Sign In
homemodelsimagesvideos3D Modelshubsarticlescomicschallengesupdatesshop

Tiers of AI Image Models for Photorealistic and Stylised Art: A Qualitative Personal Perspective

2

informedviewz's Avatar

informedviewz

Sep 10, 2026

(Updated: a day ago)

musing
ai model comparisonstylisedphotorealistic
Tiers of AI Image Models for Photorealistic and Stylised Art: A Qualitative Personal Perspective

This is a live (WIP) document that shows my personal qualitative opinion about the effectiveness of different AI image models, primarily in terms of quality of visual rendering and aesthetics, as well as prompt following and adherence. This is a subjective perspective based on using different models in online generators and in ComfyUI. The percived rankings are obviously limited: In photorealism, the focus has been on computational documentary photorealism and cinematic film still. In stylised arts domain, the focus has been mostly on digital/phantasy illustrations, graphic novel styles and comics. I have listed the tiers and rankings in separate categories: Commercial (API) versus Open models, listed separately for photorealism versus stylised art. The reason for separate ranking of Open vs. Commercial model is: Out of the box released open-weight checkpoints are usually far from competitive against commercial models, whereas community fine-tunes and workflows developed by the community for these open models can push their quality to a level higher than best commercial models in specialised application domains. Only the newer generation models that operate based on natural language prompts (as opposed to the tags soup) are taken into account.

Commercial Models for Photorealistic Art:

Tier A:

  1. ChatGPT Images 2.5 Flare/Sunburst (High/Medium Quality)*

  2. ChatGPT Images 2.0 (medium quality)

  3. Nano Banana Pro/2

Tier B:

  1. Seedream 5.0 Pro

  2. Groke Imagine Images 2.0

  3. Muse Image

  4. ChatGPT Image 1.5

Tier C:

  1. Wan Image 2.7

  2. Flux 2 Pro

*ChatGPT 2.5 seems to suffer from occasional "hiccups" (e.g. distorted limbs) when trying to deal with prompt following versus quality, especially in high-quality mode (like the high-quality mode issues in GPT Images 2.0). This makes it necessary to do more trial and errors to get the right image.

Commercial Models for Stylised Art:

Tier A:

  1. Nano Banna 2/Pro

  2. ChatGPT Images 2.5 Flare/Sunburst (Medium Quality)

Tier B:

  1. Seedream 5.0 Pro

  2. Groke Imagine Image 2.0

  3. ChatGPT 2.0

Tier C:

  1. Muse Image

  2. Groke Imagine Image 1.5

Open Models for Photorealistic Art:

Tier A:

Krea 2 Turbo Ecosystem. Note-worthy fine-tunes: Lustify! Krea 2 (v10), CielBleu Krea 2 (v1), Sick Ollie Krea 2 . Or When used with proper LoRAs.

Tier B:

  1. Qwen Image 2512. (when used with proper LoRAs)

  2. Flux 2 Klein 9B. Note-worthy fine-tunes: Flux 2 Klein 9B True (V3). Or When used with proper LoRAs.

Open Models for Photorealistic Art (WIP):

Tier A:

Krea 2 Turbo+Raw+LoRA Ecosystem.

Tier B:

  1. Qwen Image 2512. (+ LoRAs)

  2. Flux 2 Klein 9B/Base + LoRAs.

  3. Anima (?)

For a more general and quantitative assessment of AI Image models. There are excellent articles by the community, as well as formal recognised leader boards:

https://arena.ai/leaderboard/text-to-image

https://artificialanalysis.ai/image/leaderboard/text-to-image

Thanks for reading, and any comment is very welcome.

T.B. Continued ...

@informedviewz

2