Results oversaturated, smooth, lacking detail? No result at all? Please read the following!
This is a terminal-snr-v-prediction model and requires a configuration file placed alongside the model and the CFG Rescale webui extension! See further down for more information and equivalent for ComfyUI.
HOW TO RUN THIS MODEL
- This is a terminal-snr-v-prediction model and you will need an accompanying configuration file to load the model in v-prediction mode. You can download it in the files drop-down and place it in the same folder as the model.
- You will also need https://github.com/Seshelle/CFG_Rescale_webui. This extension can be installed from the Extensions tab by copying this repository link into the Install from URL section. A CFG Rescale value of 0.7 is recommended by the creator of the extension themself. The CFG Rescale slider will be below your generation parameters and above the scripts section when installed. If you do not do this and run inference without CFG Rescale, these will be the types of results you can expect per this this research paper.
- If you are on ComfyUI, you will need the sampler_rescalecfg.py node from https://github.com/comfyanonymous/ComfyUI_experiments.
HOW TO PROMPT THIS MODEL
- Natural language has a strong impact on this model as of V7. It won't understand everything but it can get you places. And then when you can't squeeze any more out of it, you can use e621 tags to help refine your prompt. (or you can just use tags, either way will work.)
- Start your prompt with a short natural language prompt describing what you want, then pad it with e621 tags to refine specific concepts. As this is a v-prediction model, prompt interpretation can be a bit more literal.
- Many flavor words and artists from SD 1.5 work again.
- PolyFur is trained on MiniGPT-4 captions, so try being really flowery with your prompts and use sentences even. (check the generation info on the preview images to see how)
- If an established character isn't coming out accurately, try increasing the strength of their token and adding a few implied tags that describe their appearance. Keep in mind that characters that aren't very popular or don't have many images in FluffyRock's dataset typically won't fare that well without a LORA.
- Avoid weighting camera angle keywords too strongly, especially close-up.
- Resolutions between 576 and 1088 should work reasonably well as that is the range of FluffyRock.
Model Info and Creation Process
V8.1 now available. Last major version for a while. Currently the quality tag LORA from Feffy is not baked in and I recommend using it as a LORA. As of this V4-V6, @Feffy's Quality Tag LORA was baked into the FluffyRock part of this model. Add masterpiece, best quality and worst quality, low quality, normal quality to your positive/negative prompt.
This model is a "Train Difference" mix (see https://github.com/hako-mikan/sd-webui-supermerger/blob/main/calcmode_en.md#traindifference) of FluffyRock and PolyFur onto a diverse and versatile base model to expand the text encoder's range of prompting and understanding of concepts outside of e621 tags. This process differs from traditional merging and block merging and shouldn't be confused with "Add Difference" that we've had. The current base model is a mix of 6 non-furry models at varying ratios, which are mostly merges as well that cover a lot of ground in terms of prompting capability. As of V6 the process has gotten more intricate with CLIP tweaks and block merging.
Credits go to Lodestone Rock and company for training such an expansive model. https://civitai.com/models/92450?modelVersionId=124661 https://huggingface.co/lodestones/furryrock-model-safetensors/tree/main
And pokemonlover69 over at https://civitai.com/models/124655?modelVersionId=136127 for humoring my insanity on this process.
And Feffy for their Fluffyrock Quality Tags LORA. https://civitai.com/models/127533