Regards for all your love and support ❤
avoiding negatives can improve result on SDXL
Description
Merged DMD2 Lora for faster generation and realism boost
Settings
CFG: 1 - 2 | Steps: 4 - 8 | Sampler: LCM (or Euler A) | Schedule Type: Exponential
FAQ
Comments (156)
A miracle!
Wow! "Last Fame" is a tremendous improvement! Thank you so much!
Please keep working @epinikion, don't quit please! You are making history!
still sdxl has no knowledge about 'tremendous' of compositions , human poses , even human real anatomy. he will only make history when working on fixing this missing knowledge in sdxl.
@amazingbeauty This model series survived SD3, SD3.5 and the Flux family. Still alive after almost 2 years of sdxl. That means something. Epinikion is doing a fantastic job and the images are incredibly authentic, have more character, in their domain, than Flux and 3.5. Its quite sad the t2i hype is dying out like that, but this makes me even more grateful Epinikion and a few other Creators keep pushing SDXL
@PrecisionMaster yes , still not conflict with what i'm saying , i try to say truths that been confronted by known creators! by blocking me! , it just a truth some believe and listen to it and some not. sdxl lacks a lot , still way more to even get in sdxl knowledge , creators must push into that.
interesting! So you are convinced that SDXL bears a lot of untapped potential? I'd love to hear more!
@PrecisionMaster no , i have no clue about this field , but if it can be trained then yes , if it cannot be trained on missing compositions and poses then there's no untapped potential
Waooooooooov
Can't use your checkpoint anymore:(
Please be back soooon on CIVITAI!!!!
Absolutely incredible
Weird comment:
I wish someone would make a DMD2 without the DMD2. I love the imagery that it gives, but I don't like the images that LCM. gives. Still, great initiative even if it's not for me :-)
Whats the difference between the DMD2 version?
DMD2 is a faster version. Less step, pretty much same quality as original.
@Reelai Ive only tried LCM and I thought it was terrible for image quality, so Im a little sceptical of of DMD2. Is there any place that shows comparisons?.
@fox23vang226 I did the comparisons in the showcase images. Same seed, prompt and resolution so u can compare between the LF models.
@fox23vang226 no - DMD2 Lora is best of the bunch, its invaluable if you take your personal free time seriously
@Narz Yeah, I thought why download a giant harddrive space sucking checkpoint when I can just download a 300MB lora and have the best of both worlds. Ive downloaded the DMD2 lora, but just havent had time to use it yet. I'll report back any noticeable prompt degradation or quality loss later.
@fox23vang226 What was your schedular while using LCM ? If it is not expo, it may give really bad quality. Also, because of less step requirements, you can go higher resolution to get even more crisp results.
@Reelai Just started using DMD2, and using the settings posted here, euler a and exponential. Im blown away by the results.
Ive posted my review, its shocking and mind blowing.
lastfame is the same lastfamedmd if dmd added manually as lora ?
You keep spamming checkpoints XD
(ok I just noticed the last one was from november...it is not exactly "spamming"... Sorry)
Love your models btw ;-)
Will be the last Version for this Model, need to do something else 😅 And thank you!
And yes it was kinda rapid releases, I switched the model versions a bit so the best performing models are at the top. So the last one was from March 4th
@epinikion thanks for your kind reply... You know my point... I don't like to "spend" 30mn on an image, to realise in the end that This image will be seen by nobody because everybody switched to the new-super-epic-thing ... Maybe you should just stay happy with your latest Epic CP, at least for a while, you really should, it s really good ;-)
@epinikion and I promise I won't do img2img with another CP XD I guess it s not cool for you...
@BretChampagne Haha everything fine, do whatever you want. I’m happy for everyone that has fun at using my models. I’m not here to tell anyone what to do or not. 😅 also appreciate your kindness
Edit: Found one major problem with DMD2, just as with lightning and LCM. Long prompts past 75+ tokens causes the output to completely fall apart. While on regular checkpoints I could use 300 tokens and it would be fine. Thats a huge bummer, 75 tokens is not enough.
