Provides more defined characters and backgrounds and can improve character/style LoRA accuracy compared to the base Pony V6 model.
Characters included in the base model have their own style which you can override with artist prompts or style loras.
Recommended positive prompt: score_9, score_8_up, score_7_up, source_anime
For western style replace source_anime with source_cartoon
If using a character or style LoRA and its style is not preserved remove source_anime. Can also remove score_8_up, score_7_up if the LoRA style is still too weak but this usually simplifies the image too much.
To use artist tags shown in the examples use art by followed by the artist name.
v2 is a 50/50 merge of confetti4you and Comradeship v5a DPO. v3 is the same but with Con5tiXL. Some prefer the 2D style of v2. v3 seems to work better with character LoRA especially with 3D ones.
This model is also available as a LoRA extract designed to be used with the base Pony V6 model.
All images are generated with Euler A sampling method, 6 CFG Scale using WebUI Forge.
WebUI Forge is built on top of Automatic1111's backend but has lower VRAM usage and higher performance. Try it out!
Description
Merges con5tiXL https://civitai.com/models/184813?modelVersionId=392194
View x/y plot images for differences. Not sure if better.
FAQ
Comments (11)
My favorite pony base model for anime!
how would you say v3 has improved over v2? i saw some comparisons in the preview images but wasnt sure which was what version if thats what its comparing
As I mention in "about this version" I'm not sure if it is improving. Since this model is a simple 50/50 merge it exists as a convenience for people who don't want to merge the latest versions of confetti tune and comradeship.
I've attached x/y plots that compare the two base models and the three versions.
animeconfettitunexl_con5tixl and comradeshipXL_v5aDPO are the base models that were merged.
anim3conf3tticomrade is v1
confetti4comrades is v2
confetti5comrade is v3
The second plot also demonstrates the effects of different negative prompts.
https://civitai.com/images/8112489
https://civitai.com/images/8112483
I'm interested in hearing what version people prefer and what prompts have worked well for them.
probably still is my favorite and most used checkpoint
Wall of text incoming. I have been searching for a more flexible "AutismMix Confetti" for a while. There were 3 main aspects I wanted:
convincing 3D (Confetti morphs into 2.5D on long contexts), good LoRA compatibilty and non-wrecked anatomy.
I was testing this checkpoint for the past few days and it seems to have all three. LoRAs trained on Confetti weirdly perform better here, more details of the style are preserved (not cherry-picking). Same for multiple LoRAs.
Among 10+ Pony-based mixes tested this has the best 3D and background consistency out of all of them. Anatomy is a bit worse than Confetti, but way above base Pony. There is also no "character loss" that Confetti has, the full Pony list is preserved.
I do not know what black magic or random chance was involved in mixing this beauty of a checkpoint, but bravo regardless. I will train a style LoRA on this soon and report the results.
Thanks for sharing. I'm interested in hearing your training results. For style and character LoRA I always train on the base pony v6 model and not on the merge. For example the overwatch and disney 3D LoRA are trained on the base model but work well with this merge.
Really have no idea why is this model so recommended, using it with LORAs trained on pony it has a very bad habit of breaking the style.
Literally turns into base PDXL, don't know what I'm doing wrong, if it's a merge it shouldn't work differently with the same tags and loras.
v2 is really good.
Does this checkpoint need the sdxl vae or is vae baked in?
baked in i think. i never select one in a1111 / forge.
where was this model from my life man, gg well played, legit the most accurate for self-made loras.
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.