Pony ports of <a target="_blank" rel="ugc" href="https://civarchive.com/models/606854/hybridmixsdxl">HybridMix_SDXL</a> by merging in the CLIP text encoder of the <a target="_blank" rel="ugc" href="https://civarchive.com/models/751465/everylora">EveryLoRA</a> checkpoint by advokat. Will be doing further Pony merges going forward, which will be posted onto a new model page. This model page is for models ported from SDXL with some NSFW capability.<hr /><h3 id="edit-methodology-used-jhxeg73ty">Edit - methodology used</h3>(Added to description for more visibility since it was asked in a comment)Simply merge the CLIP (text encoder) with the one from EveryLoRA, advokat hasn't released any details on how the CLIP for that model was created but it works well for the purpose of merging SDXL with Pony6 models. This is controlled by BASE (Supermerger) / CLIP (ComfyUI). UNET (image generation model) are the remaining layers (Supermerger) / MODEL (ComfyUI).Both 1a and 1e are simply HybridMix_SDXL Mod7_050 UNET with different ratios for the CLIP. Merge ratio depends on what you want to achieve, higher Pony / lower SDXL ratio is more unstable but makes the model align much more to what would be expected of Pony models, whilst lower Pony / higher SDXL helps to retain the style of the model as with what I did in this merge.However, UNET definitely requires more Pony to be merged in to get better results for NSFW, unless the SDXL checkpoint(s) being merged already has good NSFW knowledge, which mine do not.<hr />

HybridMix_Pony6_Port

hybridmixPony6Port_mod23Alt4.safetensors

Mod23_alt4_Intermediate1e

Mod23_alt4_Intermediate1a

creator: aaronchmPony ports of <a target="_blank" rel="ugc" href="https://civitai.com/models/606854/hybridmixsdxl">HybridMix_SDXL</a> by merging in the CLIP text encoder of the <a target="_blank" rel="ugc" href="https://civitai.com/models/751465/everylora">EveryLoRA</a> checkpoint by advokat. Will be doing further Pony merges going forward, which will be posted onto a new model page. This model page is for models ported from SDXL with some NSFW capability.Edit - methodology used(Added to description for more visibility since it was asked in a comment)Simply merge the CLIP (text encoder) with the one from EveryLoRA, advokat hasn't released any details on how the CLIP for that model was created but it works well for the purpose of merging SDXL with Pony6 models. This is controlled by BASE (Supermerger) / CLIP (ComfyUI). UNET (image generation model) are the remaining layers (Supermerger) / MODEL (ComfyUI).Both 1a and 1e are simply HybridMix_SDXL Mod7_050 UNET with different ratios for the CLIP. Merge ratio depends on what you want to achieve, higher Pony / lower SDXL ratio is more unstable but makes the model align much more to what would be expected of Pony models, whilst lower Pony / higher SDXL helps to retain the style of the model as with what I did in this merge.However, UNET definitely requires more Pony to be merged in to get better results for NSFW, unless the SDXL checkpoint(s) being merged already has good NSFW knowledge, which mine do not.

Files with SHA256: 09a258fb969ca1c453d084da7b4c9f3005a360f89817f2fb3f5c2b382d166009