[August 2026 Edition] Major Local Image Generation Models and the Position of Krea2

A new image model called Krea2, which had its open version released on June 22, 2026, suddenly became a hot topic and has taken its place as one of the major local image generation models.

With so many local image generation models available, I will summarize and discuss the major local image generation models as of August 2026, including my own biased opinions on what position Krea2 has occupied and which models have become obsolete.

2026/8/10: As of August 2026, the evaluation of Krea2 has been finalized, so predictive writing has been changed to definitive writing (no new models added, Mage-Flow is...)
2026/7/25: In the description of FLUX.1 Krea [dev], corrected "based on Krea1" to "based on FLUX.1"

Currently Major Local Image Generation Models

List Table

This is a summary of the major local image generation models as of August 2026, based on my own biased opinions. While there are many others not included, I have prioritized the following models.

• Models supported by Forge Neo

• Models that were just released recently and whose evaluation is not yet settled

• Models with unique features or use cases

The criteria for the "NSFW" item is that the base model meets the following conditions.

• Yes: Has learned NSFW images and can also generate them

• No: Has not learned NSFW images

• Limited: Has learned NSFW images, but generation is restricted

Personally, with the arrival of Krea2, I have stopped using Qwen-Image-2512 and Z-Image-Turbo.

Elements Required for Local Image Generation Models

The conditions for a model to achieve "hegemony" in local image generation are that it must satisfy all three of the following (I assert).

• Ability to generate NSFW images (at least having learned them)
  For SFW images, cloud models are easier, higher performance, and higher quality, so there is no need to run them locally. Even if the base model has not learned NSFW, LoRA or merged models claiming NSFW support are often released, but you cannot achieve significant expression with individual-level additional training. I have heard many times that "you cannot output XXX NSFW images with QIE-Rapid-AIO." I think QIE-Rapid-AIO is a wonderful model for an individual development level, but the variations that can be added by LoRA merging are limited. The minimum requirement is that the base model, or a fully fine-tuned model, has learned NSFW.
  

  
  

• Light operation and resource consumption
  Unless it can generate 1024x1024 class images within one minute (within 10 seconds for upper-middle class or using acceleration technology) in an entry-class gaming PC environment, specifically with 32GB of main memory and a Geforce RTX 3060 (VRAM 12GB) to RTX 5050 (VRAM 8GB), the number of users will be limited.

• Ease of LoRA creation
  Since the base model alone has limits on the variations of images it can generate, there is no potential for development unless you can create LoRA at an individual level.

While there are other important factors such as image expressiveness, prompt adherence, and licensing, all currently released models meet these at a high level, making them difficult to use as differentiating factors.

Krea2

Features of Krea2

Krea2 satisfies all three elements required for the aforementioned local image generation models.

• NSFW image generation
  Unlike Flux, Qwen-Image, Z-Image, etc., Krea2 is also trained on NSFW images and is merely censored to prevent their generation. In that respect, it is similar to the cloud service Grok. Custom nodes and LoRAs to bypass censorship are being researched, and new ones are released almost daily. It seems there is some performance degradation, such as the inability to output Japanese or vertical text due to bypassing censorship, but the quality of the NSFW images themselves is high.
  

• Lightweight operation and resource consumption
  The Krea2 bf16 model has a total size of about 32GB for DiT and TE, but basically, you can use an 8-bit quantized model, which has a total size of about 16GB. Even so, it is large, but thanks to ComfyUI's memory management, it can be generated without issues even with 8GB of VRAM if you have 32GB of main memory. Regarding generation speed, generation with the TURBO model at 8 steps is the standard, which is a realistic speed for the model size. Also, with the implementation of INT8 ConvRot support around the same time, high-speed generation is possible even on Geforce RTX 30 series with minimal quality loss.
  

• Ease of LoRA creation
  musubi-tuner was quickly updated to support Krea2, so the difficulty of creating LoRAs is on par with other local models. The number of LoRAs for Krea2 on Civitai is also increasing. However, training seems to take time commensurate with the model size.
  

The position Krea2 has come to occupy

• Photorealistic images
  It is particularly strong in generating photorealistic images. Photorealistic image generation has shifted from FLUX.2, Qwen-Image, and Z-Image-Turbo to Krea2.

