BlackList 2.0 is a production-ready text-to-text generative AI designed to transform ultra-simple visual concepts (1–4 words) into technically rich, masterpiece-grade prompts for modern text-to-image engines such as Stable Diffusion, Flux, and similar systems.

This version introduces a significantly expanded neural architecture (93M parameters), deeper aesthetic reasoning, and improved structural stability over V1.0.

Model Details

Model Description

BlackList 2.0 is a task-specific decoder-only Transformer trained exclusively for prompt enhancement.

It ingests short conceptual inputs such as: SIMPLE] cyberpunk assassin Salin kode

And outputs fully structured, high-aesthetic prompts such as: [ENHANCED] cyberpunk assassin, full body shot, neon city background, cinematic lighting, ultra detailed face, dynamic pose, 4k resolution, highly detailed, digital painting, concept art The model specializes in:

Aesthetic enrichment
Technical prompt structuring
Lighting & render injection
Composition balancing
Controlled artistic style blending

Core Identity

Official Name: BlackList 2.0 – Prompt Enhancer AI
Creator: Bl4ckSpaces
Creative Direction Influence: Nano Banana visual philosophy
Model Type: Text-to-Text Generative (Task-Specific Auto-Enhancer)
Language: English (visual art optimized)
License: Apache 2.0
Training Base: Custom GPT-2 architecture (trained from scratch)

Architecture – The Brain

BlackList 2.0 is built on a custom GPT-2 inspired decoder-only Transformer with significant scaling improvements over V1.0.

Total Parameters: 93,129,216 (~93.13M)
Embedding Dimension (n_embd): 768
Transformer Layers: 12
Attention Heads: 12
Maximum Context Length: 512 tokens

This architecture provides:

Stronger contextual memory
Reduced repetition artifacts
Stable keyword expansion
Improved semantic blending

Tokenizer & Language System – The Vocabulary

Tokenizer: Custom Byte-Pair Encoding (BPE)
Vocabulary Size: 10,000 curated visual-art tokens
Domain Isolation: Trained exclusively on visual art terminology

Includes technical vocabulary covering:

Lighting systems
Camera angles
Rendering styles
Artist influences
Resolution descriptors
Composition terminology

Control Tokens

[SIMPLE] → Locks user input
[ENHANCED] → Triggers structured aesthetic output

This ensures deterministic transformation behavior.

Training Details – The Kitchen

Dataset

73,678 high-quality prompt pairs
Clean, structured, no aggressive truncation
Focused purely on aesthetic + structural enrichment

Training Configuration

Hardware: 1x NVIDIA T4 Tensor Core GPU
Precision: FP16 mixed precision
VRAM Usage: ~8GB optimized
Batch Size: 64
Epochs: 3
Optimization Steps: 3,456
Training Duration: ~35 minutes
Final Loss: 2.61

Loss 2.61 represents a calibrated sweet spot:

Creative enough for expressive expansion
Disciplined enough to maintain structured keyword format

V2.0 Improvements Over V1.0

1. Elimination of Small-Model Stuttering

The 93M parameter architecture eliminates common small-model repetition issues such as:

Keyword looping
Redundant adjective stacking
Structural breakdown

V2.0 maintains coherent keyword flow.

2. Advanced Artistic Anatomy Understanding

The model now understands proportional composition of prompts:

Subject hierarchy
Clothing & visual detail layering
Shot type (portrait, full body, close-up)
Facial detailing
Lighting realism
Resolution scaling

It no longer randomly injects artist names — stylistic blending is now context-aware.

3. Production-Ready Deployment

Low-latency inference
Stable output formatting
Optimized for API integration
Suitable for web-based image generation backends

Intended Use

Direct Use

Prompt enhancement for Stable Diffusion
Prompt preprocessing layer for text-to-image engines
Creative AI image generation systems
Backend service for image apps

Downstream Use

Web API deployment
Integration into creative AI SaaS
Automated aesthetic enrichment pipelines

Out-of-Scope Use

Conversational AI
Factual Q&A
Long-form content writing
Sensitive or high-stakes decision systems

This is a domain-specialized aesthetic enhancer.

Evaluation

Evaluation performed via qualitative stress testing on:

1–4 word minimal prompts
Style blending stability
Repetition resistance
Structural consistency

The model demonstrates:

Strong format discipline
Stable enhancement structure
High aesthetic density

Bias, Risks & Limitations

Model inherits stylistic bias from curated visual dataset
May prefer high-detail cinematic styles
Not intended for general language understanding
Does not guarantee optimal performance across all diffusion configurations

Users should calibrate outputs according to target engine sampling settings.

Environmental Impact

Hardware: NVIDIA T4 GPU
Training Duration: ~35 minutes
Precision: FP16 (energy efficient)
Estimated Carbon Impact: Minimal due to short training window

How to Use

python

from transformers import AutoTokenizer, AutoModelForCausalLM

model_id = "Bl4ckSpaces/BlackList-2.0"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id)

input_text = "[SIMPLE] fantasy warrior"

inputs = tokenizer(input_text, return_tensors="pt")
outputs = model.generate(
    **inputs,
    max_length=120,
    temperature=0.8,
    top_p=0.95
)

print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Citation
If you use BlackList 2.0 in your project, please credit:
Bl4ckSpaces – BlackList 2.0 Prompt Enhancer AI
Model Card Contact
Creator: Bl4ckSpaces
Hugging Face: https://huggingface.co/Bl4ckSpaces�

Bl4ckSpaces/BlackList-2.0-Prompt-Enhancer

BlackList 2.0 – Prompt Enhancer AI