About:
I lost my job last week so more free time to improve this model! I have another one lined up so don't feel sorry for me.
I'm not happy with the hands in v2.0, I feel there was a backwards step there. Will try out some different strengths (and epochs) for my LoRas to see if that helps and maybe release a v2.1 in the next few days.
v2.0
New base blend:
Uncanny valley 'UncannyValleyilxl1.0+noob'
ILustREAL 'v1.0'
And a new custom trained realism LoRa with a far more diverse set than v1, as well as the custom LoRa from v1.0.
v1.0
My attempt to push an Illustrious v0.1 merge towards realism while maintaining strong prompt adherence and compatibility.
This is comprised of:
large part LobotomyMix v1.0 Perun (I love the aesthetics of this model)
small part Nova Reality XL Illustrious v1.0 (good quality realistic textures)
a custom trained LoRa for realistic textures and less exaggerated female body proportions (more work required here!)
Recommended Settings:
Settings used for the showcase images were:
Positive prompt:
{descriptive tags from DeepSeek-R1 e.g. 'attractive woman in dimly lit room, mild facial expression of embarrassment, soft warm lighting, simple stylish clothing, sipping from a teacup'},
masterpiece, best quality, amazing quality, very aesthetic, high resolution, ultra-detailed, absurdres, scenery, photorealistic, BREAK,
{tags provided to DeepSeek-R1 e.g. 'attractive woman, flirty,'},
BREAK, detailed eyes, detailed skin, photorealistic details,Negative prompt:
worst quality, old, early, low quality, lowres, signature, username, logo, bad hands, mutated hands, mammal, anthro, furry, ambiguous form, feral, semi-anthro, lowres, jpeg artifacts, bad art, blurred eyes, undetailed eyes, ugly face, blurred, grainy, cut off, oversharpened, sign, amateur, extra limbs, dull colours, boring, lackluster, abominations, more than 2 legs, more than 2 hand, more than 5 fingers, fused hands, switched feet, bad proportions, text, multiple views, split image, split screen, multi panel, (belt), mosaic censoring,cfg: 3.5
sampler: euler
scheduler: beta
steps: 35
Description
FAQ
Comments (21)
I've tested this -- I believe you are on the right path, I didn't get any images that were up to the standards that I'd post. But a lot of this is really quite good. Eyes are probably the weakest area.
Thanks, yes I have noticed a few lazy eyes, and some malformed irises. The faces in general can also get a little weird at medium distance or above
it's not quite there yet but this is the best illustrious realistic model i've tried since it plays along fairly well with character specific prompts and LORAs compared to the others
I found this model to be superior to several other realistic ILs, but the fact that it only outputs this woman may hinder its use. At any rate, she has a very large chin. No matter what we do, her chin will always stand out, and there is no way to control it, so we will always be looking at her huge chin.
Yeah that's an interesting one, I'm not sure where specifically the face is coming from yet. The training set I used for the realism Lora was pretty diverse, but maybe it's just been averaged? Or could be from one of the base models I merged? It might just be too much manga face coming through still. I need to do some testing
v2.0 has far more variation in the faces, but it still has a very strong preference for women.
@psychologicau it would be better to get your samples with prompt metadata. Judging by the standard post by Lizardon this version gives overburned images at CFG 7. Semi-real at CFG 4.5. Sensitivity is too high.
Full metadata is in the images, I will test out another save method which might make importing to the interface here less tedious.
As for CFG at 7 causing saturated images, I'll have a look. I use lower CFG for most models and I guess I never tried.
@psychologicau Do you use ComfyUI or what?
If ComfyUI, then a Save Image node from Comfy.Core is the only node which shows a "Nodes" button in CivitAI. Other nodes metadata cannot be recognized by Civit.
@homoludens Thanks, I see what happened there. Images were saved with the 'Save Image With Prompt Data' from mikey_nodes, as the positive prompt is partially generated in the workflow and hence is not saved by the core nodes. I'm looking into another node to attach the metadata
@psychologicau Mike's Save node never worked for me.
@homoludens I found a solution which works with my generated prompts using comfyui-lf
@psychologicau More details please?
@homoludens I realise now that previous comment was ambiguous. I found a solution for saving the metadata so Civitai ingests it properly in https://github.com/lucafoscili/lf-nodes
The showcase images for v2.0 should have the full workflow attached using this.
As for higher CFG causing saturation, I think the new version is better, but lower values (I use 3.5) still give a better result.
@psychologicau Some long time ago there was a LORA like "add_contrast" which can help you to overcome the CFG problem. What you need is to merge your CKPT with any LORA specifically developed to reduce the contrast ("add_contrast" LoRa with negative weight do the trick). After applying low contrast LORA your model will work better on higher CFGs
@homoludens Just had a look through some of those comment threads on your models and I'm gonna go out on a limb here and say people complaining about high CFG causing saturated images are setting their CFG too high. Your v1 was fine. I'm less a fan of the latter ones.
I'm not going to specifically address it. Just made sure to include my recommended value in the about section.
@psychologicau On most of my models good realism is in range of CFG 3...5. For me it is too low. Most of the tools use CFG 7 as default. This is where realism should work (my goal).
So, I accepted that critics.
The problem with CFG near 3 is that A1111 WebUI for example have a step of 0.5 programmed. And it is practically impossible to have precise control of the CFG in the range around CFG 3. Differece between CFG 2.5 and 3 can be too big. I can open the Dev Console of the browser and set any step and Min / Max, but nobody wants to hack the UI.
@homoludens This one? Control LoRA Collection - cfg scale BOOST | Stable Diffusion XL LoRA | Civitai
@samwyse link don't work for me, but I remember this lora to be used some time ago. Not the best one but seems working.
@samwyse
In v2.1 I mixed in:
https://civitai.com/models/1027924/illustrious-contrast-adjuster @ -0.5
https://civitai.com/models/1093089?modelVersionId=1227749 @ 0.75