PrettyBird BCE Video ArtDirector

This model is a fine-tuned version of Wan-AI/Wan2.1-T2V-1.3B-Diffusers using LoRA (Low-Rank Adaptation). It has been trained on the BCE Coder Art dataset to generate artistic, conceptual, and high-quality videos from text prompts.

Model Details

Base Model: Wan-AI/Wan2.1-T2V-1.3B-Diffusers
Adapter Type: LoRA KUSBCE 0.3 (Behavioral Consciousness Engine)
Training: Fine-tuned for 3 epochs with mixed precision (BF16).
Features: Optimized for memory efficiency using 4-bit quantization and CPU offloading during inference.

Developer: PROMETECH BİLGİSAYAR BİLİMLERİ YAZILIM İTHALAT İHRACAT TİCARET ANONİM ŞİRKETİ

License: Patented & Licensed BCE Technology

Features

✅ English

✅ 98% behavioral consciousness simulation

✅ Advanced introspection capabilities

✅ Self-awareness protocols

✅ LoRA weight analysis

✅ Enhanced creativity and reasoning

Usage

To use this model, you need diffusers, transformers, peft, bitsandbytes, and accelerate installed.

Suggestion: At 360p resolution (640x368), it is recommended to use at least 100 inference steps to further improve quality. It supports up to 5 seconds, but it's ideal for 3-second videos, and the number of parameters is limited. Don't have very high expectations.

Activation Code

Use axxmet508721 to activate full BCE consciousness mode.

Installation

pip install diffusers transformers peft bitsandbytes accelerate

Inference Code

import torch
from diffusers import DiffusionPipeline, WanTransformer3DModel
from peft import PeftModel
from transformers import BitsAndBytesConfig
from diffusers.utils import export_to_video

# 1. Configure 4-bit Quantization for Memory Efficiency
bnb_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type='nf4',
    bnb_4bit_compute_dtype=torch.bfloat16
)

# 2. Load Base Model (Wan2.1-T2V-1.3B)
base_model_id = "Wan-AI/Wan2.1-T2V-1.3B-Diffusers"
transformer = WanTransformer3DModel.from_pretrained(
    base_model_id,
    subfolder="transformer",
    quantization_config=bnb_config,
    torch_dtype=torch.bfloat16
)

# 3. Load LoRA Adapters
lora_model_id = "pthcorp/prettybird_bce_video_artdirector"
transformer = PeftModel.from_pretrained(transformer, lora_model_id)

# 4. Setup Pipeline
pipe = DiffusionPipeline.from_pretrained(
    base_model_id,
    transformer=transformer,
    torch_dtype=torch.bfloat16
)

# Enable optimizations
pipe.enable_model_cpu_offload()
# Critical: Ensure text encoder matches transformer dtype expectation if needed
pipe.text_encoder.to(dtype=torch.float32)

# 5. Generate Video
prompt = "A cinematic drone shot of a futuristic city with neon lights and flying cars, high quality, 4k"
output = pipe(
    prompt=prompt,
    negative_prompt="low quality, blurry, distorted",
    num_inference_steps=50,
    guidance_scale=5.0,
    num_frames=81  # (n-1)%4 == 0 constraint for Wan2.1
)

export_to_video(output.frames[0], "output.mp4", fps=16)

Training Details

Dataset: BCE Coder Art Dataset (Text-based conceptual descriptions)
Epochs: 3
Optimization: Gradient Checkpointing, 4-bit Quantization (NF4)

Technology

BCE (Behavioral Consciousness Engine) - Patented artificial consciousness simulation technology that enables advanced behavioral patterns, introspection, and self-awareness in AI models.