| PROJECT OVERVIEW 
VIVA is an independent research and framework-development project exploring how professional visual judgment can be translated into a more structured, consistent, and explainable approach to evaluating AI-generated and AI-edited imagery.

​

Developed from my experience integrating generative AI into professional visual workflows, VIVA provides a method for identifying, categorizing, and communicating visual issues while evaluating whether an image successfully fulfills the requirements of its original prompt.

 

| WHY VIVA
As generative AI became part of my daily creative workflow, I found myself repeatedly identifying and correcting similiar types of visual problems, from anatomical and object errors to unrealistic environments, text failures, reference drift, and unintended changes during image editing.

 

The challenge was not simply recognizing that something was wrong. It was determining what was wrong, how significant it was, and whether the image still satisfied the original request.

​

VIVA grew from an exploration of that challenge.
 

| THE EVALUATION CHALLENGE
Generative AI images can appear successful at first glance while still containing errors that affect accuracy, realism, or prompt adherence. Evaluating these images consistently requires separating individual visual issues from the overall success of the image.

​​

| THE VIVA FRAMEWORK 

VIVA organizes visual findings into seven evaluation categories:​

• Anatomy

• Composition

• Color

• Environment

• Object

• Text 

• Other

 

Each finding is evaluated using a 1-5 severity scale, allowing minor imperfections to be distinguished from issues that significantly affect the success of the image.
​

A Prompt Requirement Audit provides an additional check against explicit, visually verifiable requirements within complex prompts, helping identify details that might otherwise be overlooked during evaluation.

​​

| VIVA IN PRACTICE - CASE STUDIES
The framework is being tested through a series of case studies examining different challenges in AI-generated and AI-edited imagery. Each applies the same structured methodology, testing a different aspect of prompt adherence, visual accuracy, or production usability.

REFERENCE PRESERVATION

Evaluating whether a requested image edit was successfully integrated without introducing unintended changes to the supplied source image.

​​
 

COMPLEX PROMPT ADHERENCE

Testing whether a dense, multi-element prompt can be systematically evaluated for complete visual requirement fulfillment.
 

​​​​

PRECISION EDITING + REFERENCE FIDELITY

Evaluating a production-style image edit against supplied source assets and an explicit deliverable requirement.

 

 

| OBSERVATIONS FROM DEVELOPMENT
Developing and applying VIVA has reinforced that visual AI evaluation requires more than determining whether an image looks convincing at first glance. An output can appear successful while still missing explicit prompt requirements, altering source elements that were meant to remain unchanged, or failing production requirements that affect its usability.

The case studies have also demonstrated the value of separating what is wrong from how much it matters. Categorizing individual findings, assessing their severity, and checking them against the original request create a more consistent and explainable evaluation than relying on an overall impression alone.
 

| PROJECT STATUS
VIVA is an independent research and framework-development project currently being refined through iterative testing and case-study evaluation. The framework is intended to explore how professional visual judgment can be translated into a structured methodology for evaluating generative AI imagery.

ViVA is not presented as a commercially validated system or finished industry standard. The methodology will continue to evolve as it is applied to additional image-generation and image-editing scenarios.​​​​