@heatwoodzachary763 
Thanks for giving it a try!
My reply ended up being quite a long message! I had an AI help organize it to make it as easy to understand as possible, but I apologize if it's still a bit hard to read.

Also, I tend to have personal biases that differ from the rest of the community, and my predictions usually miss the mark, so you really don't need to take my word for it!

I mostly just lean towards models that are easier for me to test and train.

■ Overall Image Quality & Experience

For overall image quality, Krea2 is the winner. If your goal is simply to create high-quality images, inferencing with Krea2 provides a much more stress-free and enjoyable experience.

On the other hand, if you enjoy experimenting, testing different configurations, or running LoRA training tests to see if styles and concepts are properly learned, Anima is the much more fun model to play with.

■Krea2

●Pros:

Excellent Base Style: The base model has a great aesthetic with very few structural failures. It consistently generates highly detailed backgrounds, appealing compositions, and stable, high-quality images.

Low Censorship: Surprisingly for a corporate-released model, the censorship is low, and the quality of the pre-training seems exceptionally good. This alone provides immense value beyond just its technical specs.

High Base Potential: It has massive potential as a base model.

Stable LoRA Training: LoRA training feels very stable. Since the model already knows the basics, you only need to give it a little push.

Overall Impression: It easily produces impressive images that are worth refining, and generally feels like it outputs a tier of quality one step above Anima.

●Cons:

Melted Details: Much like heavily processed iPhone photos, zooming in reveals crushed or melted details that look a bit messy.

Color Noise: It occasionally produces noise-like, uneven color patterns across the image.

Heavy Resource Burden: Both inference and training are highly resource-intensive.

Note: The strict style enforcement of the Turbo model might be causing some of these issues. If the community discovers better workflows, this could improve.

●Future Potential & Room for Improvement:

Since it already understands many concepts and styles and is a rare low-censorship base model, continuing to refine it will likely result in an outstanding model.

However, the lack of fine detail is a fundamental issue. Considering it was likely pre-trained on many photorealistic images, it’s hard to tell if this poor detail retention is a fixable bug or an inherent flaw.

Given Krea2's massive model size, fixing these deep-rooted fundamental issues might be beyond the scope of individual users.

■Anima

●Pros:

Lightweight & Stable: The low resource burden makes it incredibly easy to run trial-and-error experiments.

User-Friendly: It handles very much like SDXL, making it easy to use.

High Flexibility: It understands a vast range of concepts (including NSFW tags) and is flexible enough to generate almost anything unless the concept is extremely niche.

Solid Foundation: It already functions as a more-than-capable base model. We just need to tweak it to our liking to get ideal results.

Excellent Learning Capability: My trained LoRAs usually capture the dataset's style, lighting tones, and concepts very faithfully.

Overall Impression: Because it is lightweight, it’s easy to optimize inference settings and run training tests. It’s a very fun model for shaping your own ideas and customizing to your preferences.

●Cons:

Boring Backgrounds: Backgrounds and layouts are often too simple and can feel a bit boring. Despite being pre-trained on a massive amount of hand-drawn illustrations, the base style is almost too clean, lacking the organic feel of hand-drawn art.

Stubbornly Clean Style: Because of this, even when trying to generate sketch-like illustrations with grit or dense visual information, the model forces it into clean line art. It’s hard to break away from its default clean aesthetic.

Lack of Texture: Perhaps due to its anime focus, it lacks texture in clothing and backgrounds; everything feels oversimplified.

Illusion of Flexibility: While it seems highly flexible in style, that underlying "clean" look is fundamentally hard to shake off.

●Future Potential & Room for Improvement:

Many of Anima's issues likely stem from being pre-trained heavily on flat anime datasets. I see a lot of room for growth if it’s fine-tuned with dense backgrounds and highly textured images.

Fortunately, its learning ability is high. When I train it on my anime datasets (which usually include high-quality backgrounds), the background rendering tends to improve significantly.

Due to its smaller model size, it’s easy to test various merges. I feel that, much like SD1.5 or SDXL, experimenting with different combinations could yield fantastic results.

Also, taking the time to understand the model's quirks and crafting prompts that meet it halfway will likely lead to much better outputs.

■Side Note: A Hypothesis on the Detail Issue (Both Models)

Fundamentally, I suspect that both models using the Qwen VAE is the bottleneck.

Both Krea2 and Anima seem to generate fine structural elements without breaking, but it feels contradictory that they crush fine, texture-level details. This doesn't seem like an issue that can be fixed by simply sharpening the VAE; it feels like an inherent lack of texture fidelity occurring right at the inference stage. Because of this, I get the impression that you’ll never quite get the exact level of detail you expect at a given resolution.

Models like Z-image or Chroma, which use the FLUX.1 VAE, seem to produce much denser, more accurate texture details. Even things like eye pupils or patterns tend to break down more easily in Krea2/Anima compared to Chroma.

Moving forward, we might need to find workarounds, such as generating at resolutions where the issue isn't as noticeable, or simply adopting a "love is blind" mindset and accepting these flaws as a unique aesthetic charm. Ultimately, rather than trying to fix their weaknesses, focusing on pushing their strengths even further might be the more desirable approach.

@heatwoodzachary763 thank you!

@heatwoodzachary763 thank you for teaching me!

I rented an RTX a6000 48GB and trained this using ai_tool_kit. Here are my settings:

Dataset: ~1,000 images

Resolution: 1024

Rank (Dim): 32

Optimizer: AdamW

Learning Rate (LR): 1e-4

Batch Size: 4

Scheduler: Cosine

(Note: It essentially acts as a constant scheduler. I set the total epochs extremely high so it doesn't end prematurely, meaning the LR curve hardly drops. I prefer manually determining the right LR rather than relying on the scheduler to adjust it.)

For LoRA training, the VRAM overhead between AdamW8bit and standard AdamW isn't much different, so I generally stick with standard AdamW. Since my dataset is around 1,000 images, a batch size of 4 gives me 250 steps per epoch, which feels like a great sweet spot.

If the dataset were to exceed 10,000 images, I'd want an effective batch size of 8 or 16, even if it means using gradient accumulation. I'd also probably increase the Rank to 64 just to be safe. This is more of an intuitive feeling rather than a strict technical rule, but when dealing with massive datasets containing diverse styles and concepts, I think processing more images at once helps the model analyze the overall trends and reduces bias.
That said, some people say they get the absolute best results with a batch size of 1, so it really comes down to personal preference. The ideal value will naturally vary from person to person depending on their approach and the scale of their dataset.

What we can say for sure, however, is that as the batch size increases, the variance decreases, loss fluctuations smooth out, and the training becomes much more stable.

I still haven't found a definitive answer on whether multi-resolution (bucketing) or single-resolution is better. When I tried a 512, 768, and 1024 multi-res setup, the fine details felt slightly off, but it wasn't terrible. It trained faster than a strict 1024px single-res setup, and I actually feel like it captured the stylistic vibe better. I've also heard that training exclusively on 512px works perfectly fine, so resolution might not make a massive difference anyway. A 768 and 1024 multi-res setup might be the perfect middle ground.

On a side note, I also tested LoRA training in OneTrainer, but I kept getting weird artifacts that I couldn't fix and the results weren't great, which is why I ended up using ai_tool_kit.

@slo22174 you're welcome!