Hey I can help if you need

I love this project and would gladly help you with running any tests. :)

Can confirm this version works way better with Loras! More accurate outputs as well.. Great job!

Much improved. Good work!

The current version live on Civitai is amazing! People usually prefer overly beautiful youthful faces for their generations but me working in Ad/Design industry - faces with imperfections, less beautified ones are bread and butter. I will make sure to test the updated model but there is a massive chance I will keep using the one I currently have.

(Please refer to my Image post; in this post I'm solely focusing on embedding and beauty portrait as those are what I'm more familiar with. I'll leave LoRAs and Lycoris for someone else.)

Summary: Right off the bat V5fix is already looking much better than V5. All the masculine bias is significantly reduced (although it's still there imo). V5fix has better contrast, and a reduced level of haze at low CFG. With that said I still prefer V4 for portrait generation. V4 produces better looking female portraits, with more dynamic poses, and most importantly to me, better likeness when a TI is applied.

These are the impression I'm getting with regards to V5fix vs. V4:

V5fix has higher contrast and reduced haze: This change is perhaps the most crystal clear. It's not better or worse, it's a change in aesthetic, and I do love this change. Imo the skin texture isn't necessarily better in V5fix vs. V4, but the increased contrast does make it seem like so. This change comes not without consequences as it makes people look slightly older as well.

V5fix seems to still have a slight bias toward a particular look: A masculine look. A square face, bigger jaw, longer chin, smaller eyes, and smaller separation between the eyes. Even with Lizzy Caplan who has this particular look, V5fix exaggerates it. It is important to note that these effects are significantly reduced in comparison to V5, but the minute differences are still there. And multiple minute differences can compound to create a noticeable difference.

V5fix poses are more rigid: See sample images. All V5fix generations conform to a particular pose. V4 is more flexible and creates more dynamic poses. Although there are many things that could affect this and it may not be a bug but a feature. Perhaps V5fix just listens to the prompt better, or it follows the aesthetic of a particular seed better (all images were generated with same seed with the exception of hannahowo).

As someone who mainly makes TIs I can only speak for myself as my interest is very particular. V4 suits my needs better. It listens to my prompt like no other checkpoint. When I throw a TI into the prompt, I KNOW it will generate a realistic depiction of what the SD1.5 base model would. So much so that it became my go-to checkpoint for TI likeness testing.

For the integrity of the V4 version to be kept intact, whatever process that is making V5fix more bias toward a particular look has to not only be minimized, but eliminated. Sure the boost in contrast and skin texture is great. But no amount of that benefit could trump the realistic rendition of real people, whose unique facial features deserve to be highlighted accurately. Personally, I don't want AI generated blemishes as much as I want real looking people.

So far the V5fix is already doing a MUCH better job than V5, and we only have you to thank for that Mr. SG. I admire your contribution to the community, 20 revisions sounds like an awful amount of work. Hopefully this provides some feedback if you do choose to move forward with this project.

i just started testing and it seems this one respect my lora, i used like 6 lora at once .. zoom slider and other "slider" + my own made loras, and it gave me good result, and i used controlnet openpose , and it was working fine, thank you a lot for this
btw i used (Code_755_noVAE_fp16-no-ema.safetensors)

@SG_161222 Iv just tested it, my god you've somehow brought back the magic and amped it up 10x as good!! my lora's now work beautifully and characters are extremely accurate! SG I salute you and give you a virtual high five for figuring it out you're a masterclass and i want you to know how much we respect and cherish your incredible work. (used Code_755_VAE_fp16-no-ema.safetensors)

First of all, I admire your commitment to making quality models. I tested my custom Loras with VAE, worked well with ControlNet, TIs, and Loras. Loving the contrast in this test model and I don't need to use the add_detail Lora anymore which I did for v4. Something I tried and made some progress in fixing the square jaw bias - when generating Asian female portraits, I counter any square jaw/ masculine features by adding (toned face) in the positive prompt and man in the negative prompt with 1.3 weight. The square jaw bias is still there but nothing that can't be fixed with prompting. I am trying a few ideas, will update here. Overall, I really like this iteration over v4 and the previous release of v5 for sure. Can't wait for the magic you'll bring to SDXL :)

I tested the new fix and I can say that for me it didn't made much difference, I can see that the fix respects a little more the prompts than the V5 base model, but besides that I could not find difference enough to be able to say wich is better, one thing that amazes me is the capability of upscaling from both models, for people upscaling it too much makes the images kind of too soft, but for animals with fur it does an incredible work.
Anyways both models are amazing, I can't say that the fix is better.

@leonnn1 I used Invoke ui for my tests. The new changes to Automatic1111 give me acceptable but weird results.

hi, on the files section the one with file name "Pruned Model fp16 (1.99 GB)" is it the same as the v5? only less in size?

Thanks so much for your hard work! I have the feeling the code_1284_vae follows my prompts better than code_755_vae and RV50_vae using Drawthings on iPhone 14pro max. My rating so far: RV50_vae = good, code_755_vae = better and code_1284 my top favorite of all models i have ever used on my phone!

So I've been working with 1284 and 755 (both are bounds better than the original 5.0)

And I've been trying to determine which is better by producing about 5 30 image batches (15 each) at 768x768, all generating pictures of young women.

This resulted in 1284 being the clear winner because it's:

Slightly less cleft chin (seems like the last remaining overly masculine trait)

Moderately more likely to produce a more conventionally attractive and feminine face

Less likely to produce overly large craniums

Consistently more appealingly balanced lighting

.

The only negative tradeoff I noticed of using 1284 over 755 is that pupil deformation seems mildly more frequent. I personally find the benefits to significantly outweigh this and will be using 1284 going forward.

>I see skin pores on big boobs
>happy
simple as. Take your time and have fun

Code 755 with ADetailer also gives nice results, saves alot VRAM, without highres fix. Will try 1284 thanks.

I will keep this short but THANK YOU for all you do and your dedication to fix this problem. No more weird changes to the faces of trained famous people. Much better contrast. Skin tone looks amazing, and everyone looks beautiful again!

1284 > v4 >>>>>>>>>> 755 >>>>>>>>>>>>>>>>>> v5

Only negative thing to say about 1284 is that it adds too much makeup to peoples' faces? I need to test this more.

Code_755 works very nicely for me. Here's a stress test with 277 frames of animation:

https://civitai.com/posts/442010

That latest one is such an improvement!! Best one so far. Just one thing I've noticed. The face quality does go down a bit when you go full body or when you do a 16:9 aspect ratio standing up.

Is v5.1 the same as Code_1284?

@SG_161222 Sweet. I've been getting better results with Code 1284 but didn't really compare it extensively with v5 or 755. Is DPM++ 2M Karras with 30 steps alright? Or is Euler a / SDE Karras better