GREAT SCOTT! Vivid insanity is here!
===========================================================================
KR34M - 11/25 - Let's give FLUX/KREA the DR34M makeover. This checkpoint excels at photorealism and abstract artwork. We hope you enjoy what's likely our last FLUX release!
Note: NF4 is actually K4_S in GGUF format. FP8 is actually Q8_0, also GGUF!
We recommend the BF16 variant + T5 BF16 if you have the VRAM.
Note: Use LoRA with C4PACITOR at slightly lower than normal weights for best results and to avoid anatomy blending.
===========================================================================
C4PACITOR models are created by DR34MSC4PE with all the trappings you've come to expect as well as some cool bonuses:
Enhanced realism and photo-realistic concepts, trained on high quality datasets with the latest techniques.
Specific anime/illustration tuning to introduce some lost artistic concepts
NSFW tuned and capable of realism and artistic/anime images. Female anatomy is well represented with additional fine tuning in the works.
Exceptional performance with character/other/stacking Lora
Like our work? Buy us a coffee: https://ko-fi.com/dr34msc4pe
===========================================================================
DR34MSC4PE is
c0ur4ge (training/qa/inference code) /
eraser851 (training/captions/tooling code/data/qa)
===========================================================================
Recommendations by base model:
dev - (d_v1/d_v2/a_v)
===========================================================================
CFG: 3-7
Steps: 22-60
Sampler: RES_2M
Scheduler: Beta/Beta57/Bong Tangent
========================================
Hyper8/L16HT - (x-series)
===========================================================================
CFG: 3-6
Steps: 16-32
Sampler: DEIS
========================================
schnell - (s_v0/s_v1):
===========================================================================
Note: 100% Schnell Base
Steps: 2-14
Sampler: DEIS
===========================================================================
Description
Happy New Year! As we wade into 2025, we're going to start sharing more variants from v2 (and beyond) that show off some promising runs and subsequent interpolations. Starting off with a taste of our forthcoming "X" (NSFW-first tuning) series: X-L16HT (light)
X-L1GHT is a 16-step Hyper style model that has a wild side while keeping the vivid madness you've come to expect from DR34MSC4PE.
Limitations:
Female centered tuning with limitations at extreme resolutions. If you notice body parts (particularly female NSFW) appearing where they shouldn't, adjust the resolution of the image slightly and try again. If it continues, increase the step count at the respective resolutions and it should usually give you something great!
CFG > 5 not recommended.
Recommended Setup:
CFG: 3.0 - 4.5
Samplers: DEIS (Preferably RES2M by ClownsharkBatwing)
Steps: 16-? (Experiment with this, particularly when targeting NSFW concepts along with aspect ratio)
FAQ
Comments (13)
c4pacitor D v2.
We've been using this for awhile now.
Nothing yet matches this.
The composition, color separation, concept art with regard to illustrations is outstanding. I continue to find new ways to use this and it continues to surprise me.
Thank you so much! We work super hard on our fine tunes and it's feedback like this that makes it all worth doing.
Hey, just a thanks for the NF4 models. I've been using them for the past few days, and all my most recent posts for dV2 and xL16HT have been made with the NF4s in Forge. I really like the quality of the images, and of course, the faster load times. The smaller size lets me go with larger text encoders and clips (fp16 and 32, respectively). You've done some really nice work with these. Rock on!
Glad you like them! I was really surprised with their quality holding up (and sometimes producing better(?) images than the others, in terms of aesthetics. Thanks for sharing!
So, i'm a noob. The NF4-all-in-one, should it not be able to run without vae's? And when I try to download the different models, they are all named the same?
They shouldn't be named the same, I don't believe but I do know that Civit messes with how the naming works when the files are uploaded. The NF4 variant I typically was testing in Forge (and created therein). The VAE is one of the things included but, admittedly, I've not tested if you need to bring-your-own VAE or if it was quantized and included in the Forge output. Do you have a workflow you're trying to using specifically?
Hi. If it helps, the NF4 is not an all in one, but a full model, meaning it was quantized from the main model to create a compact checkpoint. To use it, I include ae.safetensors, clip_l.safetensors, and whichever t5 I happened to have loaded. They all seem to work pretty well, and you can find them on civitai.
As for your other question, the c4pacitor models are all named the same by version, but you can change their file names. I use Forge so I just download them into different subfolders inside the Stable Diffusion folder.
DEIS and KL_Optimal are my sampling and scheduling methods. I favor high numbers of steps, but I prefer high quality to quick generation times. If you check out my profile, almost everything I've posted for the past few months has been with the c4pacitor NF4 model, so those posts are examples of what it can do.
My prompts and seeds are always included with the images, so if you like anything you find there, you'll have a place to start. It's pretty much all SFW, so no worries there. In any case, good luck with your generations! If you can, post images on the model pages to give back to their creators, and share prompts in the posts to give back to the community.
@arnesacknesson553 Thanks for the clarity - Arne! I use the fp8/BF16 models normally and am always impressed with what you make the model do! A few "successors" being tested so hopefully will have something new for you to try soon! I remember why I crossed my wires here:
I originally made AiO checkpoints but they were checkerboarding on certain larger resolutions for some reason so I ditched them (since the rest of the models here already followed a non AiO pattern) in favor of the slimmed NF4.
@arnesacknesson553 Yes, i got it working. Including ae.safetensors, clip and t5 did the trick. Thanks!
Hey which CLIP and VAE models do you recommend for this model?
Flux-dev standard are fine here but you could technically use a different CLIP-L for fun results!
@c0ur4ge what clip could you recomend and what is waht the clip L does??
@sergiogmx use the flux standards - T5 and Clip-L.
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.