-- v2 --
Better facial variety
Easier prompt aging
General Improvements (there's a lot of little tweaks over v1)
Better prompt adherence (not perfect at all... we're only at v2 here... keep the faith.)
-- v1 --
Boy this one was a PITA to train.
Because this was brought up in the comments, I'll post this here too:
It's been tough to dial some attributes in with the new Illustrious XL 1.0 release. I have been working on that. Illustrious was originally designed for anime/illustration where key distinguishing features of age are more related to physical size. Many anime characters have very similar faces - and so training for Illustrious 1.0 (link) has more challenges in (or at least more challenges for me) with:
1. Variety of faces
2. Dialing in Age
3. Things like blemishes, wrinkles, etc...
4. Masculine features (most anime characters - even male - tend to be more effeminate looking)
So this is definitely something that is going to take a few generations to dial in. I did have some success with getting aged appearances with A LOT of weighted prompts that described aged features - like wrinkles or scruffy-ness etc...
For this image, I added:
Old woman, grey hair, wrinkled face, glassy stare, very old, wise, wearing a sweater,
Noteworthy is the result of months of work trying to see if I could train Illustrious to do what I wanted. Illustrious is a bit of a bugger to train for (at least for me), it took longer than I thought. I did recompile my work for the Illustrious 1.0 base. (some of the original work was done on the 0.1 base). This is definitely a 1.0 version. There are things that I'm happy with and things that still bug me a little. I'm hoping the feedback I get from the community will help me dial this in for future versions.
It's important to know that Illustrious kind of has a "base face" and you have to prompt your way out of it. It also tends to be harder to prompt for male subjects. You'll have to wrestle with it a bit.
Things to know about Illustrious.
At least with the recent Illustrious 1.0 base, LoRAs can be a bit of a bugger to work with. I've tried to mitigate some of that with fine-tuning... but I'd keep in mind that LoRAs that were originally trained on the Illustrious 0.1 base, may take some extra arm-twisting to get results with. Illustrious is kind of notoriously difficult to work with on Realistic modeling and training, so keep that in mind. If you're just looking for the best realism, you're probably going to find that with SDXL. You can try my QuadPipe model or my MegaChonk models for that.
When it comes to settings for this one, I've had good results with a wide spectrum of CFG 3.5 - 18.5 have given different and solid results. It depends on your prompts of course.
I have found that going too heavy on prompt tags like:
high-quality
ultra-detail
etc...
These tags are probably aren't necessary and they definitely confine what it produces. Again, sometimes you have to twist its arm a bit, but you can get there. I did find that Negatives (especially media-type) give you an extra boost here on realism.
Check my prompts to learn about settings etc... but in general this is what I was using when I created the demos:
768x1152
CFG 3.5-18.5 (super wide, but also it follows prompts at higher CFG without going crazy)
4x_foolhardy_Remacri (upscaler)
30 hires steps
0.27 Denoising Strength
Upscale by 1.5
Sampler: DPM++ 2M SDE Heun
Schedule: Karras
Sampling Steps: 50
Description
FAQ
Comments (17)
small tip to match the age ? adult, middle age, old, senior?
It's been tough to dial that in with the new Illustrious XL 1.0 release. I have been working on that. Illustrious was originally designed for anime/illustration where key distinguishing features of age are more related to physical size. Many anime characters have very similar faces - and so training for Illustrious 1.0 has more challenges in (or at least more challenges for me) with:
1. Variety of faces
2. Dialing in Age
3. Things like blemishes, wrinkles, etc...
4. Masculine features (most anime characters - even male - tend to be more effeminate looking)
So this is definitely something that is going to take a few generations to dial in. I did have some success with getting aged appearances with A LOT of weighted prompts that described aged features - like wrinkles or scruffy-ness etc...
Testing now, the , model seems to struggle with the prompt even with high cfg, struggle with eyes as well
I'd be interested to know what your sampler/scheduler are and number of steps etc...
If you feel comfortable enough to share the prompt you're using here, I'd be happy to try it to see if it's something inherent with the Illustrious XL 1.0 base (new base) or if it's something with the fine-tuning.
