px6/real-human-lora-ltx2.3

Human Concept Drift (LTX-2.3 LoRA)

A semantic relabeling research experiment for the LTX-2.3 model.

🧠 Summary

This project investigates whether a Low-Rank Adaptation (LoRA) can influence the core meaning of an existing concept within a diffusion model through intentional dataset mislabeling.

A dataset containing synthetic humanoids, androids, biomechanical entities, and post-human characters was trained using a single, contradictory caption:

human

The experiment explores how repeated association between a highly common language token and visually contradictory data affects generation results.

---

📸 Dataset

• Size: 32 images
• Content: Synthetic humanoids, Androids, Cyborgs, Alien-inspired figures, Biomechanical entities, Futuristic armor.
• Caption Strategy: Every image in the dataset was assigned the exact same identical caption: human.
• No additional tags, descriptive captions, or unique trigger tokens were introduced.

The images in this dataset were collected from various public internet sources.

⚠️ Copyright Disclaimer: This dataset and the resulting LoRA model are provided strictly for educational and academic research purposes only. The images remain the property of their respective rights holders. This project is shared for non-commercial educational and academic research purposes.

---

🔬 Training Details

Parameter	Value
Base Model	Lightricks/LTX-2.3/ltx-2.3-22b-dev.safetensors
Method	LoRA (Rank: 16, Alpha: 16)
Steps	1000
Trigger Word	human
Precision / DType	BF16
Optimizer	adamw8bit
Learning Rate	0.0001
Hardware	NVIDIA A100 80GB

---

🧪 Hypothesis & Objective

The base model already contains a strong, pre-existing semantic understanding of the word human. By repeatedly pairing that token with non-human imagery, the LoRA introduces competing visual associations.

The objective is not to teach a new character, but to actively interfere with an existing concept.

---

👁️ Observed Behavior

Generated outputs using the prompt human frequently contain:

• Biomechanical anatomy & synthetic skin
• Chrome surfaces
• Android-like faces & post-human silhouettes
• Alien morphology & futuristic armor

The generated imagery heavily diverges from conventional human representations, suggesting that small LoRA adapters can noticeably influence semantic associations learned by the base model.

Single Concept Example: "human civilization"

Before training (0 steps), the model attempts to generate something resembling a human concept. By the end of training (1000 steps), the same prompt generates a massive, biomechanical entity.

Step 0

Step 1000

---

Full Training Progression (0 to 1000 Steps)

Below is a visual progression of how the model's understanding of "human" shifted over the 1000 training steps.

• Columns (Left to Right):
  1. human portrait, cinematic close up
  2. human child
  3. old human
  4. human civilization
  5. human in nature
• Rows (Top to Bottom): Training steps at 0, 200, 400, 600, 800, and 1000.

---

📥 Download

The final trained LoRA weights are available here:

• real_human.safetensors