I've been using this model to generate background images for my AI driven Text-based RPG game at www.forgesaga.com and enjoyed the results so much I had to share it with the community too.
Sample Prompt with a bunch of loras to get the results I am getting:
POSITIVE:
painting of ..., late night, dusk, night, dark, starlit, fantasy, medieval, loading screen art, <lora:epiNoiseoffset_v2:0.2>, <lora:howlbgsv3:0.5>, <lora:artbytokiame:0.5>, <lora:add_detail:0.3>, ink outlines, wallpaper, vignette, matte painting by (Kazimir Malevich:1.1), simple lighting, dynamic shading, complementary colors, HDR, absurdres, cinematic, muted colors, limited color pallete, (masterpiece:1.2), (best quality:1.1)
NEGATIVE:
realistic, character, people, photograph, 3d render, blurry, grayscale, oversaturated, bad-hands-5, easynegative, photozoov15, structurezoov15, verybadimagenegative_v1.3, cropped, (watermark, logo, signature, text, signature:1.4)
Settings:
"cfg_scale": 7, "sampler_name": "Euler a", "steps": 40
To get the results above, I personally run my generations through a second img2img pass at 0.45 noise with the same settings and a canny controlnet of 0.3 to keep the general shapes intact.
I then do a final upscale to final resolution.
Enjoy!
Description
FAQ
Comments (8)
Hmm, I wonder if "painting of " is almost a required keyword to get relevant stuff?
It would be great to see more of your original prompts in the example images @knarb.
Yeah most likely helps trigger landscapes from paintings.
All my prompts are generated from python scripts, so they follow the same basic formula from the one posted above.
A part describing the subject (exterior, forest, whatever) and then a part on the style. Things I change here are the artist and modulate some loras depending on interior/exterior.
Credits to the used models for that merge ?
I would love to, but this is one of those merges I wasn't really planning on, and I don't keep records of what I used to mix it. Was never planning on uploading it, TBH. Just sort of a happy accident.
So... I'ven ot been able to reproduce anything similar to this. It doesn't help you are not sharing the prompts of the image you are using as an example.
Added some more details on my workflow. It really benefits from a second img2img pass with canny controlnet on the initial generation, then a final upscale to 4x
Looks great! It would be helpful if you added the LoRAs you are using in your example prompt as Suggeted Resources, so folks don't have to hunt for them based on the name in the prompt.
Hi! This model only gives me images that looks like paintings. Nothing like your images. How did you manage to remove the painting style and instead get the digital illustration style in your images?
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.