This is a merge of Everclear and some loras that I made in the past few days, those loras are mainly trained on photographs and 3D realistic renders, my goal was to have a realistic feel to PONYV6, while using ponyV6 loras to get diverse results. the hands can be fucked, and heads are mostly okayish. There are small differences in composition between Everclear and Hecate Realism.
Basic prompts and recommended settings
positive: zPDXL, 3d:0.5, realistic, raw, photo, uncensored, 8k uhd, ultra-detailed, score_9 BREAK
negative: zPDXL-neg, source_anime, source_cartoon, blurry, simple background
clip skip 2, CFG scale:5, generate on classic SDXL resolutions and with pony recommendations
The embeddings are available there:
https://civarchive.com/models/332646?modelVersionId=372656
Contributions
4 loras that I trained were used, and two extract were used:
https://civarchive.com/models/288584/autismmix-sdxl weight:0.2
https://civarchive.com/models/325820/incuriosponydiffusion weight: 0.1
Base model used for merging:
True base model for all others:
Socials
I'm mostly doing LORA for fun, if you have a request to make and want to share your creations using my LORAs or not, you can join me on my discord server.
This is the better place to ask questions about my LORAs and how to generate better images using them.
Description
FAQ
Comments (16)
By chance, does this still fucks up teeth? Other realistic models have that as a main problem. Still thanks for the model mang
Still fucked up sadly, especially with smiles, not so much with open mouths, people who trains on Japanese idols don't necessarily tag crooked teeth and that's where the problem can come from I think
@Hecatonchireart Oh well, still is progress towards a true photorealistic model, then I personally just hope we get some horny furry to adapt it and we have a indigo-like model for pony
I guess, I have found a (somewhat) Solution....have a look at the Discussion I´ve started (because I don´t want to copy everything around).
Impression is that the model has only one face. Needs variety.
I've collected enough images from different artists, people and ethnicities to improve the diversity but this is a long process, I'm also learning better techniques for merges since this is my first one, every critique is welcome though, thanks !
@Hecatonchireart I hope you will make a great model.
What would you say is the difference between thid and everclear, since you admit them to be so similar? As a selling point
The composition is different and the faces are a bit more realistic in usage, There is also more 3D realistic hentai in my dataset, and that helps making more extreme images with a realistic texture, but currently the merging is only from 4000 images so it's not extremely noticeable on most gens.
@Hecatonchireart tested it to do some nsfw, it seems to struggle a bit to meld nsfw with backgrounds, but, its very promising!
also struggles a bit to do futas, but, you know, its a mix :) hope you continue to work on it! i dig it
After nearly 6 Hours of testing your Checkpoint, (in ComfyUI), I have found my Solution for "Pony" -related (Realism) Problems:
- Hair (check)
- Eyes (check)
- Teeth (check)
I used the following Lora´s:
DreamshaperRealisticDetailer: Strength: 0.7
Https://civitai.com/models/328911/dreamshaperrealisticdetailer
Vixon's Pony Styles - detailed painting v01: Strength: 0.7
https://civitai.com/models/352581?modelVersionId=399443
...but the real "Gamechanger" was, to switch to:
"dpm++_3m_sde" Sampler. ("dpmpp_3m_sde" in ComfYUI)
Created a Batch of 5 Portraits...and they are all good. I have uploaded one of them, have a look.
Kudos to you, and a big "Thumbs up!"
Cheers,
D.
I experienced with it a bit, and it seems that the Vixon's style worsen the result and dreamshaper really improves it, also the sampler improves the results too, thanks for sharing !
About that "Vixon´s Style"...you are right. It worked in that "Face close-up" Scenario, but otherwise, not that much. I reduced the Strength of it to 0.3. This works for me.
The model does many good things, like different aged character and a huge knowledge of different action tags. But I noticed that it does not deal with light/shadow well. I tried to get more epic lighting and darker theme, but the model always gave me a very bright character. Although the environment is dark, the character is always very bright.
Will try to have darker images with my dataset for the next iterations thanks
@Hecatonchireart I just found a lora called https://civitai.com/models/380539?modelVersionId=424827 Zavy's Rimlight. It's a SDXL1.0 lora but can work well with the model. With the use of this lora together with your model and hiresfix can give you incredible lighting and skin texture details!
Details
Files
Available On (2 platforms)
Same model published on other platforms. May have additional downloads or version variants.