HOLY MOLY?!!! is this sorcery?. How is this possible to get this kind of quality and speed with DMD2?!. After 7+ months of seeing DMD2 touted around I finally jumped on the band wagon today because Ive been so skeptical because lightning and LCM put a bad taste in my mouth because they sacrificed quality for speed. But I'm absolutely blown away by DMD2. It somehow manages to keep the same quality as a regular checkpoint while massively improving the speed of generation. Ive been generating pictures at 3 minutes each for 7+ months now like an idiot!. Now it takes like 10 seconds max and Im getting actually better quality pictures. HOW?! is DMD2 not standard going forward with all checkpoints?.
Granted Ive just started generating with DMD2 just for about 30 minutes, but just in 30 minutes Im blown away. Its like magic or sorcery, it shouldn't be possible to sacrifice that much speed and keep the same quality, its a glitch in the matrix!. Its so good that it feels like cheating and Im angry at myself for not trying it much earlier. I got some real FOMO debt anger going on now...the same feeling as when I missed the BitCoin hype train in 2010.
My settings -
Lora DMD2 at 0.8 strength
Euler A
Exponential
8 steps, 1 CFG
Kohya HiresFix
WTF?! I still cant get over how insane DMD2 is, how the F has this gone under my radar for so long?!. Normalcy bias? because lightning and LCM let me down?...
Funny! Do you think weight at 0.8 is better option instead of 1.0 - do need to test it out tomorrow
yes, for me 0.8 is best, at 1.0 the image starts to fry
Holy smokes ... I don't know what magic you are using but LastFame is something i have never seen before. The quality and prompt adherence is out of this world. You really cooked with that LastFame Model !!! Thank you so much for this Checkpoint. Incredible work...
Thank you ❤️ Appreciate your kind words. Hope you have fun creating
Yeah, this model is so good, that it has achieved the feat of being the only model I have ever left a comment on in the two years that I have been generating images. I enjoyed your models for SD 1.5, found better ones when I switched to XL, but this XL model right here has got to be at the top right now, and I don't know when someone else will come surpass it.
Appreciate it!
Can we never be done with SDXL? It's still the reigning king of realism, for all its flaws. The opportunity for creativity with SDXL is so much more than anything we've been given before, and Flux is great but it just ain't there yet, and much of its potential is locked behind the stupid t5 encoder.
Can you explain what you mean by "much of its potential is locked behind the stupid t5 encoder"? This sounds interesting, haven't heard about this yet.
@Celesta its still all magic to me but i have heard people say that the t5 encoder is an obstacle for training, uncensored gens, etc. I hope the smart people eventually crack it lol
really good! does anyone use this with flux? Im using generate img Flux > Upscale with realdream(ultimate sd upcaler), what config its recommended? to fix eyes/skin of flux model
No.
Not exactly sure what you're asking. This is an SDXL model, Flux is a different model entirely. Maybe you meant to comment this on a Flux LoRA? That's what it sounds like.
Same here. I used Epic Hades to refine Flux pictures for months, I'm doing few test for Lastfame DMD2 for same purpose
@sgmgong28901119 How do you make Flux into NSFW? Or u just make a lora of a character and use other basemodel to make the nsfw?
@TropicalCat try flux models that already included nudity, such as Flux unchained https://civitai.com/models/645943/flux-unchained-by-scg ,my personal favorite one
Has anyone managed to get this model running decently in Draw Things APP? We seem to be missing the correct scheduler (Exponential)? Would love to get this running.
Just try another one. I've tested one of the older model version and it works perfectly with other scheduler on drawthing.
Yes I know, some older models work perfectly with DT but they are older models and therefore do have their flaws. I´m particularly keen to work with @epinikion latest versions (Last Fame), as he is a true legend serving us well since SD 1.5.
