THRILLustrious
By pAInCREAT0R
"Thumbs Up Reviews" are appreciated as it is the only way creators not in CIVITAI's creator program get their work noticed. - Thanks!
Putting the THRILL and more 'real' into Illustrious!
V3 Real ILL - This is a generational jump forward for Thrillustrious.
Better overall visuals, better details, and more realism. Give this one a try. It was the result of targeted training over many hours. IMO, one of the biggest leaps any of my models have taken between version releases. It is finally starting to compete in my mind with my PONY and SDXL models. This version is V2 very heavily trained and then merged back at .85/.15 with V1.
V2 - Just added a lot of training on a subset of about 8000 images while I was traveling. Enhanced details, faces, and overall quality.
For V1 Genesis, I started a completely fresh training on the base of Illustrious XL. I then merged that new checkpoint with BETA and trained on the resulting Checkpoint. The idea was to continually reinforce the concepts in the base model while pushing the aesthetic toward realism. This approach diluted the residual impact of the models merged into BETA.
I am not great at prompting Illustrious models yet. I spend more time in FLUX and PONY. But I thought the early results were worth sharing. Hope you enjoy. If you want this one to keep being trained, give it a like or leave a comment. Better yet - post some images you made with it in the gallery!!! (I don't need your buzz!!)
I have been using Euler A, Beta, 40-55 steps, 2-5CFG. However, I have not experimented with it enough to say that is optimal. The model benefits from HiresFix pass.
How BETA was made:
All training was done in Kohya starting with the Illustrious base model. No text encoder training at all - only UNET. With the recent Kohya updates, you can do higher batch and gradient sizes at faster speeds than ever. So on my 4090 I could batch 5 or 6, gradient 5 or 6 to get a virtual batch size of 25-36. This made it much faster to train than in the past. (Honestly, this was the only reason I attempted it after finishing PONY model.) There was some DareMerge of a small amount of my FULLY_REAL_EXTREME model and Redcraft Illustrious model to smooth out aesthetics and add a little more flare. About half the training was done at 1024 and half at 512. It could use another pass at higher resolution to help the faces. I preserved the clip from Illustrious 100% though and did not train encoders at all. (I figured I would mess it up.)
Disclaimer: This model is provided for educational purposes and for the advancement of AI art in general. The creator of this checkpoint assumes no responsibility or liability for the end user's outputs. Any illegal, unethical, or deep fake use of this model is the sole responsibility of the user engaged in those activities. The end user assumes all liabilities from use of the checkpoint.
No permissions are given for anyone to distribute this model on other sites. I do not give permission for it to be used on any image generation service other than CIVITAI.
Description
V4 Minor training enhancements. Refreshed VAE as the community noticed it was causing some color bars in some of the images.
FAQ
Comments (20)
Wow, this seems surprisingly realistic for an Illustrious-based model. Kudos and thank you. Is there a list of tags for your photo dataset?
Please do continue! Illustrious needs a realistic push. Your model excels at the following: It heals the I call them "ugly illustrious faces", It still understands some character features and Is not totally random ;-). As you say, it struggles with anatomy and I would add that it also has problems resolutions some non 1:1 resolutions. Please keep it up :-)
Only messed with this for a couple hours, but it seems pretty good "out of the box" so far. Looking forward to see how far you can get with this model. Keep up the good work.
Excellent work, definitely the current champ in the realism category of Illustrious in my opinion, great prompting, amazing backgrounds, great creativity, anatomy could use more training, and its a bit weak at fantasy characters, like non-humans, anime, videogame characters, but Im having a lot of fun with it so far, thank you for the model, keep up the awesome work!
Thank god someone is doing some work towards proper realism instead of semi-realism.
Please continue on this and get it better. One feedback so far is that the eyes seem a little too 'bright', or maybe it's just that they look a little too big.
Either way, getting towards realism is what's up.
Maybe this could use a merge with the pony URPM model to make something nice but i'm not sure how well pony/illustrious merges play out.
Pretty incredible that you were able to make Illustrious look this realistic. Every other realistic Illustrious finetune right now has a sort of 2.5d look to it but this one looks almost on-par with realistic SDXL models. Hope you keep working on this.
This is crazy awesome, your a pro, man!
Hope you keep it up.
What are your recommended hires fix settings?
Which prompt base should I use? Can the model make an Asian face? Will there be a version based on NoobAi?
amazing model, please continue updating it!
Excellent!
Thank you for sharing.
Damn this model is so nice. It is ecactly what i was looking in a illust model. Sometimes I need a few more details to make it more realistic, but the base is always good enough!
Maybe some tips which embedings or detail-loras to enhance, when the result needs just the nudge in the right direction?
Now that's the first illustrious realistic model I see that has reasonable output, great work!!!
how are you guys getting such sharp images? i take the exact inputs from the gallery and somehow it comes out quite blurry. is it an upsampler?
V3 Looks awesome! Great work there.
Looks awesome man! Great job!
Between your Pony and Illustrious models, which one you like the most? What are your thoughts in the Illustrious vs Pony capabilities?
Thanks For all your work!
best cg model v3 ,my third girlfriend
Surprisingly good model. Does all kinds of 3d art really well.
Will you add online generation to Civitai?
tested a few sampler/scheduler combos, most actually work fairly well. dpmpp 2m/3m sde, deis, euler a cfg pp, etc. I don't go above 30 steps (with the sgm uniform scheduler). nice model, thumbsup.
as a sidenote, less is more when it comes to the negative prompt. too much shit there=body horror.
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.