An experimental attempt to convert Pony v6 to a V-prediction model with fixed Zero Terminal Timestep noise schedule, as well as an attempt to bring some anime artists and characters while not breaking Pony's knowledge using an extremely small dataset.<h3 id="v0.1-9ujnld763">v0.1</h3><h2 id="you-must-set-sampling-type-to-v-prediction-and-apply-zero-terminal-snr-patch-otherwise-you-will-get-noise.-ptnk4abq7">You must set sampling type to V-prediction and apply Zero Terminal SNR patch, otherwise you will get noise.</h2>Artists: <a target="_blank" rel="ugc" href="https://files.catbox.moe/ajszqh.txt">https://files.catbox.moe/ajszqh.txt</a>Characters: <a target="_blank" rel="ugc" href="https://files.catbox.moe/hg99to.txt">https://files.catbox.moe/hg99to.txt</a>They don't work mostly due to how underfit this model is, while the main goal was to convert it to vpred. But you still can try raising weights on them. You can use this model in either ComfyUI or reForge.Lemme go on a tangent for a second. While baking this, I've come to conclusion that Pony v6 is actually a decently-trained model if you don't count the fact that it nuked most of SDXL's knowledge. While it's a lot harder to train for it, the end result is very stable and well-generalized. I've also started thinking that most people who tried to train Pony were doing it wrong mainly due to assumption that LoRAs are a fine and problem-free solution for SDXL. As to why this is not the case, I think one of the main reasons is well-explained here: <a target="_blank" rel="ugc" href="https://github.com/kohya-ss/sd-scripts/discussions/294#discussioncomment-10081465">https://github.com/kohya-ss/sd-scripts/discussions/294#discussioncomment-10081465</a> and below.

V-pony

vPony_v01.safetensors

v0.1

An experimental attempt to convert Pony v6 to a V-prediction model with fixed Zero Terminal Timestep noise schedule, as well as an attempt to bring some anime artists and characters while not breaking Pony's knowledge using an extremely small dataset.v0.1You must set sampling type to V-prediction and apply Zero Terminal SNR patch, otherwise you will get noise.Artists: <a target="_blank" rel="noopener noreferrer nofollow" href="https://files.catbox.moe/ajszqh.txt">https://files.catbox.moe/ajszqh.txt</a>Characters: <a target="_blank" rel="noopener noreferrer nofollow" href="https://files.catbox.moe/hg99to.txt">https://files.catbox.moe/hg99to.txt</a>They don't work mostly due to how underfit this model is, while the main goal was to convert it to vpred. But you still can try raising weights on them. You can use this model in either ComfyUI or reForge.Lemme go on a tangent for a second. While baking this, I've come to conclusion that Pony v6 is actually a decently-trained model if you don't count the fact that it nuked most of SDXL's knowledge. While it's a lot harder to train for it, the end result is very stable and well-generalized. I've also started thinking that most people who tried to train Pony were doing it wrong mainly due to assumption that LoRAs are a fine and problem-free solution for SDXL. As to why this is not the case, I think one of the main reasons is well-explained here: <a target="_blank" rel="noopener noreferrer nofollow" href="https://github.com/kohya-ss/sd-scripts/discussions/294#discussioncomment-10081465">https://github.com/kohya-ss/sd-scripts/discussions/294#discussioncomment-10081465</a> and below.

Files with SHA256: 13efbf38a7a687f45eb366bdc0e631cc10c6d5fb571f2b456018e7b8ab7f6f7a