@ClassicMood4 for fp16/8 models, (optionally add your loras) and then stack Hyper-8 at 60-100%, DMD2 25%, TCD sampler (sss 40%), shift at .60, CFG doesn't matter. < works with most XL/Pony/Il models out the box or with a little tweaking.
otherwise, use the LCM sampler with the Lightning model versions with low CFG, or choose a DMD2 version with an SDE sampler (or bastardize the above DMD2/Hyper configuration without the DMD2 lora).
@Narz Thanks a ton for your valuable tips, will keep experimenting.
What u recommend for best realism? have any lora recommendations?
Finally had a chance to try out LastFame. It's every bit as great as I expected from the sample images. Hopefully the admins will turn the on-site generation back on: I much prefer the upscale system here to the one on the other website that I used.
They said, probably this week the will roll out the new checkpoint feature.
When will this get back on the generator 😭I'm unable to use my fav checkpoints
It should be this week, no confirmation of what day though. They're working on shoring up the features
And thank you for your support ❤️
@epinikion Thank you for your response, and thank you for creating excellent versions :)
You've given the community some of the best realism models for both SD1.5 and SDXL... I think you've pushed it to the limit with this latest update, amazing photo quality and more diverse faces. Thank you for your service, you are honorably discharged from Stable Diffusion.🫡 Haha What's next?
Takes a bit of prompt engineering for consistent results, but the fidelity and variety in faces, poses, and locations is really impressive in VXVI.
This might be one of the best realism models I have ever used. Amazing job :D
Thank you!
DMD2 version - AWESOME Images - but occasionally - I get smeared images. It make happen on even subtle prompt changes.
Make sure your cfg is 1, steps between 4-8 and sampler is LCM (or EulerA, DDPM at steps 4)
Any prompting techniques to stop nipples from slipping out of bikinis?
Try in Neg Prompt “ nipples through bra/fabric”
Easily one of the best realism models I've ever came across, having a lot of fun and finding some methods.
absolutely great work, but I hope you ditch the very stupid roman numerals in naming.
You know what? This (as a illustrious model) will decimate all, after, you put about fifteen grand in it or more. If we have to, overnight parts from Japan.
Can it do consistent sex, poses, or just lucky results?
Yes I noticed its doing 'decent enough' nsfw poses as well.
can this generate males ?
These are really good. Did you use Flux to refine this model because in some ways it seems better than Flux.
All images generated in Forge UI, mostly 30 Steps, CFG 3 - 5, 2M SDE Exponential and hires. Recently posted images I use hires with same prompt and dmd2 Lora at 2x upscale and 1 or 2 Step with cfg 1. that’s it. No Flux, but dmd2 turns people towards flux style on hires if u go on cfg 4
And no, refining is based on training and trained photo Lora’s with high resolution
I am new in the game, what are the best settings in auto1111 to get such realistic results as the owner? I tried this one in comfyui but the results are not that good...
Better use Forge UI since A1111 is pretty outdated.
Just use "copy all" from the image, paste it into prompt textbox (or drag the image into textbox) and use the arrow under the generate button to read all settings into the UI. Generate! (also works for A1111)
What? A model with good to great output quality without having cleft chin butt chin on every single person? You broke the rule! Well done, very good model
Cleft chin is for "Flux checkpoints" reason to still exist! SDXL has had a long journey!
thank you for all your hard work over the years
The skin details and lighting without having to prompt for it is top notch. Great job.
For the gooners out there, the answer is yes.
Great model that is not just for tits and bums! There needs to be more like this. Cheers for your effort and dedication.
Thank you!
The effect of matching with DMD2 is very good. Will there be an SFW version? If you don't add costume prompt, there will be too many nudes.
Since I’m done with finetuning there will be no sfw version. Sorry.
Which settings do you use for the hi-res fix in Web ui?
LCM Exponential, 4 Steps, cfg 1, same prompt with dmd2 Lora added
Or Euler A Karras for better details, but less restoration
Sometimes. Need lower steps on faces for Persons with LCM
Having VAEDecodeTiled error?
