V4 is out!

This version uses the same models as V3 (with updated epoch versions) while adding a new base, Deliberate V2! Deliberate and 526Mix are combined with 67% 526Mix and 33% Deliberate average weight merge, and then used just like 526Mix was used in previous versions. Additionally, this mix used block weight merging in Comfy to merge together the e6-laion + Polyfur part with the FluffyRock NoPE part at 0.5 CLIP merge, and the result is train differenced onto P.A.W.F.E.C.T just like V3. One last thing, the name is now officially changed to FluffyMix to make it easier to remember.

V3 is out!

This version uses fluffyrock-NoPE-e144-terminal-snr-vpred-e63.safetensors, e6-laion-576-704-832-lion-e81-terminal-snr-vpred-e30.safetensors, polyfur-lion-e87-terminal-snr-vpred-e36.safetensors, pawfect-alpha-e48-None.safetensors, and v1.5 of 526Mix. e6-laion operates similarly to Polyfur and includes the same MiniGPT4 captioned photographs, but e6-laion is also trained on anime style images from booru sites. e6-laion and Polyfur are merged together at a 0.5 weight average merge for the purposes of this mix. FluffyRock NoPE also operates similarly to FluffyRock Vpred, and while both models use the same dataset and are both v-prediction, FluffyRock NoPE messes with the text interpretation part of its model in order to remove positional encoding, meaning it doesn't require a token limit. It isn't known if this is passed on through merging, however NoPE did have a different and more preferable aesthetic from its sibling, which is why it was chosen. Finally, P.A.W.F.E.C.T is another v-prediction model trained on images from FurAffinity. The unique aspects of this model are its aesthetic and tags, as FurAffinity tags are often diverse and sparse, meaning images are often incompletely tagged and sometimes use uncommon tags. This means P.A.W.F.E.C.T both provides unique tagging and a unique aesthetic to this mix.

V2 is out!

This version uses polyfur-lion-e76-terminal-snr-vpred-e25.safetensors and fluffyrock-576-704-832-960-1088-lion-low-lr-e126-terminal-snr-vpred-e99 mixed with 526Mix to provide the model with even better knowledge and aesthetics. Polyfur is special, as it attempts to include both high quality photographs autocaptioned by MiniGPT4 in its dataset alongside furry images, so that you get the best of both worlds. The model is early in its lifespan, but contains enough specialised knowledge that it significantly improves the quality of this mix.

Unofficial merge of fluffyrock-576-704-832-960-1088-lion-low-lr-e126-terminal-snr-vpred-e99 and https://civitai.com/models/15022/526mix-v14 using train difference with the supermerger extension in automatic1111.

Train difference is explained here https://github.com/hako-mikan/sd-webui-supermerger/blob/main/calcmode_en.md#train, and is what allows this model to exist. TLDR, this model uses the a1111 supermerger extension to merge 526Mix with FluffyRock vpred using these settings: 526Mix is set as model A, fluffyrock vpred as model B, v1-5-pruned (sd base model) as model C, add difference as the merge mode, train difference as the calculation mode, alpha is set to 1, and the output is saved as fp16 safetensors.

This model requires a config and CFG Rescale. Download the config file and put it next to the regular model in your stable diffusion models folder. If you use a1111, install the CFG Rescale extension here: https://github.com/Seshelle/CFG_Rescale_webui. If you use comfy, download the sampler_rescalecfg file from here https://github.com/comfyanonymous/ComfyUI_experiments and put it in your custom nodes folder. Recommended settings are 7.5 for CFG and 0.7 for CFG Rescale.

Credit goes to zatochu on the furry diffusion discord for figuring this out, and credit to 526christian for creating the 526mix and lodestone for creating fluffyrock.