“`html
Real photo / image generated by chatgpt / image generated by Flux-Krea (by Drew Breunig)
AI Image Generation Shifts Focus: From Generic to ‘Opinionated’ Models
Table of Contents
San Francisco,CA – August 15,2024 – A new approach to artificial intelligence image generation is gaining traction,prioritizing distinct artistic styles over broad capabilities. Developed by Drew Breunig, a computational photographer based in Oakland, California, the “opinionated model” technique, exemplified by his Flux-Krea system, is challenging the current dominance of generalized AI image generators like GPT-4.1.
The Problem with Current AI Image Generators
Current AI image generators, while capable of producing technically accurate images from text prompts, often lack a unique aesthetic. They tend to produce images with a characteristic “AI” look – often overly bright, soft, and featuring artificial bokeh. This stems from being trained on massive, diverse datasets without a strong emphasis on a specific artistic vision.
Flux-Krea: A New Approach
Flux-Krea tackles this issue by first undergoing broad training to establish a strong foundational understanding of image creation. However, it then undergoes a refinement process focused on a highly defined aesthetic. Breunig describes this as cultivating the “signature” of a photographer.This focused training allows the model to consistently produce images with a recognizable style, even when given short or generic prompts.
The key, according to Breunig, is to avoid mixing numerous artistic styles during training. A diluted approach results in a “watery” and unfocused aesthetic. Instead, concentrating on a specific look allows for a more powerful and consistent output, reducing the need for extensive post-processing or complex prompting.
Comparative Results: Flux-Krea vs. GPT-4.1
In direct comparisons, both Flux-Krea and GPT-4.1 were given identical, detailed prompts. The results highlighted the differences in their approaches:
- GPT-4.1 generated “correct” images but exhibited the typical “AI” characteristics – excessive brightness, softness, and artificial bokeh.
- Flux-Krea produced more natural-looking portraits,believable urban scenes,and images with the feel of authentic snapshots.
The rise of Specialized AI Models
This trend signals a shift towards “opinionated models” – AI systems trained with a specific aesthetic or identity. This approach has potential applications across various creative fields, including animation studios, fashion brands, and individual photographers seeking to replicate their unique style. Breunig’s work suggests that specialization and customization will be crucial for advancing the quality and diversity of generative AI.
the progress of Flux-Krea,built using a custom training pipeline and approximately 200GB of curated image data,demonstrates the feasibility of this approach. The model is currently available for limited access through Breunig’s website (https://www.dbreunig.com/).
This move towards specialized models promises to not only improve visual quality but also restore creative diversity to the field of generative AI, mirroring the advancements seen in text-based AI and chatbots.
Via | dbreunig.com
Image | Marcos Merino through AI
