It might be released just as people start thinking it’s outdated…

Architecturally, models like Neta-Lumina already seem more advanced and promising, especially with video models improving at text-to-image too.

I truly hope v4.5 won’t be overlooked when it's released—since it's a specialized architecture, it depends on the community to build the ecosystem for it to even run.

Thank you for sharing all the information.
Sorry for the long message, but I wanted to share my thoughts as well...

I haven’t used NovelAI v4.5 much yet, but it’s impressive that the vectors can go beyond 10 without issues.

Just a guess, but maybe there’s some kind of weight regularization like in A1111.

I also got the impression that Enhance was just doing basic i2i-style modifications when I tried it in the past.

I find NovelAI’s prompts, inference techniques, and tools like Vibe Transfer and Director Tools fascinating too—I'd love to replicate them locally, so I often wish they shared more papers or technical details.

NovelAI v4.5’s rendering is impressive. I suspect the U-Net + 16ch VAE setup offers both flexibility and detailed structure.

True, since NovelAI has tweaked the VAE before, they might have done something similar in v4.5.

The unpredictability of NovelAI v4.5 likely comes from the diversity of its large-scale training.

Local models like Pony or Illustrious often rely on biased merges or style LoRAs instead of base models, which makes them easier to control.

With closed models, you can’t adjust them the same way, so you need to learn their quirks.

It’s a bit of a shame—if we could use them locally, they’d probably be even better.

Also,Models like SD1.5 and SDXL seem to handle rough prompts well, which might be why some prefer NovelAI v3 over v4 or v4.5.

T5 or LLM-based models often need more precise prompts unless heavily biased like Flux.

Some models, not just NovelAI v4.5, are underrated simply because they don't perform well with simple prompts.

That said, without some guidance, it’s hard to craft good prompts. If NovelAI shared more about prompt building, it might help users avoid guesswork. Even powerful models lose their potential if kept closed, so listening to user feedback could really help.

Also,NovelAI is a closed model, but it started it all—and we usually follow in their footsteps. So rather than rejecting it, I think it’s valuable for the community to experiment, share insights, and find ways to replicate it locally. That was especially true back in the SD1.5 era with the leaks.

Neta-Lumina is promising—its potential may even surpass NovelAI v4.5.

Every architecture is a rough gem that needs refining, but that takes community effort.

Sadly, many get ignored early and forgotten in favor of what's already proven.

If people explore Neta-Lumina’s possibilities, even in small ways, it could become something great.

Here’s a bit of a side note, but my take on NovelAI:

Every version of NovelAI is highly refined—their fine-tuning is on another level.

They train only the U-Net until tag recognition is near perfect, without losing flexibility.

Their dataset creation and training methods are highly refined.

Even with great new base architectures, it's extremely hard for the community to surpass NovelAI through fine-tuning alone.

NovelAI was already impressive from v1, and v2 showed how far a small model like SD1.5 could go.

With v3, they explored optimal SDXL training early on—some had noticed the need for VPred-like methods, but the community largely ignored it until NovelAI revealed details.

I was deeply impressed by the NovelAI_v3 paper.

It used many improvements that a few had long advocated, showing they recognized their value early on—and proved their effectiveness through the model's fine-tuning quality.

VPred has only recently become widely recognized and practical.

With SDXL, unless you're using true VPred models like CosXL, Playground v2.5, or Terminus-XL, training for VPred can be tough—but projects like NoobAI are doing great.

I find much to agree with in the new architecture choices of v4 and v4.5.

Honestly, I like NovelAI v4.5’s architecture—it’s close to my ideal setup.

U-Net + T5 feels like a solid evolution, improving prompt understanding while keeping familiarity.

It’s more stable than switching to DiT, and personally, I still get better results from U-Net. The 16ch VAE is also a big plus.

I saw that as a solid and proper direction for evolution. Kolors, with its U-Net + LLM setup, seemed like a worthy SDXL successor, but it was sadly overlooked—dismissed due to licensing and its 4ch VAE. I still see that as a missed opportunity.

That's why I feel NovelAI consistently trains each of their models in a near-optimal way, and I truly respect that.

Many fine-tuned local models now produce high-quality images, but I often question whether their training, hacks, or base models were truly optimal.

but,Models like Neta Lumina, Pony v7, and Chroma—trained with LLMs or T5—seem like cleaner, more solid starting points with fewer fundamental flaws, so I have high hopes for them.