• Original 2D images, anime, and illustrations
  The number of 2D copyrighted characters and art styles that can be generated is overwhelmingly smaller compared to Illustrious or Anima. However, since the basic performance of 2D image generation is higher than Anima, the use of Krea2 might increase for original 2D images and major copyrighted characters. Also, as for erotic content, it is currently far from reaching Illustrious or Anima, which were trained on hardcore 2D erotic images from Danbooru. Therefore, unless a fine-tuned model like Illustrious for SDXL 1.0 is released,
  the status of Illustrious and Anima will remain unshaken..

• Edit-based
  Krea2 does not have an official Edit function like Qwen-Image-Edit. However, unofficial Edit functions and compatible models such as Krea 2 Identity Edit (ComfyUI-Krea2Edit) and ComfyUI-Krea2-Ostris-Edit have been released. Although the quality seems high for being unofficial, the evaluation has not yet been settled as of August 2026. Considering the ease of introduction,
  I don't think it's enough to replace QIE-2511 or FLUX.2 [Klein] yet.

Evaluation in Design Arena

Design Arena Krea2 is also highly rated on the evaluation site called Design Arena.

In the ELO Rating based on blind tests of generated images, Krea2 is slightly inferior to Ideogram 4.0, but it has obtained an outstanding score among open-weight models (local models). On the other hand, regarding generation time, Krea 2 Turbo is more than three times faster than Ideogram 4.0.

In June 2026, Ideogram 4.0, which was released about three weeks earlier, was a bit of a hot topic, but Krea2 has completely overtaken it in popularity.

That is all for Krea2.
The following is a summary of Hugging Face links and personal thoughts for each model in the list.

SD1.5 / Stable Diffusion 1.5

Released in October 2022

Personally, I only used derivative models like AbyssOrangeMix3 for a few hours when I first started image generation. I still have the images I generated back then, and they are quite nostalgic, filled with what we call the 'masterpiece face.'
It is said that the detailed depiction of backgrounds and patterns is better than SDXL, and some people still use it today.

SDXL

Stable Diffusion XL 1.0

Released in July 2023

It serves as the base for excellent derivative models such as Animagine, Pony, and Illustrious. Since it leans toward photorealism, some popular photorealistic models today are based on SDXL 1.0 rather than Pony or Illustrious.

Pony Diffusion

Released in January 2024

It was the dominant model until Illustrious appeared, but after Illustrious took over, it feels like a model exclusively for furries. Since v7 failed spectacularly, a comeback is likely impossible.
Personally, I only keep WAI-ANI-PONYXL v14.0 along with LoRAs for Pony.

Illustrious XL

Released in September 2024

Still very much in use. Or rather, future models will continue to become more sophisticated, and I don't think we will see another model like Illustrious that produces 'nice' images with so few tags, so I believe Illustrious will continue to be used forever.
Illustrious is very popular, but OnomaAI is disliked because their management was handled too poorly. It is unlikely that Illustrious XL v3.0 or later will ever be open-sourced.
Personally, I use WAI-illustrious-SDXL v16.0 as a benchmark, but it seems many people prefer v14.0.

NoobAI XL

Released in November 2024

A model fine-tuned based on Illustrious XL v0.1, characterized by having v-pred models that are strong in light, shadow, and color expression (there are also e-pred models).
Currently, most major Illustrious derivative models have also merged NoobAI, so most models that call themselves Illustrious are actually NoobAI as well. It makes me wonder when I see a v-pred model listed with Illustrious as the Base Model on Civitai, but since Illustrious also apparently became v-pred starting from v2.0, it's not something to worry about too much.
Personally, I occasionally use Obsession v-pred v2.0.

FLUX.1

There are so many variations that I don't know which is which. I wasn't generating images when it was first released, and I haven't used it or its related models, so I have almost nothing to say about it.
I think its role has ended with the release of FLUX.2 and Qwen-Image.

FLUX.1 [dev]

Released in August 2024

FLUX.1 Kontext [dev]

Released in June 2025

I only know that it seems to be an edit model.

FLUX.1 Krea [dev]

Released in July 2025

It was reportedly developed jointly by Black Forest Labs and Krea AI, based on Flux.1. It has no direct relationship with Krea2. You might need to be careful not to confuse it with FLUX.1 Krea [dev] when researching Krea2 on the web or with chat AIs.

