A Merge of the High Quality models and some loras for fine-tuning.
Focused on high-quality 2.5d and semi real images.
FP8 VERSION USES WAY LESS MEMORY! Bringing this one to FP8 cut the size in half but did not lose any quality. Way better suited for not so high end GPUs than usual FP16 Checkpoints.
VAE required for v1.0, there is no VAE baked in! Otherwise images look like negatives
Description
This is a new mix that is made in FP8! High quality and half the memory usage!!!
Recommended Settings
Scheduler: Euler / Euler a with Uniform or Polyexponential
CFG: 7 (5-10 )
steps: 25 (min. 10)
FAQ
Comments (21)
How did you do the fp8 conversion?
a little embarrasing, but I dont actuall know. It happened at some point and I only noticed it some merges down the line when merging loras did no longer work because of some fp8 calculation problems. Everything I merged into it afterwards was automatically fp8
Love the fp8 variant.
On another note - what I miss on all/any/every pony based checkpoint, is the "support" for more integrated backgrounds - and faces/races.
This is not a "your checkpoint" issue, but I'm sure those things could help to stand out of the crowd more.
Keep it up and thank you for your effort and sharing it!
PS:
I'd happily go into more detail / be more specific - if you want, but wanted to keep it short for the first read.
I actually merged a lora for better race depiction, African for example looks way better IMO. What backgrounds are you thinking off? many work just like that, but you have to prompt for them (minimum outdoor/indoor)
For background as example "waterfall" is one of the backgrounds that stands out on "repeating" itself.
Have to admit, since this initial post above, i've finaly got another one.
Until then (yesterday), I've gotten extactly the "very same" (very minor nuances, depending on seed and checkpoint) background sujet.
A very small pond in front, a small... erm.. translation error.. "uphill offset of earth".. before the grass floor continues in the forrest along the mini-"river" into the distance.
Maybe that "waterfall" is mostly (as in: more often) triggered when combined with characters, idk / not sure.
For races, have to admit i've recently only used my "bias" dynamic prompt, which "tends to show more light skined" characters. Dont get me wrong, not limited to caucasion, but every line has a priority for caucasion, or asian or middle-eastern, before allowing any other race or "combo" thereof (mixed race).
(where i live, there's more of those (asian/middleastern/caucasian) than blacks/africans).
So yeah, africans/blacks were shown better than expected.
Anyway, my reference to race was aimed towards middle-eastern.
And here again, just today I've got a (1) very good one, but could not get her to survive upscaling (she became "white").
By middle eastern, i mean egyptian, arabic, and people from like lebanon and iraq.
I made this image to test arabian phenotype. to me it looks authentic, but I am not overly familiar with it. What do you think? Also put some waterfall in the background that looks different than others IMO :)
It's a good idea to use FP8, but how to run it?
I've tried Automatic1111 and sd.cpp, both do not support data type 'F8_E4M3'/'float8_e4m3fn'. Any ideas how can this model be used?
It works with auto1111 1.9 and 1.10. it could be a GPU problem, older GPUs do not support FP8. Make sure your drivers etc are up to date.
@JohnnyB1 Tried the model with auto1111:
AttributeError: module 'torch' has no attribute 'float8_e4m3fn'
I was getting the same error.
AttributeError: module 'torch' has no attribute 'float8_e4m3fn'
Edit:
There is a FP8 setting to enable in the settings tab.
Video drivers were up to date, but there is also a CUDA toolkit: https://developer.nvidia.com/cuda-downloads?target_os=Windows&target_arch=x86_64&target_version=10&target_type=exe_local
Also, I renamed and moved the venv folder and added "set COMMANDLINE_ARGS=--reinstall-torch" to the webui.bat and then removed it after running webui.bat once.
@avscdf Did it help with the float8_e4m3fn error?
I'm using auto1111 with it right now
I realize this may sound dumb, but have you tried re-downloading the model?
I had (I'm guessing) a download issue on another model
This model is SO GODDAMN GOOD
The main drawback is it doesn't know a lot of concepts
I learn/test models by adding more stuff until it gets unstable, and you practically trip into amazing results with this
I have now spent a good while with this model, it is wildly underrated. This model performs exceptionally well with the lightning 8 step lora, and is very image stable with respect to other loras. Stable enough for some pretty sloppy wildcarding to work out well too!
it does have a strong style, and I have yet to try pushing too hard against it
It has an okay grasp of concepts, I am unsure which set of tags is best to use, but I've been testing with danbooru tags and it's been pretty successful
Highly recommend
Thanks for the kind words :) I use mainly danboor and derpibooru tags (with the Autocomplete extension). The derpibooru tags work best imo.
Oh, is the model explicitly trained via derpibooru? I read somewhere that danbooru tags were the way to go, any other tag types that work well?
@EliuLohm the base model is still pony, and as far as I know they used derpibooru. I am not aware of any other tag sets, that doesn't mean there arent any
Doesn't want to work for me for some reason. As soon as I select it and it loads, it will switch back to the previous checkpoint that I had selected.
gotta give us more information than that if you want help
Automatic1111? tried redownloading? what does your console say?
@EliuLohm I think the reason it's not working is because I use an AMD GPU version of A1111. I decided to give up because trying to enable FP8 bricked my install.
@Bureaucracy Sometimes this happens when the model has not been downloaded correctly. When you click the update button in the 'Checkpoints' section before it has finished downloading, it appears as downloaded, but when you select it it automatically switches to the model you had selected before it. Try downloading it again and make sure it has finished and you have it complete on the disk.