Fixing Identity Drift in AI Image Generation with a Deterministic Constraint Layer (Minimal PoC Inside)
Everyone building image or multimodal applications knows this pain:
- you generate the same character twice → the face changes
- change the angle → body ratio collapses
- add more iterations → drift becomes uncontrollable
- and the model “just does whatever it wants”
People call it randomness.
But it’s actually a missing structure problem.
Today I’m sharing a small, fully reproducible PoC that stabilizes:
✅ identity
✅ body ratio / geometry
✅ scene semantics
✅ lighting context
❌ without fine-tuning
❌ without LoRA
❌ without prompt hacks
Just a deterministic constraint layer.
✨ Why this PoC works
Most generation pipelines have:
- a model
- a prompt
- a hope that it behaves
This PoC adds a missing piece:
4. A deterministic constraint layer
The constraint layer defines:
- identity anchors
- ratio + geometry targets
- allowed semantic drift
- forbidden alterations
- reproducible prompt assembly
- stable configuration YAML
The model keeps its creativity,
but the “rails” prevent drift.
🧩 Core idea
Control doesn't come from more parameters — it comes from structure.
The model predicts.
The constraint engine limits the space of acceptable predictions.
This hybrid gives you:
- stable identity
- consistent proportions
- fixed scene context
- reproducible runs
- simple code anyone can extend
📦 GitHub Repo (Minimal, Safe, No Hidden Tricks)
👉 https://github.com/yuer-dsl/vertex-deterministic-agents-poc
This is a small, practical PoC — not a full framework.
I designed it so normal engineers can read the code in minutes.
Files include:
-
structured_multimodal_constraints.yaml -
constraint_loader.py -
multimodal_prompt_builder.py -
deterministic_agent_stub.py -
parse.py
Everything is deterministic and reproducible.
🔧 When should you use this?
- building a multimodal app
- stabilizing an image pipeline
- doing character-consistent content
- prototyping product features
- improving reproducibility for demos
- replacing fragile prompt engineering
If you ever felt:
"This model is strong, but impossible to control."
Then the deterministic layer solves exactly that.
👤 Author
Yuer — independent engineer exploring deterministic AI, multimodal stability, and structured reasoning layers.
💬 Open to discussion
If you want a version for other models (Claude, OpenAI, PixArt, stable-diffusion, etc.), just drop a comment or open an issue in the repo.
Let’s make AI powerful AND predictable.
Top comments (0)