Easy image quality, more creative output
V1.3 Update (creativity upgrade)
Merged in the amazing Paradox2 model for it's creativity. Managed to keep the BPN1.2 hands intact(mostly). How far can you push the model before it stops to really understand what you want?
All preview images are pure txt2img, no lora's used.
A lot more creative imagery
Listens to prompt a lot better (some prompts won't work anymore because it might listen too well)
Has some repetitivity issues when it comes to buildings in wide shots (1280+px width, to be fixed in future merge)
Less character, more creative focused
Because the model is a lot more creative now, sampler choice becomes more a factor too. Use DPM++ type samplers if you want cartoony styles, or Euler/Heun type (which includes Restart) samplers for photorealistic. If consistency is what you want, i advice to use LCM and the LCM lora(Lora weigths:~0.8, 2 cfg). FreeU and SAG were used for images, but they add little/no extra generation time.
If you have questions or remarks on how to improve the model, join the CivitAI general chat on Discord, I usually check in there quite a bit.
V1.2 update (Massive improvement)
Massive upgrade in quality. Both image stability and creativity have gone up, and errors have gone down. Shoutout to the amazing Paradox by Thaevil1 model which significantly helped increase the comprehension abilities of this merge.
All posted main post images by me are txt2img only, no inpainting, no loras. (so you know what the model can do out of the box). Only extension used is: https://github.com/ljleb/sd-webui-freeu. Remember set the template to sdxl, it'll improve the quality of your output even further.
Negatives are generally not needed anymore, unless you want to block something specific.
Most images made with A1111:
50 steps restart sampler
2x upscale, 15 steps restart or DPM3 SDE Karras
V.1.1
Just trying to reduce the jaggies a bit, as the image quality feels a bit unstable.
V1.0
A modelmix as an attempt to get Protovision to become a bit more creative in it's output. Before it was very biased towards people. I've tried reducing that habit, and replace it with creativity. All images posted by have no loras or regional prompting used, this is all pure txt2img prompting. No refiner is needed, i've managed to keep the great image quality from Protovision mostly intact.
Commonly used negatives by me (not needed anymore since 1.2)
paint, blurry, swirl, circle
These are negatives i commonly use. Paint and blurry to get the image output crisper. swirl and circle are remnants of the merge i've nto been able to get out which are often seen when the model is grasping at really abstract concepts. Having those in the negatives certainly improves the output in those cases.
Description
Massive quality upgrade. Both creativity and quality have been improved.
FAQ
Comments (11)
If y'all haven't given this a try in a while, or at all, please check it out! It's insanely good. Creative and cohesive with pristine quality, and I haven't seen such versatility outside of 1.5 dreamshaper. This takes it to the next level.
Incredible. The best merge/checkpoint/model that I have used so far. Very creative, recognize artits and styles very well while keeping the exquisite quality. 11/10
Thank you for the kind words. I hope if i can get around making a new merge, it can do the same, but even more!
I agree with what has been said so far about this model--exceptionally clear, well-defined output.
Which optimizer did you use for the merger?
Thank you for the kind words.
Optimizer? Just as always i went with gut feeling when mixing models in supermerger. If it's someting different from supermerger and the options i use there, i'm pretty sure i'm not using it. As for the mergring workflow, i just look for models i think have weights that i can use, and then figure out which blocks to keep and which blocks to discard.
So basically it's just yoinking the best parts of other models and adding them specifically onto mine. Doesn't always work, sometimes weights seem incompatible. If there's better ways to merge, i'd gladly learn, to even make a better model merge!
Prompting with GPT-4 prompts often leads to very good results.
Very nice compatibility.
Can you give me some pointers on how you prompt?
@imagemage sure, in plain detailed image descriptions, like in the cyborg images:
This is the original image description of another image rendered by Midjourney, with GPT-4 you can upload this image and get exactly this:
A close-up portrait of a female cyborg in a sci-fi setting. The image is to be highly detailed, showing the contrast between her human features and cybernetic components. The left side of her face reveals advanced technology with metallic structures, circuits, and glowing orange lights, suggesting she is part machine. Her human side should display elegance with clear skin, a focused eye, and well-defined eyebrows. (The backdrop is dark and blurry with the suggestion of rain streaming down:1.1), giving a reflective sheen to the figure and contributing to a moody and intense atmosphere.
@imagemage i added the (1.1), because there is no rain on the image
@imagemage this is the original image
https://www.fotocommunity.de/photo/erinnerst-du-dich-noch-klaus-tesching/47890995
@imagemage or try this :)
The photo shows a figure from the fantasy world, reminiscent of a hobbit. The figure is sitting in a dark room with a window in the background that lets in daylight. It has large, pointed ears, bulging eyelids and a strong, childlike-looking facial structure with a snub nose and full cheeks. Her hair is light brown and neatly combed forward. She wears clothing reminiscent of the 18th or 19th century: a dark blue jacket with a white shirt underneath and a vest. The figure's hands, which appear smaller and chubbier than human hands, are holding a finely patterned teacup, with a second cup and an old, gold-colored teapot on a table next to her.
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.