Let me start off by saying THANK YOU to all of you that spend your time creating amazing tools, models, LoRas, workflows and tutorials. Whether it's for hobby or to put food on your table, your work is greatly appreciated and adored!
v1.2 The Finale
This version I took a deep dive into model merging and painstakingly tested individual block values with merging DigitalAF_XLP_v11 with Zonkey - Realism - v5.0
Shoutout to bot_i_cell for an amazing model!
Like with everything ComfyUI was the laboratory. I focused on pulling as much structure, detail and clarity out of each model.
Most of the images in this version's showcase are with hires fix, face and eye detailing and final upscaling, solely with this model. Some are just with hires fix. Plays well with all the favorite LoRAs while maintaining the realism.
This will possibly be the final version of this model (possibly an impaint and lightning version). I feel like I've done all I can to pull out what I intended when I started, better realism, more quality, versatility and better overall features. Hope you all enjoy <3
Recommended settings:
Sampler: Euler, Normal/SGM_Uniform
CFG: 5-7
Steps: 30-40
v1.1 4 Step Lightning
I got a request to make a lightning version and wasn't quite sure how to. I had issue getting the lightning loras and models to merge and still maintain the pony attributes. (I'm not an expert at this, I throw things at a wall and see what sticks). So I used a simple model merge DigitalAF-XLP_v11 with SDXL 4step Unet model and a touch of the lighting lora. But the workflow below is what I finally got to work. The image generated very similar to one generated by the other versions, using the same seed and that's what I was looking for, along with quality generation. I was very surprised and pleased with the results, with it being only 4 steps.
Recommended settings:
Steps: 4
CFG: 1-2
Sampler/Scheduler: DPM++ SDE; Normal, SGMUniform or AYS
v1.1
Trained several different LoRas for color, lighting, details, quality and merged with v1.0. Had several iterations that were great, but at some point along the way it lost some of that Pony NSFW quality, I turned to a god of a model Sateluco's SatPony-NoScore and mixed in ~30% to bring back what was lost and bring out even more detail. Major kudos to Sateluco for a great model!!!
Recommended settings:
Use the baked VAE, seems to be much better than the default sdxl.
With SatPony's no score, it seems to now be just as responsive without the score_9, 8, 7 etc. But does help with some prompts still.
Sampler: DPM++ 3M SDE Karras
CFG: 6-7
Steps: 20-35
v1.0
DigitalAF-XLP is the result of my experimenting with merging models/loras in ComfyUI using a few of my favorite Pony models mixed with some SDXL models and loras. I've tried several different combinations over the last few months, learning what worked well, but mostly what was an absolute mess of a merge, ultimately starting over from scratch. I also started training my own LoRas, which I plan on releasing when I feel they're ready and not something that's trying to reinvent the wheel. I'm more interested to see what the community thinks and to see what creations you can bring from it.
Model DNA:
There were several iterations and combinations of the above models, so it's hard to say what the % of each is.
Added LoRas:
Some person WIPs of my own for skin texture, quality and detail.
Settings I've found work best for me (still experimenting):
Sampler: DPM++ 2M Karras; Euler A++ Exponential; DMP++SDE Exponential;
CFG: 4-8
Steps: 20-35
I only use ComfyUI, So do your own experimenting in A1111 or elsewhere for the best results.
What's it good at?
SFW and NSFW: has no problem with either
Does dark scenes, low light really well
Prompt adherence
How can it be improved?
As with every SD model ever, anatomy, but specific things I've noticed, better eyes from a distance
Can be somewhat hit or miss with hands, a majority of the time they're great, others not so much
I've noticed some artifacts in some generations, they show up as hot spots or grey orbs, usually on or around an area of the body that's in the prompt.
I'm always open to suggestions and love learning. So if the elders of AI have any suggestions please, I am all ears :)
Enjoy <3
Description
4 step Lightning version of DigitalAF XLP v11
Recommended settings:
Steps: 4
CFG: 1-2
Sampler/Scheduler: DPM++ SDE; Normal, SGMUniform or AYS
FAQ
Comments (11)
me love you long time.
Hopefully, you will improve east asian support.
You can refer to Korean or Japanese entertainers, especially ones popular for 10 years or longer.
I will take that into consideration for the next version.
@DigitalAF Please note that there is a significant difference between typical East Asian aesthetics and the aesthetics of Asian women in Western countries. We generally do not think highly of the way Asian women are styled in the West.
That is why I said:"You can refer to Korean or Japanese entertainers, especially ones popular for 10 years or longer."
@daixiaoluck261 your input is appreciated. Can you give some example Korean/Japanese entertainers and for contrast/comparison, some western styled ones?
@DigitalAF I don't know whether CivitAI accepts URL of another website. Let me try. This is an extreme typical western styled Asian women sample for me: New York Times Opinion on X: "“A single person can become a vehicle for so many others’ most personal hopes,” says @conniewang. “They can find, in her, a sense of belonging that’s otherwise in short supply.” Follow the link to read more. https://t.co/AY1ueOBuce https://t.co/aklFIPNKih" / X
If you can see my link, please notify me. For example, "like" my comment.
@DigitalAF OK. Then, I will talk about two Korean entertainers who fit the typical East Asian aesthetic: Yoona and IU. The former has two Instagram accounts, while the latter has one. Considering their ages, please look through their posts before they turned 30, especially for Yoona. Also, do not ignore their advertisements for Korean companies. I emphasize again: Korean companies (although East Asian companies are generally good, not only Korean ones). A good example for Yoona is Innisfree. A good example for IU is Jinro. These two brands have collaborated with the respective stars for many years.
CivitaAI already overflooded with asians, just saying. Just because of numbers, not quality.
@ArnAbbas That's a great point and there are lots of models that are more influenced with that ethnicity. I have been testing mine for responsiveness to ethnicity and they work quite well, just a few that don't seem to respond, but Asian, Korean, Japanese, Chinese, etc, seem to, with distinguishable changes in features from one to another. It's all about the prompting. If it's not the look of a certain person, then I'm sure a Lora or embeddings would be easier and get better results. It seems to be much more challenging to 'train' a model to something that specific with out changing the entire model's look and responsiveness
@DigitalAF thats right... Lora's exist for granularity. A base model needs to be a jack of all trades in a style (be it drawn, photoreal or inbetween) -- rather than a race or specific concept. However this being Pony, training IRL people is going against the point based on Pony's origin training. You can mimic irl people by bruteforcing but its never going to be perfect unless the origin training is solely irl datasets
@DigitalAF I tried your Lightning checkpoint. In terms of Scheduler, I can find SGM Uniform(there is a space in the middle) yet I can't find Normal or AYS.
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.