// chapter 02
understand the technology
Diffusion, GAN, autoregressive token-based. Three architectures, each with implications for what a generator can do and what training data it needs.
architectures primer ->// independent reference :: rev 2026.04
The 2026 buying guide
to AI image
generators.
Because "best" depends on your use case, budget, and licence needs. This is a framework for evaluating any generator on the market, not a ranking of products that ages the moment a vendor reshuffles tiers.
// chapter 01
Six axes that separate one generator from another.
Listicles call this "features". We treat each as a measurable dimension you can score any generator on yourself, with published research where it exists and a self-test protocol where it does not.
read the framework ->01
#C6FF3D
resolution
native output size, upscaling, megapixel ceiling.
02
#B8A9FF
prompt adherence
what you wrote vs what the model produces.
03
#FF5B22
text rendering
readable text in the output. architecturally hard.
04
#7CFFA8
style control
presets, LoRAs, references, transfer APIs.
05
#FFB145
photorealism
lighting, materials, depth, photographic fine-tune.
06
#FF4A6E
subject consistency
character sheets, seed locking, adapters.
// chapter 03
understand the legal stack
Commercial use, copyright protection, and indemnification are three separate questions. Most guides collapse them into one boolean. That is how people get sued.
the legal stack ->// chapter 04
evaluate any generator
Fifteen questions covering capability, licensing, provenance, cost at scale, and exit. Print it. Reuse it on every new generator that launches.
the checklist ->// site.is()
What this site is.
- [+]An educational reference written like a buying guide for cameras: explain the technology, the dimensions, the legal stack, send you to vendors to verify today's specs.
- [+]Sourced from arxiv papers, vendor legal pages, and court filings. Every claim links to a primary source with a verification date.
- [+]Maintained quarterly. The /legal-landscape and /licensing pages get tighter cadence because case law and policy shift.
// site.is_not()
What this site isn't.
- [x]A ranking. We don't publish "top 5" lists, "winner" badges, or performance claims we cannot substantiate.
- [x]A pricing table. Prices move monthly. Every time you need a number we link to the vendor's own pricing page.
- [x]A gallery of sample outputs. We don't reproduce any named generator's visual signature.
// primary sources
What we cite, and why it matters.
// arxiv
verified April 2026Ho, Jain, Abbeel 2020
Denoising Diffusion Probabilistic Models. The paper that established the architecture behind Stable Diffusion, Imagen, and most current generators.
// arxiv
verified April 2026Rombach et al. 2022
High-Resolution Image Synthesis with Latent Diffusion Models. The Stable Diffusion paper.
// arxiv
verified April 2026Goodfellow et al. 2014
Generative Adversarial Nets. The original GAN paper, the previous-generation architecture.
// gov
verified April 2026US Copyright Office
AI guidance and registration policy. Human authorship required for registration.
// vendor
verified April 2026Adobe Firefly legal page
The vendor's own statement on training-data sourcing and indemnification.
// gov
verified April 2026EU AI Act
Training-data transparency obligations and the text-and-data-mining opt-out under Article 4 of the Copyright Directive.
// about comparisons
We don't do "X vs Y".
Here's why.
Most comparison content ranks named products on undemonstrable criteria like "best for text rendering" or "highest image quality". Those rankings are stale within a quarter and invite trademark complaints from the products ranked lower.
Instead, our /compare page explains the three meaningful ways generators differ, lists the commonly-discussed tools as a directory with links to their own pages, and stops there. If you arrived here from a query like "Midjourney vs DALL-E", what you actually want is "what is architecturally different and which fits my use case", which is the page we wrote.