USAGE TIPS AND BEST PRACTICES AND SETTINGS:
Keep CFG at 5 to 6, as this model is a tad overbaked. Only in rare cases do certain esoteric prompts demand normal CFG (e.g. wikihow articles).
~40-45 steps recommended when not using img2img to upscale as your high-res fix.
img2img at 0.47 to 0.55 strength and same step count is recommended for high-res fix on most prompts to upscale from a small resolution. InvokeAI users should enable High Res Optimization, ideally at the same strength. With that enabled, you can get away with 30-35 steps on DDIM, k_euler_a, and DPM samplers to get clean and fully-detailed results.
highly recommended for boosting photorealism and 3D art in many cases, though may occasionally behave oddly for certain prompts (e.g. interior images, spaghetti demons). Download the learned_embeds.bin, and place in your embeddings folder. You may want to rename it. The trigger in the Automatic1111 UI is whatever you change the filename to, while in Invoke it’s .
Best results at at least 768x resolution in either direction.
When prompting for a person’s age, try typing their age in words, not numbers. Numbers are tokenized separately, so prompting for a 42yo instead of a forty two year old can give you a man-sized kid.
If your images are a bit dark, try putting low light in your negative prompt to kill dark lighting effects, or dark to kill dark coloring.
DPM samplers like DPM++ 2M Karras are more liable for artifacting and weirdness in this model, and may need extra steps. DDIM is more stable, though may not be as good for some prompts, such as high-detail photographic prompts.
Other negative embeddings besides (e.g. bad-artist-anime, bad-hands-5, bad_prompt_version2) are generally not recommended, as they impact art style and model expressiveness and have little, if any, actual benefit.
k_heun is excellent for photorealism, but is very slow. Some prompts can work fine at only 15 steps, but others may come out wobbly until 30 steps.
For some very high-detail images and cozy interior design, a higher img2img/InvokeAI HRO strength like 0.65 can do good for detail level and sharpness. 0.75 can also work, but that high risks introducing a bit of the crunch effect experienced at 7+ CFG. Use wisely.
Update V1.4.5: This version tries to somewhat reduce iris and pupil distortion in peoples’ eyes and improve general coherence somewhat, on average. Also, it’s a touch less dark. This was accomplished by building a sub-mix with similar behavior to V1.4, but with better eyes and 50% noise offset, then mixing that in at 20% and hoping for the best. Seemed to work okay. Has occasional interpolation weirdness in a few prompts.
Anime / 2D animated-optimized version here: https://civitai.com/models/35893?modelVersionId=42086
Surreal / vivid illustration mix: https://civitai.com/models/44596
My personal merge of Stable Diffusion 1.5 custom models using the noise offset to improve contrast and dark images. An inpainting model is provided to make inpainting in the model’s styles and detail easier.
I like to think of this model as being like base SD 1.5 hyped up on energy drinks and ADHD. Maybe some artist names aside (I don’t use them ever, anyway), it has a lot of 1.5’s general-purpose power, but has a much higher baseline detail and quality and can be very responsive, provided you’re using appropriate settings.
This model is meant to be:
Artistic and elegant
Drop-dead easy to work with
Good at making cool characters and landscapes
Not bound or leaning towards any single style
Killer at digital and conventional art in many aesthetics
And above all, fun
It’s not so great at explicit sexual content and anime*, including anime-based embeddings. There’s a million other models for those if that’s what you’re after.
*There is some ability to bring out a neat anime aesthetic when you prompt for ‘anime style’, which I find to be quite cool to look at, although it can be a bit finicky. If you try to make anime-esque art with this model, do not put ‘portrait’ in your negative prompt, or use ‘close’ or ‘closeup’ in your positive prompt, as those seem to force it into a 3d-like style even if you add more weight on the anime style.
I want to also bring attention to whosawhatsis’ verisimilitude, which is great at readily making wallpaper-quality photorealistic stuff.
I also want to shoutout coreco and his seek.art MEGA v2, which was responsible for much of the composition of V1.3-V1.4, and is an excellent update to his Mega model.
This model has a baked-in VAE based on Waifu Diffusion 1.4 by hakurei. You can switch this out (such as to vae-ft-mse-840000-ema-pruned) if you wish, though I used that one specifically because it helped improve detail quality in some art styles and situations.
Example images were generated in Invoke AI. This means unless you use Invoke AI, you likely won’t be able to recreate my images exactly. Just learn from the prompts and modify the weighting in prompts as needed for the UI you use (if you use the A1111 UI, any (plus sign)+ is equal to one set of parentheses).
By downloading, you agree to the creativeml-openrail-m and Dreamlike-art licenses.
Credits (V1.4 / V1.3.5):
Roboetic’s Mix – Roboetic
seek.art MEGA v2 – coreco
RPG V4 – Anashel
HeStyle V1.5 - krstive
Movie Diffusion - Dalle2Pictures
Analog Diffusion and Portrait+ - wavymulder
RealSciFi - AIfriend
Foto-Assisted Diffusion - Dunkindont
fantasy-art-style v1.8 - kasukanra
Vintedois Diffusion - Predogl and piEsposito
noise offset – Nicholas Guttenberg
Visit Official Website