AcornXL V3

I had a bit of a break as I had to have a heart surgery. I am back but working on stuff is probably going to be slow for a while. Any follows help as well as checking out my YouTube channel or most of all any direct donations on my BuyMeACoffee.

https://buymeacoffee.com/wadebabcoc1

https://www.youtube.com/@musicinthemachine/videos

I use some of my models for videos on my YouTube (subscribe, like and feel free to comment on videos) page if you want to check it out.

AcornXL V2

This is my first update for AcornXL. It's been a while, but I don't upload unless I believe I have made considerable improvement. I believe that is now the case.

https://www.youtube.com/@musicinthemachine/videos

I use some of my models for videos on my YouTube page if you want to check it out.

Acorn has always been primarily a photo model. It still is, which I think my samples will show, but I have included a variety of styles in my initial examples to show the model is not only among the best, if not THE best, photo model available, but also a versatile model capable of many styles.

There have been a lot of changes in Stable Diffusion since my first XL model. Updates to Automatic1111 which is what I primarily use, as well as others such as ComfyUI which I use some as well. I still get the best results in A1111. I find for photo quality images, it is best to use CFG in the 3.5 to 4.5 range. My favorite samplers are:

DPM++ 3M SDE Karras

DPM++ 2M Karras

and DPM++ 2M SDE Heun Exponential

Although good result can be achieved at around 20 steps, I find that taking a little extra time and doing 60 steps tend to have good payoffs.

AcornXL V1

This is my first XL model. It's definitely not perfect as there just isn't a lot to work with yet for SDXL. It is pretty good though. I use a mixture of samplers, but so far my favorites are DPM++ 2M Karras and DPM++ SDE Karras. Normally CFGs from 4.5 to 7. Steps between 20 and 40. I have experimented using it together with the base SDXL refiner and so far it seems to improve results. If you use the up to date version of Automatic1111 it allows you to have the refiner kick it at whatever percentage you want. Some of my example pic use it starting at 50% up to 90%. Just use some trial and error and see what works for you.

AND as always, my sample images use NO loras, textual inversions or negative TIs. Model and webui extensions only.

sdxl_vae is baked in

Feel free to post your gens down below. As always, I love to see what you make.

Acorn Photo Diversity V5

After releasing my Diversity V4 of the Acorn Photo models, I wasn't quite happy with the photo quality level. I kept testing and I believe this new version does a better job.

Prompting is different with this version. No longer is the (photorealistic) prompt needed and overall you often need less prompting and less negative prompts. As always, my sample images contain no Loras or TIs, either positive or negative.

Samplers I use mostly DPM++ 2M Karras and DPM++ 3M SDE Karras although many work fine, these are my favorites.

CFG anything from 6 down to 2 work good. I have found that if facial expression are too strong at 5 or 6, running it again at 2.5 or 3 still generates a great image but tones down exaggerated stuff.

Steps are usually 20 to 30.

vae-ft-mse-840000-ema-pruned baked in to all 1.5 models of the Acorn Series

Acorn Photo Diversity V4

Updated to generate a greater range of ethnicities.

Acorn Photo V3

It's getting hard to tell these generations are not actual photographs. Not sure how much more we will get out of SD1.5. For prompting, version 3 stays pretty much the same. I use DPM++ 2M Karras mostly, but also DPM++ SDE Karras and DPM++ 2M SDE Karras. CFGs from 4 to 6 mostly but depending on preference it doesn't get bad well up to 10.

Post your creations down below. I would love to see what you make.

Acorn Photo V2

This model increases the photo realism significantly over the previous version while still being a very versatile model. It is great for portrait type people pics but also for more action oriented things like Fantasy RPG type shots.

Post your creations down below. I would love to see what you make.

Similar to previous version, seems to help a lot to use (photorealistic:1.4) in the prompt.

I find it works best at CFGs from 4.5 to 7, but gets some good results at other levels.

As with a lot of models, the lower CLIP Skips seems best, but I have gotten very good results with 1,2 or 3 and good images that drift farther away from the prompt as you go higher.

As before I can't accurately state exactly what is in the mix or how much weight each has. A lot of trial and error testing. The list of models included that I am sure about are:

Photon_v10

Aamon_v10

Colorful_v31

ConsistentFactor_euclidV51

AlmostHumanMix_v10r

EroticVision_v4

NextPhoto_v20

LOFI

QGOPromptingreal_qgo10a

VirileReality

RPG

Acorn Photo

This is not a replacement of Acorn Is Spinning, but an addition with a little different use case. Feel free to play with both and see what you can make. Please post any examples you like as well.

My goal is to get a model that is versatile enough to generate any kind of content equally well. To date, this is the most photorealistic model I have made. Acorn Is Spinning seems slighty more versitile in what it will do, but Acorn Photo seems to be a little better at just pure photos of people.

I recommend using the term (photorealistic:1.4) in the prompt. Also 1.5 or 1.6 work well.

Recommended negative: anime, cartoon, paintings, glossy skin, (worst quality:2), (low quality:2), ((monochrome)), ((grayscale)), (skin tags:1.5), skin spots, acne, skin blemishes, Watermark, Text, deformed, disfigured, poorly drawn face, mutated, extra limb, poorly drawn hands, missing limb, floating limbs, disconnected limbs, disconnected head, malformed hands, long neck, mutated hands and fingers, bad hands, missing fingers, cropped, huge calf, fused hand

The examples I upload to demonstrate the model are done with no Loras, Textual Inversions or LyCos. Nothing but the prompts and upscales.

Acorn Is Spinning

This model is intended to produce the best photo real images that I can get. I am no expert in models and this is a merge of multiple other models created by those who know a lot more than me. What I can say is that in using a lot of models, this merge has given me the best results I have achieved.

My true goal is to get a model that is versatile enough to generate any kind of content equally well. Most models are designed to make attractive women, which I certainly appreciate, but I would like to be able to just as easily generate several people, both men and women without having to fight it.

The examples I upload to demonstrate the model are done with no Loras, TIs or LyCos. Nothing but the prompts and upscales.

This model is obviously not perfect. I can't accurately state exactly what is in the mix or how much weight each has. The list of models included that I am sure about are:

Photon_v10

Aamon_v10

Colorful_v31

ConsistentFactor_euclidV51

AlmostHumanMix_v10r

EroticVision_v4

NextPhoto_v20

LOFI

QGOPromptingreal_qgo10a

There might be a few other in the mix, but the heaviest weight started with Photon. On it's own, it seems to be about the best single model I have found but mixing in some of the others made some big improvements to the overall outputs and versatility.

I can take no credit for creation, so if you like the model, enjoy it and mix in other new stuff and make it better.

I get the best results with CFGs from 5 to 6, with DPM++ 2S a Karras and DPM++ 2M Karras. Normally around 25 to 27 steps. Best to use the highest resolution you can as well. I do a lot at 1024 X 768 as well as 768 X 768 and recently started using the the ADetailer instead of the Restore Faces.

For Clip Skips, I have gotten some great results with all of them up to 12, but the best are probably 1,2,3 and 6 most of the time. It seems to get nice results all the way up, but the higher you go, the farther it gets from the prompt while still giving pleasing images. Just maybe not the one you were looking for.