V8 is a mix of the past v7 v6 and v5.
V8 mix B has better knowledge of characters and need less weighting to get colored skin for example. Not that it is always correct but it has a better idea of who a character is.
Either way both can be interesting, curious which one you prefer.
Rule of thumb for both hyper versions, if you get artifacts either our cfg is too high or your steps too low. Hyper versions are Euler and Euler a only most other samplers don't really work.
Version 6 seems to behave with most samplers, not sure if I would recommend Euler a however.
Removed 5.1 it was too buggy so I removed it.
Not sure if Clip merging is such a good idea especially with Pony models.
Did remove the T5 clip merge as it looks like it does not work, or only has an effect with A1111. If you are using Comfy UI there is no difference between v4 with or without T5.
The real version 9 of my fast Pony merge.
I decided to publish this under a new name as it will where I try to get the look I want (more then in the other one where I tried to keep everything Pony intact).
I don't think I broke anything Pony like yet...
Will leave the name however as "fast" as it is based on version 7.5 and you can get decent results with 10-12 steps.
Don't use Euler a, especially with version 2 (at least for me it causes issues)
DPM ++ samplers are fine
I hope nobody minds the prompts I have been borrowing to se how mine compares.
Description
Merged with 12 steps hyper cfg lora.
Sampler: Euler or Euler a with SGM Uniform.
FAQ
Comments (35)
i don't know which one is better between this checkpoint and pony realism. Really good
Thx for your Work.
I have not much experience with pony models etc. lets see what your model does.
Please consider to give us more information about each update,cos of the recent fast updates its a little confusing.
To download and organize 7gigs of stuff every day is a little bit annoying. I am not sure about the idea to publish different hyper-steps-models eg. cos everybody can download their own hyper-step-loras. greetings from germany
Hi, yes sorry my merging process is a bit confusing and very iterative and I don't take notes what goes into every step. I will try to do better.
True about the hyper ones, I was not sure if I should add them or not.
I sometimes prefer merged ones myself instead of loading the Loras.
Also don't worry I will take a break with this one fairly happy where v5 and 6 is at.
@TiwazM hey i do not FORCE YOU to stop working!!!! ;-) A short break is okay and then back to work . Stay safe bro
@mscorpion7285 hehe sometimes I do down a rabbit hole and keep working :)
Hyper models perform WAY better than hyper LORA. That's why.
Also, hyper/LCM lora don't play well with other lora.
@Drakeni this info is new for me,in my opinion a seperate hyperlora would give you more flexibility.Right now I am playing arround with pcm that seems to outperform hyper.
Does this work in Forge UI?
It does, I use Forge. I'll upload a sample image in a minute, just working out the details. It's kind of finicky, but it can do nice things.
@Gouda if you are using the Hyper version yes it is
Yes. I have switched to Forge for all image generation and haven't had any issues with SDXL or Pony models.
v4_T5xxl, v5(T5), v6 all worked fantastic with:
—dpmpp_sde, karras, 8 steps, cfg 3-4
The hyper versions however are slower and output lower quality in comparison in all sampler/scheduler/cfg configurations I've tested it with.
I've compared those against the 12 step version, which already has a higher output quality than the 8 step variant.
Your model is too fast and furious already to be improved by hyper it seems. :D
If you want to look into further acceleration methods though, I can recommend the PCM LoRA a lot, which seems far more stable than hyper and maintains a higher quality even when deviating from recommended settings.
Hi thanks for the comment, I agree not totally convinced with Hyper, in my test it sometimes adds details but it is very finicky to use with high res fix. I find the non hyper versions "easier".
Thanks for the PCM tip I will have a look.
did you use these https://huggingface.co/wangfuyun/PCM_Weights ?
@TiwazM Yeah, hyper has this semi-fried effect that looks very detailed at first glance. Sometimes it can be offset by using dpmpp_2s_a sgm_uniform for the first pass, cooking the image even more, and then going for euler_a sgm_uniform in the 2nd pass to smooth that back out again. But doesn't always work and often still leaves this cloudy kind of perlin noise in the image.
The only really successfully implemented case I know is the Creaprompt series. That creator somehow managed to reduce hyper to only its pros and got rid of all the cons. That's magic, no idea how that was pulled off. Maybe that creator is willing to share some secrets.
About PCM, yes exactly that link was where I got mine from!
