AI_Photo_Generation_with_Python_A_Developer's_Guide.pdf
0 likes92 views
The document is a developer's guide on AI photo generation using Python, presented by Nithish Raghunandanan at PyCon+Web 2025. It covers generative AI models for image creation, including techniques like stable diffusion and fine-tuning methods, and addresses applications like outpainting and inpainting, as well as ethical considerations regarding AI-generated content. Additionally, it provides resources and links for further learning on related tools and projects.
AI_Photo_Generation_with_Python_A_Developer's_Guide.pdf
- 1. Confidential and Proprietary. Do not distribute without Couchbase consent. © Couchbase 2025. All rights reserved. AI Photo Generation with Python: A Developer’s Guide Nithish Raghunandanan Developer Advocate PyCon+Web 2025
- 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 ● Live in Munich, Germany ● Like to Hack stuff :) @nithishr
- 3. 3
- 4. 4 This was 6 years ago 🤨
- 6. 6 Do you remember Prisma?
- 8. 8 Lots of Generative AI Models for Image Generation
- 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/ SkalskiP/FLUX.1-inpaint-dev
- 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 Games?
- 27. Confidential and Proprietary. Do not distribute without Couchbase consent. © Couchbase 2025. All rights reserved. 27 ● 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/SkalskiP/FLUX.1-inpaint-dev ● ControlNet ○ https://huggingface.co/spaces/hysts/ControlNet-v1-1 ● QR Code ○ https://antfu.me/posts/ai-qrcode Resources
- 29. 29 Q&A