NEWTON

Latest version: HALCYON

This is a ~~complex~~, "FAST", experimental SD1.5 merge I've been working on for some time and now I've decided today I'm releasing it into the wild. There is still much more work to be done, but please try it out and let me know what you think about it!

☼ ABOUT NEWTON ☼

Newton is a mix of popular SD1.5 checkpoints/merges and many LoRA of which were trained exclusively for the merge recipe that I may release individually at a later date. The main thing to note is that LCM was incorporated into the recipe, which while this model is not quite the speed/quality of Turbo or Lightning*¹, still seems to allow for decent results at lower sampling steps and CFG scale. I imagine with ComfyUI this model could be very useful on lower end hardware using multipass workflows or even animations.

I've found this model to be very compatible with many different LoRA, and overall style seems to range from flat anime to shaded anime/2.5D. With influence from LoRA you might be able to achieve more accurate 3D /Realistic styles. Prompting realistic or 3D via keywords should bring about a 2.5D style, while anime or anime style in the prompt should focus more toward a 2D /flat anime style.

Some custom LoRA concepts that have been incorporated: virtualization, abstract scifi digital art, mecha musume, a distinct anime portrait style.

This is not an NSFW focused model, while it might have prominent feminine features and be capable of NSFW it shouldn't be likely to get any of those results unless you force them.*²

*¹: 42~ sec hires fix for similar results vs other popular SD1.5 models at 73~ sec on a 1080ti.

*¹.1: I was also able to generate some of the XYZ comparisons of 52~ images in less than 5 minutes on a 1080ti.

*¹.2: Okay, now I've generated a batch of 56 "mostly" decent images with the same average parameters minus seed, in 6 minutes flat. Since writing this section I've generated over 500 images.

*¹.3: Still not sold? See for yourself! (Elapsed time comparisons with popular models in sample showcase)

*²: I wasn't able to produce any results with nudity or sexual content testing with standard parameters/prompts.

☼ PARAMETERS (How TF do I use this thing) ☼

After some testing I've found the following parameter ranges to produce good results, I've also provided some XYZ comparisons along with the sample images.

But before we dive into the specifics, I made a little chart!

This is honestly just based on me comparing a bunch of XYZ comparisons, no special maths or aesthetic scoring here. While higher CFG/steps may add more details, I actually kinda prefer it somewhere in the middle, seems to have better LoRA control around that area too. Anywho:

Sampler: Any sampler is fine. (I think)

Euler ancestral is a good starting point. OR THE LCM SAMPLER WITH SGM_UNIFORM

Steps: 4-16 | CFG: 1-4 | Hires Fix: 4x_foolhardy_Remacri, 8-15 steps, 0.5-0.65 denoise

(Exceeding 16 sampling steps doesn't really help and seemed to oversample/ruin some details.)

CLIP Skip: 1 OR 2

(Sometimes results are very similar or different but not bad, I'm typically on 2 but I say go with preference. Or try both? See XYZ samples in showcase.)

Higher CFG = more contrast /deeper colors /more shading ( I usually keep it around 2-3 )

Embeddings as needed. I didn't do anything fancy with embeddings or enhancements like ADetailer in the sample images, okay maybe easynegative made it into 1 or 2 of them but its mostly just some Hires Fix. Feel free though to use all of the tools and tricks, the results could be incredible!

Recommended VAE: berrysmix or blessed2, but you can use any you like. I did not intentionally bake any VAE into the merges.

If you use ComfyUI, try the LCM sampler with sgm_uniform scheduling. Try it.

I have xformers enabled on A1111.

All 20 images in the first post made using A1111, expect the rest from ComfyUI.

☼ TODO ☼

• Variants excelling at Realism, True Anime, Surrealism, etc.

• Attempts at achieving better early samples.

• More poses /concepts.

• ComfyUI Workflows.

• More experimentation...

• More fun and creativity!

☼ NOTES ☼

If you find yourself with blurry or "low sampled" images, please review the recommended parameters here and maybe the XYZ comparisons or prompts in the samples showcase for reference. Results should be more decent around 3-4 CFG and 12-16 steps. Isn't that crazy though?! (Or maybe I'm crazy, please also let me know if you think I'm crazy.)

Results just don't look good? Don't worry, the model won't give you a Mona Lisa if you prompt "masterpiece", ultimately depending on your goal there may be a need to invoke quality or be very specific via keywords but I really do recommend using multipass ComfyUI workflows, as you should see a notable jump in quality with not many steps needed.

I started designing this merge with love still in my heart for SD1.5. While I actively enjoy XL models, as well as PonyXL and derivatives, I believe there is still more that can be done with 1.5. (Not to mention, what am I supposed to do with thousands of 1.5 LoRA?!?!)

I would love it and it would be a big help if you let me know what you think about the model, your experience using it and hopefully some of the results you ended up with! <3