Given groups=1, weight of size [512, 16, 3, 3], expected input[1, 4, 128, 32] to have 16 channels, but got 4 channels instead
Wow! SDXL stining photorealism, perfect lighting alignment, even with complex prompts, combined with the pony's "special" abilities.
Thank you so much...
I really love the SDXL model you created; it's the most realistic one I've tried for generating images. I'm wondering if you would consider training a base model with Flux. I noticed there are only Flux-type LoRA models available. Is it because the base models trained with Flux aren't as realistic as SDXL, or is there another reason? This is really important to me, so thank you for your explanation
Flux issue is there license is pure bs
@angelcookies What is the problem with the FLUX license that keeps people from creating checkpoints?
I am having this problem on google colab when loading the checkpoint:
{"error":"Unauthorized","message":"The creator of this asset requires you to be logged in to download it"}
Also getting this problem:
HTTP request sent, awaiting response... 401 Unauthorized Username/Password Authentication Failed.
How can I use this checkpoint if I am using google colab (Fooocus UI) to run it?
i turned of the authorization for the newer models, to avoid the hassle
Wow 2 years of epic checkpoints! Flux just takes so much time on my gfx-card. Hope Pony-7 will be the future.
any tips for better gen?
i'm quite satisfied with dpmpp family (2m, sde ~ 3msde with karras or AYS scheduler, at 2.0-3.5 cfg, 30 steps) for realism. LCM, DDIM, and DEIS + SGM_UNIFORM, DDIM_UNIFORM , Karras work well too for 2.0~3.5 cfg at 16-30 steps. If using hires fix, 10-16 step with 0.2-0.3 denoise plus 4x ultrasharp upscaler works great too. Definitely better than without.
Awesome
Can anyone shoot me a link to a ComfyUI workflow I could download with settings that will work well for the VXVI - LastFAME (Realism) model? Definitely struggling to get results like the examples.
https://civitai.com/posts/16470373
download the image, drag it onto your comfyui, workflow embedded
Nice :)
EpicGOD❤️
Fantastic, this is the best model I've ever used, this deserves a flux version
Cool model for realism, its quite horny. After using it for some month, i summarize it as good for generating amateurish realistic photo style (not quite what i'm looking for), but it definitely did a great job for some time. Works quite well with ipadapter too, which is another plus. There are just too many model that has too strong face bias and won't cooperate with ipadapter.
great model
This one is so good
Getting excellent results, you have brought me back from Flux...👍🏻
Glad you like it, if we just could have the t5 from flux for Prompt adherence, that would put SDXL back to Top
true words. been waisting my time with flux
this is probably a stupid question, but is this a "base" or "refiner" model? I'm new to all this, and just learned that you have to use both. I'm using automatic1111 and idk what I'm doing. Also do I need a vae?
It's a base model as the tags suggest. VAE baked in and refiner isn`t required by any "newer" models anymore
One of the best realistic checkpoints
Ghhyii
best one I've ever used for drawing people realistically
So realistic that I am kinda surprised
question.. what would be the best prompt to have a woman where she has nipple print through clothing? Thats a tough one and wanted to know if you or anyone else has answers.
IF you mean getting nipples poking through clothing then you can use LORAS such as:
https://civitai.com/models/915086/nipples-under-clothes-10
https://civitai.com/models/181510/poking-nipples-by-stable-yogi
https://civitai.com/models/550414/nipples-through-white-wet-top
great value for buzz...
I have no idea how you get these results, I'm using forge and copying your parameters and it's nothing like yours. I only get decent results when using 7 CFG
Lets figure this out.
I'm using this version: version: f2.0.1v1.10.1-previous-662-g0ced1d0c • python: 3.10.6 • torch: 2.3.1+cu121 • xformers: N/A • gradio: 4.40.0
UI Setting: XL
Diffusion in Low Bits: Automatic
Settings:
Random number generator source: GPU
Lets start with this and see if it fixes something for you
@epinikion thank you for responding, I'll try it
he has a private comfy workflow that he strips the metadata from the images before posting
@GeneralSpacey i'm not even using comfy neither do i strip any generation data 😂
@epinikion @rodj I also often get comments from users saying they can't reproduce my images exactly. I also do use Forge (reForge) with some customized settings, and I even believe the GPU differences also affect the final result.
