V2
Based on Plus after resetting a bit the CLIP and merging back a bit of the first version, it keeps the crazy and excessive style while being more respectful of prompts. It will not "replace" the Initial commit which is more "standard", but i wanted a version more unhinged similar to the now gone Psychonex model.
Showcase pictures post contains all the Forge parameters used for generation.
PS: we are stuck until pictures can get rating before the model can be accessed 😭
In any case, i am also waiting for the model to be available on-site at next reset and then, a 2k followers bounty should happen 😉
Bounty is currently running! https://civitai.com/bounties/11070
Plus
This is not yet the final form of New Mecha, but i wanted to Improve upon Experimental to bring it up to speed on my own "style". The base idea is that i mostly use DPM++ 2M SDE Heun with SGM Uniform as a sampler/scheduler pair, which is not readily available onsite or in some other tools.
So, leveraging the dataset i made to make the ZIT version of New Mecha (which was generated using this pair), i trained a LoRA using Euler A as sampler and used it upon my own checkpoint to help it generate what i wanted with a simpler sampler.
This result in a less flat style even with Euler A, which may not fit all intended needs but is closer to my personal taste 😋
Usage: 40 steps using Euler A (CFG around 5-6) + a round of img2img at 0.3 denoise should give you similar looking pictures to the showcase one.
NB: It got also a bit funky regarding prompt following, which can give nice results but also be a bit frustrating. I'll see to that before a next release (probably with a real fine tune) 😝
Also, hands should be fine most of the time, but strange stuff can still happen 😅
Experimental
Long story: https://civitai.com/articles/23172
DMD2
Update DMD2: Merge of the DMD2 fast LoRA, it alter a bit the style but is nice nonetheless (switch to lower CFG, around 1.2, and 6 to 8 steps, including for Hires.fix and Adetailer)
New Mecha
I have too many models and PostBan was my attempt at creating something new... but i was not satisfied.
HoJ is my prefered style, but it is too messy.
AnBan is nice but too flat.
Both of them aren't getting me satisfied anymore.
That's when i did some tests with PleasureChest by @Ereijtic, together with the UniversalCLIP by @Felldude : by integrating them with my earliest version of HoJ, i was getting something that really like. But the result was broken. Eyes and hands were unusable, i clearly f**ked up something during the merge. That's why, i thought this could be good opportunity to try sd-mecha. By leveraging this tool, i was able to ensure my merge is consistent and i can share the recipe.
With the additional help of Indexed by @epiTune and Artistrious by @Jedas , i feel like i got exactly where i wanted to land (or at least i think so. Before i try to do a V2 and get frustrated again :D)
New Mecha will be my new main model going forward. No further versions will be made for AnBan, HoJ, PostBan or other miscellaneous models.
Here is the sd-mecha recipe:
import sd_mecha as sdm
sdm.set_log_level()
clip_uni = sdm.model("clip\\UniversalCLIPbase.safetensors", config="sdxl-sgm")
clip_v2 = sdm.model("clip\\illus2-clip.safetensors", config="sdxl-sgm")
recipe_clip = sdm.slerp(clip_uni, clip_v2)
recipe_base = (
sdm.pick_component(sdm.model("illustrious\\pleasurechest_v1.safetensors"), "diffuser") |
sdm.pick_component(sdm.model("vae\\sdxl_vae_extract.safetensors", config="sdxl-sgm"), "vae") |
recipe_clip
)
models = [
sdm.model("hoj40c.safetensors"),
sdm.model("shinv3.safetensors"),
sdm.model("illustrious\\indexed_v1.safetensors"),
sdm.model("illustrious\\artistrious_v10.safetensors"),
]
recipe_dare = sdm.ties_with_dare(recipe_base,*models,probability=0.6,seed=42,alpha=0.5)
lora = sdm.convert(sdm.model("..\\Lora\\000000\\HoJ4PC.safetensors"), recipe_dare)
recipe_lora = sdm.add_difference(recipe_dare, lora, alpha=0.3)
sdm.merge(recipe_lora, output="NewMecha.safetensors")Some notes if you want to follow along:
UniversalCLIPbase.safetensors is the merge of CLIP-L and CLIP-G from the link above, with the keys renamed to fit the SDXL expected naming convention (i ended merging it in a very small quantity, together with the Illustrious V2 CLIP)
sdxl_vae_extract.safetensors is the SDXL "FP16 fix" VAE, with the added first_stage_model prefix before the keys for the same reason as above
HoJ4PC.safetensors is an un-released LoRA i made by substracting HoJ by PleasureChest.
This is a messy recipe, my first attempt at using sd-mecha with almost no documentation (i found this doc in an other branch... but it's a little light :D)
Regarding usage... well, i think it is un-needed to provide any pointers. Because basically, use the checkpoint as you would any other Illustrious model usually and it should work :D
My usual go-to for local gen is either Euler A or DPM++ 2M SDE Heun and SGM Uniform, 20 or 40 steps depending on the amount of details i want, CFG 4 or 7 depending of the freedom i want to give to the model.
Two grid are available in the showcases pictures for demo.
You don't really need "quality tags", add the one you like. Sometimes, i just add IllusP0s, SmoothPos or dfPos. I made a "MechaP0s" using this model CLIP since it differs from traditional Illustrious a bit.
For negative prompt, i usually leverage IllusN3g, but hey, you can go crazy with your usual neg if you want :D
Adetailer is also a good help, but whatever facefix tool you like is fine.
Enjoy! (if you do a merge, a small link to here is appreciated.)
Disclaimer: This model was posted on Civitai ONLY. If you find this model elsewhere, know that it was not posted by me (but really i don't care anymore, just be aware that commercial use of this model is FORBIDDEN).