LovelaceA 

Thanks for the reply. I also hope for a bright future for local models.

I agree with your analogy—most people just mix or season pre-cooked dishes.

Many enjoy consuming, but few are interested in actually cooking.

I hope more people get into training, more architectures get explored, and we build a more diverse community with more model choices.

Right now, the focus tends to be too narrow, and new possibilities often get ignored—it’d be great if we could explore them together.

It's interesting how NovelAI v4.5 renders shapes and details accurately even at a distance.

I don’t know their exact architecture, and they likely don’t use the methods I’m thinking of, but as an example, models like Cascade use stepwise latent upscaling.

Cascade also has UltraPixel, which adds another latent upscaling step to improve detail—making images sharper even at the same resolution.

So enhancing latent information, not just output resolution, could also be a factor.

Thanks for sharing the article—it was really interesting.

You're right, NovelAI explored a different architecture on their own.

As I mentioned, I see v4.5’s U-Net + T5 + 16ch VAE as a solid evolution and close to the best current architecture.

The T5 article was interesting, especially since it offered a different view from mine.

Some say T5 struggles with NSFW due to censorship, but I haven’t found that to be true.

Even if there is some bias, it hasn’t affected training in my experience.

I fine-tuned PixArt-Sigma, which uses T5 only, and never had trouble with NSFW learning.

It’s a strong architecture that learns concepts very well—I was able to train almost all concepts in my dataset without issue, and others have too.

I honestly have doubts about TE training.

Even models like SD1.5 or SDXL with CLIP don’t really need it.

It can speed up concept learning as a hack, but it’s a double-edged sword with side effects.

That trade-off might be acceptable for personal, single-concept LoRA, but it's too risky for large-scale fine-tuning.

If early learning of unknown concepts is essential, combining it with large-scale textual inversion like Pivotal Tuning makes more sense.

Chroma and Pony_v7 don't use TE training either, and it hasn’t been an issue.

Even Flux LoRAs often cover NSFW concepts and are popular.

As a side note, lumina image 2.0 uses Gemma as TE, which is a censored LLM, but I believe it only uses intermediate outputs, so it’s fine.

NovelAI has never fine-tuned TE, and likely wouldn’t do something that irregular.

If T5 had serious issues, they'd likely switch to another LLM as the TE.

If I'm wrong, well... that's a bit embarrassing.

Sorry for the long message again, but I really enjoyed this meaningful exchange with you.

It was a great chance to organize my thoughts—thank you.

I understand the frustration—it’s natural to feel that way seeing how impressive their model is.

But in a way, I’m also grateful. They’ve always shown us the direction to aim for, and that can be encouraging.

I felt despair when SDXL was released, due to its massive parameters, the heavy load from high resolution, complex architecture, and the difficulty of training.

However, NovelAI v3 showcased such incredible quality that many people, including myself, felt it was worth pursuing the architecture.

As a result, the community took on large-scale training and grew to the point where it could create models that come close.

I believe v4.5 plays a similar role now—it’s something to strive toward. Even if we can’t match it directly, we can look for different paths, like video models, or new base architectures.

So I try to enjoy what I can do now, while staying excited about the future. Let’s both keep exploring and having fun with what’s ahead.Even if what I can do is small and limited, none of it is ever wasted—every experience and bit of knowledge becomes a valuable asset for the future.

This model already has a built-in VAE for NovelAI v2, so there’s no need to specify a different one.

I’m not sure what you mean by “very different,” but if the colors look more vibrant, that’s likely the intended result.In my experience, the NovelAI v2 VAE has better color than v1 and is closer to the look of vae-ft-mse-840000-ema. If you noticed a similar feel, the correct VAE is likely applied.

If you want results closer to my sample images, each one includes the prompt and workflow, so feel free to use them as a reference.

I also shared a clean, general-purpose workflow with the model.

There’s nothing particularly special, so you should be able to reproduce it even with WebUI.

Some images in the gallery were adjusted using DoRAs or embeddings listed below—those might help too.

https://civitai.com/models/1253884?modelVersionId=2070815

https://civitai.com/models/1809022?modelVersionId=2047219

https://civitai.com/models/1809022?modelVersionId=2061380

If you have any other questions or issues, feel free to ask!

If you're open to sharing your image in the gallery, I could use it as a reference and try generating something similar. It might help identify if a more tailored prompt is needed for good results with this model, and I’d be happy to test and share my findings.