Kawaii MonKe (Monster/Kemono) is the anime part of the Real Moon - Anime model
The full version requires VAE
Description
Two models were updated and percentages adjusted
FAQ
Comments (10)
The demo gallery is misleading about the real capabilities of the checkpoint. I've tested the version 7.0 and I found it's an honest anime checkpoint, no more no less. Instead, the demo images are fitted with exagerated stuff, like:
Madly detailed CG unity 8k wallpaper:1.5, masterpiece:1.3, Madly detailed photo:1.2), (hyper-realistic lifelike texture:1.4, realistic eyes:1.2), (octane render, unreal engine 5), (ecstasy of light and shadow:1.5, deep shadow:1.5), (sony a7, 50 mm, film grain:1.5, 4K UHD HDR), (photo by lee jeffries, Greg Rutkowski and Magali Villanueva)
All this 'pumping' in the positive and negative part of the prompt surely turbo-charge the look of the images, but at the price of a large control drop over the real subject you asked for, with the checkpoint rarely doing what you need...
I invite the author to change method, givin us a idea more realistic of the style of his chekpoint...
I am confused by what you say. This model only generates anime images, although the prompt indicates a lot of realism, this model will only generate anime and not hyper-realism. All the realism specifications of the prompt achieve the quality of the example images. These realism prompts and the negatives are like templates that I use in all my models, they are made to increase the quality as much as possible.
@JamOnOn Probably it's a missmatch due to Google translator, because english is not my laguage: is not a problem of realism/hyperealism or anime.
Dont' you notice a certain difference between your demo images and "cretaumundu"'s images, in terms of colors, detail, and saturations?
@settima_ai I understand, it's not my language either (I speak Spanish), the problem with those images is that it uses <lora:add_detail:0.9>, they would look better if I removed it or put it with a lower value. Is a good Lora but not always.
@JamOnOn Ideally, you would post only what the model can do by itself, no loras, no upscaling. Maybe HiResFix, but nothing else. Give an honest representation, including reasonable prompts, because if a model can't follow directions, it's useless to us.
I haven't been able to recreate anything even close to your examples either in style or composition. I tried using 6 and 7 versions and your examples from them. I failed. I got the feeling that important details are missing in the prompt, but there are a lot of unnecessary ones. I tried to recreate user examples from the gallery and the composition as a whole turns out, but there is a fuzziness of the picture, which is very often associated with the fact that the prompt is overloaded with keywords that are contradictory or inaccessible to the model. I am discouraged because this is the first time this has happened on a popular model. Perhaps the model is completely incompatible with the Draw Things app
Here is my examples (prompts and settings from model card):
https://imgur.com/3KKYQeu
https://imgur.com/8QBFsYe
https://imgur.com/YfmiYDP
@Unicom I have not used that application but maybe that is it, to generate the images I use stable diffusion webui in Google Colab, and that has never happened to me.
@JamOnOn maybe later I will test with different prompts
The showcase prompts wont work they are somehow messed up as u already have noticed. U can check my pic for a similar image with optimized prompt