I've so far used low-cfg 2-,4- and 8-steps models all very successfully
Best settings I found so far:
• low-cfg-2-steps:
—steps: 3 (yes, 3 worked best, with 2 all you get is cooked)
—cfg: 1
—sampler: euler_a
—scheduler: sgm_uniform
• low-cfg-4-steps:
—steps: 4
—cfg: 1-1.3
—sampler: euler, euler_a, dpmpp_2s_a, dpmpp_sde (in order, euler being the best, same scheduler for all)
—scheduler: sgm_uniform
• low-cfg-8-steps:
—steps: 8
—cfg: 1.3-1.8
—sampler: euler
—scheduler: sgm_uniform
• normal-cfg-8-steps:
—steps: 8
—cfg: 2.5-3
—sampler: euler, dpmpp_2s_a, euler_a (in order, euler being the best, same scheduler for all)
—scheduler: sgm_uniform
@redpinkretro Awesome thanks for sharing. I agree Creaprompt is great and who inspired me to try it.
and I agree I have been playing with the normal cfg and I expected it to go higher but looks fried very quick.
oh and at least so far in my testing euler a seems to cause less trouble with the high res fix
@TiwazM euler_a is kind of the safe bet in most cases and usually smoothes things out even after going a little too far in a previous step.
2s_a is usually a bit of a gamble, like it doubles the amount of detail in your image and on top acts as if the cfg was increased by 3 points as well :D
I'm only using ComfyUI, but I can check some multi-pass configurations to uncover some more.
The PCM testing I've done was mostly during 2-pass workflows within the SillyTavern series I'm working on, which also has a lot of guiding elements like CNets and IPAdapters, so maybe that's what kept my 2nd passes intact. (hires. fix is basically a 2nd pass with upscale, right?)
@redpinkretro I have not really looked into Comfy, and not technically sure how highres fix works as it uses an up-scaling model too.
I am tempted to release an pcm lcm version, seems to work quite well with cfg 2 and 8 steps. Not sure if people are happy if I make them download another model... then again it is not just v6 with it... but maybe needs more work.
@TiwazM Ok, then it might be like the UltimateSDUpscale node, which also goes through an upscale model.
Something about using LCM with PCM was discouraged on the huggingface page. Not sure why, but they recommend euler and ddim as samplers only as if LCM would break something.
In my testing LCM was just very soft and slightly desaturated, looking like it needs more cfg, but frying when you try.
Yeah I get that, I'm also too lazy to download frequent updates sometimes, but if it's really different to the previous models, having a different combination of letters attached already does the trick to get most curious enough I think :D
Downloading v7.0 now. Very exciting to see a suggestion come to reality so quickly!
@redpinkretro I agree the LCM version is a bit strange so I did not use that. Not sure if 8 steps was the right choice or if my other Lora is too strong 😆 Perhaps I will experiment with the 4 step at some point.
@redpinkretro And thanks for the PCM idea.
I find 8s 3 cfg a bit soft sometimes too but might be the prompts, curious what you think as you dont use A1111 :)
@TiwazM Alright, I've tested some settings and am currently putting together some example comparisons to illustrate the differences between v40, v50, v60 and v70.
So far I can say v70 looks really good, but wait for the comparison to see for yourself :D
And yeah, there are some adjustments in settings I'm including as well for ComfyUI use.
@redpinkretro Nice thanks for testing, looking forward to it.
@TiwazM It's pretty incredible what distillation methods can do (especially PCM!) but I still think v6.0 offers the best quality and hope you will continue to provide un-distilled updates to this model. This and Godiva are the two best realistic Pony models right now!
@focalbluebell I can add the non PCM version of 7.0 if you want ;)
@TiwazM yes do pls v7 without any tuning :-) Cos everybody is able to download and use tons of pcm,hyper etc loras and can use them as they wish.I want the full,raw,bloody,pure realisticfantastic naughty pony. :-) Even if there is not much time for me atm to try pony-models att all.a dozen models starving on my ssds.
@mscorpion7285 sure will do, should be there soon :)
I know the SSD issue all to well :(
@TiwazM and bye the way. Who gave you the permission to call yourself Tiwaz? Justify yourself or Mistress Germania will teach you a lesson with the holly sword of Sauerkraut!
@mscorpion7285 Hehe, mmh I speak German and like Vikings :)
@TiwazM Jawohl. ok then it makes sense. :-)
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.