V4 NEW VERSION OUT!
with magic and a lot of time Cranberry is now SDXL Based. I took Cranberry V2 and force fed it to SDXL models until a decent base appeared that i trained over a new data set of 10 000 images. trained to fit an anime aesthetic with a lot of prompting freedom, it can generate great anthro/feral/human results
VAE IS NOT BAKED IN PLEASE USE SDXL VAE
Cranberry_4: Second version, trained on a larger dataset of about 10 000 images with a lot of new concepts and artists thanks to SDXL
Trained it with custom captions to still perform good on minimalistic prompts (You really can have a wide range of results for a single prompt)
Positive: Same as sdxl, simple as "amazing quality, best quality"
Negative: lowres, bad quality and worst quality are nice none but (monochrome, greyscale) are sometimes good as the model tend to generate monochrome images, you can also put (cartoon, toon) if you want to have a real anime look
PS: also for some reasons, High.res fix of 1.5 works better than 2 so use it :)
OLD VERSIONS
PonyXl trained to fit an anime aesthetic with a lot of prompting freedom, it can generate great anthro/feral/human results
VAE IS NOT BAKED IN PLEASE USE PONY VAE
Cranberry_2: Second version, trained on a larger dataset of about 4000 images for about 100 epochs per images
Cranberry_1: First version, trained on a small dataset of about 2000 images for about 100 epochs per images
Trained it with custom captions to still perform good on minimalistic prompts (You really can have a wide range of results for a single prompt)
Positive: Same as pony, no need for source:anime, but place them at THE END of your prompt to have better results
( score_9, score_8_up, score_7_up, score_6_up, score_5_up, score_4_up)
BUT
(score_9) only gives good result too
AND
(score_6_up, score_5_up, score_4_up) only act as a "good" version of the raw without positives result so try to use it too
Negative: none but (monochrome, greyscale) are sometimes good as the model tend to generate monochrome images, you can also put (cartoon, toon) if you want to have a real anime look
PS: also for some reasons, High.res fix of 1.5 works better than 2 so use it :)
Description
A semi trained version with a lot of fixing with a more 'anime' look to it. It's more of a training ground for V4 but it's good anyways
Euler or DPM ++2M works best, Euler A is also good.
FAQ
Comments (9)
Nice "default" anime style, still controllable by style Loras to a degree. Good contrast and colors with sharpspectrum VAE. Model works great at 12 steps with euler/sgm uniform ~5 CFG with Hyper-SD. Most merge models fail this test.
Great job! I am looking forward for V4.
v2 is the most unique and cutest model I have ever used. v3 seems to have changed a bit.
thx for the heads-up, v3 is still in testing phase but i wanted to release it anyway; see it as a "summer special" V4 is closer to v1 and 2
Would you consider allowing CivitAI generations with this model?
V2 is not just cute, but very stable in both style and composition/details. You can use it for stabilisation of other models in merge. Also, Cranberry style is very flexible in merges, so you can use it as base. I think it might be reasonable to use it as base model for training.
The only downside is that it might be hard to generate background sometimes (it tends to be gray/white/flat/gradient color).
Also, still hoping for V4. Great work.
maybe v4 is cooked and ready to be released ...
@N_Endo still waiting ;)
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.