@QuadPipe sure, I was planning to upload examples tomorrow to show what I meant lol, but I'm using your recommendation, everything but 25 steps and goings for 5.5 cfg like your examples to 7, but for example:
1girl,Wonder Woman, looking back at the viewer from behind, smirking, pulling down the edge of her panties with 1 finger,tall body, standing, dreamy atmosphere, intense warm colors, vivid colors, detailed background, dramatic lighting, vivid colors, saturated
I'll add that training on top of illustrious-1.0 was far more work than training on top of Illustrious-0.1(which is what most of the checkpoint merges and LoRAs are based on).
Right now, I have mixed feelings about Illustrious, simply from a standpoint of how much extra force is required to get it to do what I want (from a training perspective).
I know that this model will get better, I appreciate you trying it out and giving feedback!
@QuadPipe no worries, take all the time you need, I'm the meantime I will keep playing with your xl checkpoints Wich are super good
@P_Universe Running that now. Ok, I think I see what you're talking about.
Everything renders well (wonder woman - looking back etc...) but hand movements aren't behaving specifically according to prompt.
I ran this at both my standard:
DPM++ 2M SDE Heun + Karras
50 steps
CFG: 5.5
Upscaler: 4x_foolhardy_Remacri + 30 Hires steps + Denoising Strength: 0.27 + Upscale by: 1.5
Hires CFG: 5
896wx1152h
and
DPM++ 2M SDE Heun + Karras
25 steps
CFG: 5.5
896wx1152h
There's definitely an improvement with the upscaler - here's a link to the upscaler I'm using.
So - I found a way to get this to do what you want, and it has to do with the natural language methods of Illustrious prompting. One of the big features of Illustrious is to have Flux-style prompting OR tag based prompting. I got this to do what the prompt suggested but only after turning the prompt around a bit.
Positive:
view from behind. with her back to viewer. A smirking Wonder Woman is pulling down the edge of her panties with 1 finger. looking at viewer.
Negative:
bad proportions, low resolution, bad, ugly, bad hands, bad teeth, terrible, painting, 3d, render, comic, anime, manga, unrealistic, flat, watermark, signature, worst quality, low quality
Even so, it took a lot of generations and I think it must be something specific to Wonder Woman and how the images that were originally used in Illustrious might have been tagged. I did not specifically train the action in your prompt, so it would be falling back on the core training. I'll continue to take a look at this and see if the problem spans across other scenes etc... I'll think about a fix for the next version.
@QuadPipe damn thanks for that big reply, also noticed some lack of characters like this example: 1girl, Tifa Lockhart, posing against the wall, raining, hand behind head, armpit, sad, tummy, dynamic pose ,vibrant color palette, motion smearing, in motion
I noticed also that the model so far loves to undress the chest
And here, it was hard to do the action, even tho there is no a special character: 1girl,purple_hair, purple_eyes, mask,crop_top, underboob, jacket, black_gloves,thighhighs, torn_clothes, street, alley,fishnet_stockings, dark_aesthetic, leather _belt, choker, long_hair, fingerless_gloves, lifting shirt, hand on shirt, teasing gesture, provocative_pose, seductive_action,cyberpunk, goth glowing_eyes, vivid colors, fishEye pov, fishEye view, light and shadow
But I will try to play more with natural prompt to see is there are bigs differences
Weird faces & anatomy
I have said it in my review, but i feel this should be visible here too: I have been playing a bit with Rillusm and Noteworthy deliver so much more! Thanks for your hard work!
dope
Will you continue to train on this basis to avoid the problem of women looking very uniform? It has this problem when I have fewer prompt words and no defined race, features.
Illustrious has this challenge in the base Illustrious XL 1.0 model as far as I can tell and I have been doing some work to break away from that. I will continue to update this model. Right now, the base model seems to require a lot of specific prompting to get away from that.
Others may have had different experiences, but that's what I encountered while training this model.
I have found that a v1 model tends to help identify the major problems so I can work on improvements for v2. Things get better as we figure out what we need to work around.
.
even if imperfect, I'm a fan of the effort and progress you made
Thank you! Working on v2 - hoping to have some improvements out later in March, 2025 (month/date for context later)