"New" version? Nope, just a fix.Changed the VAE for other slightly simmilar, and added some furry, mecha and naked bodies in "dynamic poses" to dataset to fix incoherences (look at model pictures to understand). Just a few 200 pics, nothing big. Everything should be 95% the same. CFG 1.7 to 3.7 seems to be the sweet range. Steps still 10 to 16.I'm removing old version since it is worst in every aspect, and is occupying precious space in Civitai's cloud. ___________________________________________________________ What is this? Let me make it as short as I can:First, I took <a target="_blank" rel="ugc" href="https://civarchive.com/models/257749?modelVersionId=298112">Pony Diffusion V6 XL DPO </a>,then I made several pictures with <a target="_blank" rel="ugc" href="https://civarchive.com/models/378903?modelVersionId=430191">Fast Pony Merge v2</a> and <a target="_blank" rel="ugc" href="https://civarchive.com/models/563148?modelVersionId=1094115">Cognitive River Pony v2</a> and downloaded some made with FLUX over here.I digged here and there for some good and free to use Pony-compatible LoRas, then started merging all that, trial and error... a lot of time. Added a fine VAE (old but gold) into the first merge and... well... see the pictures.It's the first and only post (no upgrade in sight) so enjoy it (unless some Pony fan comes to ask for a new merge). It's called v6.9 because it had 8 stages but I liked it more like 6.9 so it's clear the model can't stand against v7.<ol><li>It can generate pictures from 768*768 up to 1152*1152</li><li>Use always Euler_a . Under some conditions, Euler or dpmpp_2m_sde could do better. Usually with style LoRas or inpainting, feel free to experiment.</li><li>10 to 16 steps</li><li>1.5 to 4 CFG</li><li>Karras, Normal or SgM_uniform scheduler.</li><li>'source_pony', 'source_furry', 'source_cartoon' and 'source_anime' still works.</li><li>'rating_safe', 'rating_explicit' are sort of needed to guide works.</li></ol>THIS is a must in positive:<pre><code>score_9, score_8_up, score_7_up, score_6_up, score_5_up, </code></pre>THIS is a must in negatives: (delete "ugly face, simple, horror" if you want to make monsters. "simple" for 2D flat or cartoon)<pre><code>score_5, score_4, censored, watermark, bad hands, disfigured, error, mutated, long chest, (long torso, long body, bad anatomy), ugly face, simple, horror</code></pre>

CaPony Diffusion v6.9

Just changed the VAE for other slightly simmilar, and added some furry, mecha and naked bodies to dataset to fix incoherences (look at model pictures to understand). Everything should be 95% the same. CFG 1.7 to 3.7 seems to be the sweet range. Steps still 10 to 16.

Files with SHA256: 5b2f6518e414d5f3ef7cd972b9d6466bd32988d1e248403aa1ea9b87e4d86928