Lumina-Image

Neta Lumina and its derivative model, NetaYume Lumina, were once expected to be the successors to SDXL as anime models. However, while they had some good points, they were basically inferior in quality to Illustrious and remained a work in progress. Since Anima has been released, they will likely not be used in the future. I have never used them either.

Lumina-Image 2.0

Released in February 2025

Neta Lumina

Released in May 2025

Qwen-Image

Qwen-Image / -2512

Released in August 2025

I think it shared popularity as a photorealistic generation model with FLUX.2 and Z-Image. Or rather, I think none of the models could take the lead simply because they lacked a decisive element (NSFW). I think it is safe to say that it has now been replaced by Krea2.

Qwen-Image-Edit / -2509 / -2511

Released in August 2025

This is the first edit model that can be used properly locally, and I think it still has many users. While FLUX.2 [Klein] is evaluated as being slightly higher performance, it seems that the transition has not progressed much.
The downside is that 2D images have a strong stylistic bias, making them difficult to handle.

Personally, I found the cosplay LoRA (photorealistic conversion LoRA) interesting, and I enjoyed comparing multiple LoRAs.

Qwen-Image-Layered

Released in December 2025

It is a one-of-a-kind local model that can analyze and decompose the layers of an input image, but it is not being used because image generation users do not have a need for layer decomposition in the first place.
Professional designers would likely use high-performance paid tools or cloud services, so the target user base is unclear.

Z-Image / Z-Image-Turbo

Released in September 2025

An honor student model that is not bad for both photorealistic and 2D images, and runs lightly. Conversely, it lacks any particular standout features, and I feel it lost popularity because the SOON™ (coming soon) scam for Z-Image (so-called Z-Image-Base) delayed the development of the LoRA creation environment.
I feel that because it has little stylistic bias, it is suitable for redrawing after upscaling.
I think it is safe to say that it has now been replaced by Krea2.

FLUX.2

Like FLUX.1, there are so many variations that I don't know which is which. Regarding Klein, I didn't even know for a while that it had been released or that it was an edit version. I wish they would make it clear by name, like Base / Turbo / Edit in other models.
At the time of release, it was evaluated as having decent quality but being too heavy, so I have continued to ignore it, including Klein, up to this point.

FLUX.2 [dev]

Released in November 2025

The expressive power seems quite high, but it is too heavy, so there are almost no users. The fact that Forge Neo does not support it also reflects this.

FLUX.2 [Klein]

Released in January 2026

As an edit model, FLUX.2 [Klein] itself is evaluated as being superior to Qwen-Image-Edit, but as mentioned above, I don't think the number of users is increasing because the marketing for the FLUX.2 series is too poor.

Anima

Released January 2026

As it is currently the most popular model, I think it needs no introduction.
While other local models are becoming massive by enhancing overall performance and features such as photorealism and text output, I believe Anima's success lies in its minimal feature and performance enhancements, its lightweight nature, its specialization in 2D output, and its full support for NSFW content.

Ernie-Image

Released April 2026

I believe this was the first local model capable of producing proper Japanese text.
I often saw people having fun creating silly Japanese poster advertisements or magazine covers that could previously only be produced via cloud services. However, there is no practical use for individuals, and for actual business use, cloud services offer higher quality.
Furthermore, a common feature of all local models released since then is their strength in photorealism and Japanese text, making it difficult to differentiate based on those features. It has been completely buried by the successive release of late-coming models with similar characteristics.

HiDream-O1-Image

Released May 2026

Expectations were high as it was developed by HiDream, which has a solid reputation as a photorealistic image generation service, but the quality was too low compared to the company's cloud service and other local models.
While future version updates are anticipated, a comeback from this point will be difficult.

Ideogram 4.0

Released June 2026

Released slightly before Krea2, this model shares many features with it.
As introduced at the beginning, although Ideogram 4.0 is slightly superior to Krea2 in terms of generated image quality, it is significantly inferior in generation time, and it seems it can only produce light NSFW images, so it is currently losing buzz and popularity to Krea2.

Boogu-Image 0.1

Released June 2026

I think there is a need for it, such as the official Edit version, but it was released between Ideogram 4.0 and Krea2, and the timing was too poor. It hasn't generated much buzz.
Regarding the Edit version, the consensus seems to be that the quality is not high enough to switch from QIE-2511 or FLUX.2 [Klein].