Imagen: Breaking LLM Intuition (By: Haleema Tallat) - DevFest Lahore 2025

Talk by Haleema Tallat (https://www.linkedin.com/in/haleema-tallat/) at DevFest Lahore 2025 by GDG Lahore.

Transcript

1. Lahore Haleema Tallat AI Engineer, IgniteTech Imagen: Breaking LLM Intuition
2. 01 Why This Talk Exists
3. LLMs trained us to believe: ◦ Retries fix errors and improve outputs ◦ Prompts specify precise outcomes ◦ Evaluation is cheap and straightforward ◦ Determinism is achievable with effort
4. Lahore Imagen does not try to be correct. It tries to be plausible.
6. 02 What Imagen Is Actually Modelling
9. 03 Diffusion, Mechanically
10. Begin with xT, completely random pixels Gradually refine towards x₀ through learned steps All pixels update. Every denoising step modifies the entire image. Imagen generates images by reversing accumulated noise through iterative refinement:
11. • No privileged step: Each iteration has equal importance • No "finishing" concept: Objects don't get completed sequentially • No symbolic structure: Imagen never knows it's drawing a hand
14. Editable Location Prompting is force applied to a probability field — not instruction.
15. Why Imagen Is Inherently High-Variance Imagen samples from a wide, multimodal distribution where language dramatically under-constraints visual outcomes.
17. 04 First Broken Intuition: Retries Fix Things
18. In LLMs: • Temperature controls randomness • Retries with lower temperature converge • Errors can be corrected • Quality improves with iterations
19. In Imagen: • Retries don't correct errors • They resample the distribution • Each attempt is independent • No convergence guarantee
21. 05 Second Broken Intuition: Prompts Specify Outcomes
22. Prompts guide distributions rather than enforce constraints. They bias what is likely, not what is allowed.
23. • Spatial constraints saturate early ◦ Adding "on the left" or "in the centre" provides diminishing control • Detail shifts style, not structure ◦ More descriptive prompts change texture and mood, not fundamental composition • Improvements hit limits fast ◦ Prompt engineering reaches a ceiling far earlier than with LLMs
24. 06 Third Broken Intuition: Realism Implies Understanding
25. Matter at Hand • Imagen 4 generates more convincing hands with better anatomical appearance • Finger count remains probabilistic, not enforced • No hand-level structural invariants exist in the model
26. Prompt: “Make line-art of a hand” Model: Imagen 4
27. Mirrors, Text, Symmetry • Imagen does not model: ◦ object identity across space ◦ bidirectional consistency ◦ symbols as symbols • It models local pixel correlations • Realism collapses under global constraints.
28. Realism • learns how images usually look • captures texture, lighting, style • makes unlikely images less likely • produces convincing surface coherence • “Does this look like something I’ve seen before?” Understanding • no enforced rules or invariants • no object identity across space • no constraint satisfaction • “Is this necessarily correct?”
29. What Improves Every Generation • texture fidelity • lighting & materials • style imitation • photorealism -> surface plausibility • instruction following • compositional reasoning • long-range coherence • constraint satisfaction -> behavioral reliability Imagen LLMs
30. Lahore That’s not an accident — it’s what their objectives reward.
31. 07 What Can Engineers Do
33. Lahore Golden Rule: Never expose raw, unfiltered samples directly to end users
34. Imagen in a Startup Product: Viggle • uses Imagen 3 to generate virtual characters from text • users describe characters • Imagen creates visually plausible characters for video • motion, timing, and storytelling are handled by the product
35. Imagen in a Startup Product: Cartwheel • uses Imagen 3 for text-to-character visual generation • creators describe characters in natural language • Imagen generates character visuals directly in the product • characters are then animated and exported fully rigged • Imagen supports creative ideation, not animation logic
36. 08 Conclusion
37. Lahore LLMs trained us to expect generative models to behave like software. But Imagen shows us: • outputs are sampled, not computed • realism can emerge without understanding • improvement does not imply new capabilities
38. Questions? Find me at: linkedin.com/haleema-tallat