This series of the models was started (v0.4) from DucHaiten-Pony-Real (beta) by DucHaiten.
Version v1.0 was trained using a LoRa mode in One trainer. Next version are trained in Fine Tuning mode (full SDXL training).
Up to v1.0: Photorealism, lighting and anatomy related LoRas from CinEro series are integrated into this model in order to amplify cinematic render.
Goal of training was to improve to stability of anatomy and make the picture more realistic and moody / cinematic.
Version 1.1: shifted to fantasy / sci-fi themes with less realistic and more contrast / unstable rendering.
Version 1.2 [SOON]: is v1.1 further development with shift to horror / tentacles theme.
PS: be responsible and don't abuse this unrestricted model.
Description
Improvements: creativity, anatomy, adherence (natural, WD).
FAQ
Comments (28)
very cool model! i highly recommend considering merging it with XenoDiffusion. its a model that specializes in this type of stuff ;)
Will try to check it. Thx.
W8 a minute... It is an SDXL.. The chances are this merge will be a total mess. Pony and SDXL are not "phenotypical". I have no good experience merging Pony with SDXL. But we can try anyway.
@homoludens I've heard that there are ways to successful merge clip between Pony and SDXL, but I have been unsuccessful in my own experiments, nor have I found out how others are doing it.
I've checked the XenoDiff... First images with the same settings as for CinEro Pony v1.2 was quite noisy. It means that CFG and/or sampler needs to be changed. And Prompt is less compatible.. Also, the prompt with human-animal hybrid resulted in less animalistic result.
BUT
In other aspects image was more horror themed and quite interesting. Worth trying to mix.
@homoludens if your a comfy user, consider block merging, ive even block merged an sdxl realism model with a pony anime model, and essentially put pony prompting into the realism model without the loss of style!
@funnelweb656 yup. i do things 24/7 as part of whats called block merging
@BakaBlitz Of course MBW, but even with that.. I tried to make elemental merge of Pony and SDXL several times. It was a simple merges like double or triple sum. I also tried a mode "Add difference" with "Extract" mode. And results was far less satisfying than in case of Pony x Pony merges.
@BakaBlitz Comfy merger is cumbersome but it allows element-wise merge instead of whole blocks. But I noticed that it produces corrupted models which confuses A1111 webui.
@homoludens strange. i have not experienced that tbh. later ill experiment with block merging your model and xeno, if i find anything useful ill let you know
@BakaBlitz It would be good to see if you succeeded...
@BakaBlitz I started to work on the merge. First attempts looks promising.
@homoludens very nice! hope it does wonders. xeno is a really fun model
@BakaBlitz I realized that we cannot have a merge with XenoDiffusion. It's license does not allow anyone to share the merges. I am sorry, but I should throw away this merge...
@homoludens aw man sorry. it sucks that us creators were given this tech for free, under a free license, but choose to create our own bad licenses over them to stop other creators :( best of luck bro ill keep up with your model
@BakaBlitz NP... I worked it around already. Uploading the merge...
What's the purpose of the separate config file and does it need to have the same filename?
Honestly, I cannot tell for sure if it really necessary. OneTrainer generated this file. And I keep it with the same name in the same folder as .safetensors file.
When I was loading this model in Foocus I saw the warning that some of the keys from model was ignored or omitted. These warnings gets produced occasionally and it can be some glitch in Foocus.
But I suspected that config may contain a better instructions how to load this model and what Python modules to use for it.
This is why I decided to add this config just in case. I recommend to grab this config also. It should not break anything.
A1111 WebUI loads this model as usually.
@homoludens Thanks for the info! Never noticed such a thing for any of the other models here and was just wondering. I'll give it a try with and without and see what it does. Been testing out all of your CinEro Pony models for months and really enjoy working with them. Think I still like the IQ/adherence of v0.8 the best, but the new v1.3 series has been pretty good as well. Appreciate your efforts!
@AFD_0 Thanks for helping. I'll think about v0.8. Maybe it worth to mix in this version back to v1.3 and make a special branch for test.
@AFD_0 I noticed that A1111 WebUI load time for this models with config is significantly higher. Didn't test the same model WITHOUT config yet. But other Pony and SDXL models load faster. Don't know why yet.
Thing I noticed was that promoting on a lot of point models for exaggerated hair color.. like neon pink hair doesn't work. Would be nice to have some extreme and unnatural hair colors too. And the token 'colored eyebrows' sometimes works on anime models but rare on photographic models. Would like to see that more on pony.
Things like that works better as a LoRa. We should think carefully before incorporating the vibrant stuff into the realistic model. It might make this model too sensitive for high CFGs.
@homoludens yeah you're right but there's no lora that does it, could be that the token structure in your model might not allow for it to be its own thing but there are sdxl models that do it without messing up the rest of the hair.
Great!
omg please make this exact thing for illustrious
To this day I still use V1.0 for most of my realistic Pony work. Other models either lack realism or are realistic but no longer work with PonyXL loras.
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.