Extra Quality: Curt Newbury Studios Stefi Model

| Metric | Description | |---|---| | LPIPS (Zhang et al., 2018) | Learned perceptual similarity (lower = more similar). | | SSIM | Structural similarity (higher = better). | | FID (Heusel et al., 2017) | Distributional distance to real‑image manifold. | | Human Quality Rating (HQR) | 1‑5 Likert scale, collected from 120 professional photographers. | | Aesthetic Preference Score (APS) | Weighted average of APS‑Model (CLIP‑Aesthetic) predictions. |

Figure 1 (below) illustrates STEFI’s modular architecture, comprising three inter‑locking components:

| Component | Function | Novelty | |---|---|---| | Multi‑Scale Texture Prior (MTP) | Learns a bank of 64 texture embeddings (e.g., fabric, metal, skin) extracted from a curated 2 M‑image corpus of high‑resolution macro shots. | Enables dynamic injection of fine‑grained texture at inference. | | Dynamic Attention Gating (DAG) | A transformer‑based cross‑attention block that modulates latent diffusion steps based on prompt semantics and selected texture priors. | Prevents over‑saturation of texture information, preserving global composition. | | Quality Amplification Loss (QAL) | Composite loss: • LPIPS‑Weighted Fidelity (λ₁) • Texture Consistency (TC) via Gram‑matrix divergence (λ₂) • Aesthetic Score Regularizer (ASR) using a fine‑tuned CLIP‑Aesthetic model (λ₃). | Explicitly drives the network toward “extra quality” as measured by both low‑level fidelity and high‑level aesthetic judgment. |

Training Details

Inference Pipeline

Original Curt Newbury studio prints of Stefi (physical silver gelatin photographs) have sold for upwards of $12,000. Digital "Extra Quality" asset licenses, due to their rarity, routinely trade for $3,000–$5,000 on secondary markets.

Before analyzing the model itself, one must understand the studio that created it. Curt Newbury Studios is not a high-volume production house. It is an atelier—a workshop where the lines between industrial design, fashion photography, and figurative art blur. curt newbury studios stefi model extra quality

Founded on the principle that a model (whether physical or digital) should evoke emotion through anatomical precision, Curt Newbury has spent decades perfecting a specific aesthetic: the balance between idealized form and natural imperfection. Unlike mainstream studios that rely on CGI smoothing or overly stylized mannequins, Newbury’s work focuses on texture, micro-detail, and lifelike presence.

The studio is renowned for:

The notion of explicitly amplifying quality was explored in “Perceptual Quality Enhancement via Dual‑Loss Optimization” (Gao et al., 2022), which combined a structural similarity loss with a learned aesthetic scorer. Nevertheless, the resulting models still exhibited occasional over‑sharpness and halo artifacts. | Metric | Description | |---|---| | LPIPS (Zhang et al

In an era of smartphone snapshots and rapid-fire social media posts, true extra quality stands out. For this shoot, we pulled out all the stops:

| Model | LPIPS ↓ | SSIM ↑ | FID ↓ | |---|---|---|---| | STEFI (full) | 0.112 | 0.938 | 12.4 | | STEFI (w/o QAL) | 0.139 | 0.925 | 18.7 | | SD‑XL | 0.158 | 0.902 | 24.1 | | MJ‑6 | 0.162 | 0.898 | 26.5 | | DE‑3 | 0.174 | 0.887 | 31.2 | | TG‑1 | 0.149 | 0.914 | 20.3 |

Interpretation: STEFI reduces perceptual distance by 12 % relative to the strongest baseline (SD‑XL) and improves structural similarity by 3.9 %. The Quality Amplification loss contributes ~15 % of the LPIPS gain. Inference Pipeline

| Configuration | LPIPS | SSIM | HQR | |---|---|---|---| | Full STEFI | 0.112 | 0.938 | 4.62 | | – MTP (random texture) | 0.138 | 0.927 | 4.31 | | – DAG (fixed attention) | 0.129 | 0.932 | 4.48 | | – QAL (only LPIPS) | 0.139 | 0.925 | 4.19 | | – All (baseline diffusion) | 0.158 | 0.902 | 4.12 |

Ablations confirm that each component contributes synergistically; the QAL alone yields the largest single‑component gain in perceptual quality.