challenges

25. Confidential and Proprietary. Do not distribute without Couchbase consent. © Couchbase 2025. All rights reserved. 25
AI Generated Videos?

26. Confidential and Proprietary. Do not distribute without Couchbase consent. © Couchbase 2025. All rights reserved. 26
AI Generated Worlds + Games

27. Confidential and Proprietary. Do not distribute without Couchbase consent. © Couchbase 2025. All rights reserved. 27
AI Generated Games

28. Confidential and Proprietary. Do not distribute without Couchbase consent. © Couchbase 2025. All rights reserved. 28
●Virtual Camera
○https://github.com/nithishr/image_gen_experiments/tree/m
ain/virtual_camera
●Flux Fine Tuning
○https://github.com/nithishr/image_gen_experiments/tree/m
ain/flux_fine_tuning
○https://replicate.com/docs/get-started/fine-tune-with-flux
●Flux Inpainting
○https://huggingface.co/spaces/ameerazam08/FLUX.1-dev-In
painting-Model-Beta-GPU
●ControlNet
○https://huggingface.co/spaces/hysts/ControlNet-v1-1
●QR Code
○https://antfu.me/posts/ai-qrcode
Resources

29. 29
Q&A

30. 30
GANs
●Generator & Discriminator competing
●Fast
●Mostly specific purpose like Person
Generator, Objects, etc