@Goofy_Ai can i ask what you used for lighting specifically in v1? Its so good ive not seen it reproduced like that anywhere else

hope one day to see mature base prefect version

大佬你好，可以问下这个底模训练lora的参数是多少吗？

Sooo, nothing changed? Description doesn't tell anything

Thank you to continue supporting IllustriousXL :)
I have a question : for lora creation, is it time to switch to Illustrious XL 2.0 (OnomAI) or should I stay with Epsilon-pred 1.1 (NoobAI-XL) ?

And here I came

Nice model, thanks for the update.

My personal test result for 5.0 vs 4.0:

*5.0 clearly has higher color saturation and contrast.

*5.0 has more flat anime style baked in than 4.0, which is similar to 3.0.

*5.0 generally has more details on background and surroundings, but it also adds extra subjects or person into the background that you may or may not want.

*5.0 has less lora support due to its own strong style backed in.

*5.0 does not contain the entire newest danbooru dataset, just as 4.0.

For me personally, I may not import this 5.0 version into my flat anime workflow, instead I would simply use the wai15 I already get used to and seemed to have better lora support. The 4.0 (or its former version2.0) is still my main model for now.

5.0 delivered! Better lighting, body texture improvements, training data more versatile, small errors cleaned up and a few improvements to NSFW.

Below I'll list my early impressions after a day or two of use. I apologize for the lengthy wall of text, but I hope my limited observations can help further this models accuracy, so that everyone can feast from the spoils!

Also to Goofy, If you happen to come across this comment, I would urge you to differentiate the model names somehow, since this, in my humble opinion should be considered a continuation of 3.0, while right now people may end up with a few question marks if they directly compare it with 4.0.

⚠️Disclaimer: These are my observations from my personal test environment which uses my own custom artist mix, as well as Lighting control and other LoRA's.

I will start with the positives:

1. More mellow lighting with good contrast.

The model is now comfortable with lower lighting environments. The lighting doesn't feel as if a floodlight was targeted onto the girl, rather enshrouds her in way which allows for the skin tone to appear more natural.

2. Body details on female characters have seen changes.

(Preference based, but I think they're better now). One very noticeable one is the veeeeeery iconic reflective spot of light on womens breasts. That has been significantly reduced.

3. Error rate seems to be lower across the board.

A notorious issue for me was teeth. They would at times show up a bit chipped or otherwise damaged. Now that has been nearly completely fixed. Other issues that would stem from limited screenspace have been alleviated by the models improved handling of wider field of view which I cover in 5.

4. Watermark exclusion seems to be better? (Though this will vary a lot based on training data and artist styles).

5. Composition is significantly better.

While this may annoy some people, I find that tag adherence is good, but the models training data is more confident with working with wider compositions, which allows for more panoramic images where a broader field of view can be used. The lessened necessity of the model to cram everything up close also means that there will be a higher deviation in images that use the same prompt data. A rather "professional" feature that maybe more casual users can find cumbersome at times. I greatly appreciate it.

6. LoRA compatibility is better.

This will require significantly more in-depth testing, but my initial testing shows a blend of marginal improvement and neutral results. Nothing broke, and some got better. The ones that didn't get better, remained the same.

7. Better expression data.

There's a wider range of facial expressions that use the same tag. Could be very anecdotal on my end.

8. X-ray effects are better.

Both LoRA'd ones and default.

Now I will cover the negatives:

1. Text seems to want to creep up more often than before.

It could be just be character or franchise specific, but speech bubbles are popping up slightly more than before (which was already fairly rare). I will curate my UC to minimize this going forward.

2. Eyes are less expressive and detailed.

This one is the biggest issues I've had so far, and unfortunately, it's a rather big one. Below I have explained my workaround, but it's a resource heavy workaround.

Also it could be just how my artist mix works with the training data. But in general a few effects around the eyes, as well as makeup seem to be less prominent this time around. Some issues are solved with increasing the weight of tags, some not. The layering of the eye socket is simpler and eyeshadow doesn't seem as high quality as 3.0.

As a solution I'm running a dual checkpoint workflow with 3.0 that handles face and eye bbox Adetailers.

3. Male data seems simpler.

Faceless men don't look as cool anymore. Before they'd at least retain some small features like more prominent cheekbones, jaw shading, small beards etc. Not a huge loss, but can be noticeable a bit.

4. Clothing textures seem to be more 2D and have heavier outlines.

Not all, but some transparent articles of clothing, or lacey undergarments are a bit simplistic in their coloring and texture quality this time around. While not an egregious issue, this can be mostly finetuned.

I appreciate everyone who bothered to read the insights of some random person like me on the internet.💖

I personally will switch to 5.0 as my main checkpoint and use a dual checkpoint workflow to handle the minimal issues I have with it.

Do you have a list of artists I can use in promts?

Dear highly educated people, which version is the best? 🥰

遠坂凛大好き

Clearly the best checkpoint around. Goofy you madman.

Anyone notice that 5.0 seems to do much less detail and "simpler" style than 4.0? At least with the same parameters it feels like a bit of a downgrade for simple usage.

4.0 plz come back

need 4.0 back. 5.0 seems like a downgrade in quality :(

like e.g hands / fingers feels worse on 5.0

5 changed too much.. everything seems to be flatter, less detailed and more cartoonish :/

Another new version that is somehow a downgrade. 2 and 4 have been the only serviceable models!

please bring back v4, v5 is clearly much worse

5 is a downgrade, bring 4 back

I have found that 4.0 and 2.0 produce more detailed, less flat images than 5.0 and 3.0, I don't know who's paying all this buzz to keep five and three on site