@Cyberdelia yes this does make a massive difference. Once put my 1050ti into the computer from a friend of mine instead of his 3060rtx and we ran 10 sdxl and 10 pony prompts with the exact same settings. The basic scene mostly stays the same but light and details sometimes closeup and full body stuff was completely different
@Cyberdelia different architectures use different math libraries/implementations and kernels get assigned differently and so on so I'm sure that even though a lot of it is standardized there are differences that account to (perhaps literally) rounding errors outside of the precision enforced by standards. If that's the case I'd imagine that causes minor variance in result on different archs, and I'd also assume some of those may compound
that's a half-educated guess, I'm actually really curious now about where the diffusion algorithm etc may be subject to architecure-based nondeterminism
@Cyberdelia @tcpstackmaster971 Random number generator source.
(changes seeds drastically; use CPU to produce the same picture across different videocard vendors; use NV to produce same picture as on NVidia videocards)
that is why we should stay wit CPU RNG, i set it back in my settings now
Can anyone share a workflow that makes it work?
this thing is incredible, i can't believe my gpu is making the things i see
Out of all the top realistic SDXL models on this site, in my opinion this is the best. Simply stunning. Great work.
TY!
Switched to Forge. I kneel.
to say it is much more difficult/worst on Forge? :)
Just wow!! It follows prompts better then any other checkpoints and the images are incredible! Thank you!
How do I get to zoom out a bit? All images seem to be very close to the character like a selfie
Try full-body or waist up
Or, try a controlnet.
@epinikion Or cowboy shot, or include an item of clothing eg belt
your example Images are SUPERIOR. Just need to download you safetenors, and put it to checkpoints folder, no need another node?
This is definitely the best SDXL model I've used. Youmust not stop with SDXL, please!
I have used LastFame extensively now. I was wondering, will you be continuing in the general 'direction' of LastFame? or next model will be something else. Hope it is the former : )
Thanks.
Best model I've tried by far. Also, the first model I've tried of 20+ models where characters can actually hold a sword without it being a handy mess.
How do I stop my character from showing her boobs. I have tried with ((nude1:2)) and words like, sfw/non nude.
Prompt for clothing
@epinikion I did that, I tried with white short on / black bikini. Also captions like white shirt covering her breasts
@elysiumfinest you would have to share prompt to see what it does
"breast" or "chest" are generally better than "boobs" or "tits". Cleavage helps. And nsfw, areola, nipples, etc in negative.
@SleepyAragorn doesnt work she still shows both her boobs
@SleepyAragorn i put everything in the negative prompts still both breasts are out
It tends to produce images with squinted eyes.
Adding "uneven eyes" and "squinting" in negative seems to help moderately.
Also, having "detailed eyes" and "detailed face" in prompts seem to be mandatory. Without them the results have bad faces.
Details
Files
epicrealismXL_vxviLastfameDMD2.safetensors
Mirrors
epicrealismXL_vxviLastfameDMD2.safetensors
epicrealismXL_vxviLastfameDMD2.safetensors
epicrealismXL_vxviLastfameDMD2.safetensors
epicrealismXL_vxviLastfameDMD2.safetensors
epicrealismXL_vxviLastfameDMD2.safetensors
AIIM_Realism_FAST.safetensors
kx7_m_e69e1424e0.safetensors
AIIM_Realism_FAST.safetensors
kx7_m_e69e1424e0.safetensors
epicrealismXL_vxviLastfameDMD2.safetensors
AIIM_Realism_FAST.safetensors
AIIM_Realism_FAST.safetensors
Realism.safetensors
AIIM_Realism_FAST.safetensors
AIIM_Realism_FAST.safetensors
Available On (2 platforms)
Same model published on other platforms. May have additional downloads or version variants.