| --- |
| license: apache-2.0 |
| library_name: diffusers |
| tags: |
| - text-to-video |
| - dit |
| - diffusion-transformer |
| - education |
| - zulense |
| --- |
| |
| # 🧠 DiT (Diffusion Transformer) Fine-Tuning Experiments |
|
|
| **Core Backbone for the [Zulense Z1 Foundation Model](https://huggingface.co/zulense/z1)** |
|
|
| This repository hosts the **Diffusion Transformer (DiT)** checkpoints trained to generate educational video content. These models operate in the latent space of our [Causal VAE](https://huggingface.co/ProgramerSalar/causal_vae_checkpoint) and are responsible for the temporal consistency and logical flow of the generated math lectures. |
|
|
| ## 📂 Model Ledger & Performance |
|
|
| We are releasing the training logs to demonstrate the optimization curve of the "Imagination Engine." |
|
|
| ### 1. `finetune_2_pytorch_model.bin` (🌟 Production Candidate) |
| * **Role:** **The Z1 Foundation Backbone** |
| * **Status:** ✅ **Converged / High Fidelity** |
| * **Performance:** |
| * This checkpoint represents our stable run. It successfully learned to align temporal attention with the "teacher's movement" and "blackboard writing" logic. |
| * **Metrics:** Achieved target validation loss on the Class 5 & 8 Math dataset. |
| * **Behavior:** Shows strong temporal coherence (objects don't disappear randomly) and adheres to the physics of writing on a board. |
| * **Recommendation:** **Use this file** for all inference tasks related to Zulense Z1. |
| |
| ### 2. `finetune_1_pytorch_model.bin` (Experimental / Deprecated) |
| * **Role:** **Initial Warmup Run** |
| * **Status:** ⚠️ **Underfitted / High Noise** |
| * **Performance:** |
| * This was an early checkpoint where the model struggled to decouple the background (classroom) from the foreground (teacher). |
| * **Issues:** Resulted in "flickering" artifacts and poor text alignment. |
| * **Archived:** Kept here for research comparison to show the impact of our improved data scheduling in `finetune_2`. |
|
|
| ## 🏗️ Architecture Context |
|
|
| The Zulense Video Pipeline follows a two-stage generation process: |
| 1. **Stage 1 (VAE):** Compresses video into latents (See: `causal_vae_checkpoint`). |
| 2. **Stage 2 (DiT):** This model (`finetune_2`) acts as the denoising backbone, predicting the latent patches over time based on text prompts (e.g., *"Draw a triangle with 3 angles"*). |
|
|
| ## 💻 Usage (Loading Weights) |
|
|
| ```python |
| import torch |
| |
| # Path to the best performing checkpoint |
| model_path = "finetune_2_pytorch_model.bin" |
| |
| # Load weights (assuming standard DiT structure) |
| state_dict = torch.load(model_path, map_location="cpu") |
| |
| print(f"✅ Loaded DiT Backbone: {model_path}") |
| print(f"Tensor keys found: {len(state_dict.keys())}") |
