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
If my calculations are correct, when this baby hits 88 miles per hour... you're gonna see some serious shit.
FAQ
Comments (92)
12-42 steps needed for a Schnell-based model?? That doesn't sound right. Is there any Flux Dev mixed in to this model? If so, then you need to change the license to the non-commercial one.
Good catch! The fine tuning was done against dev but this checkpoint is the result of starting with (more) Shnel than anything else. Updated to be better safe than sorry!
That said, you can absolutely get useable images out of this model at lower steps than recommended but I’ve done very little testing below that step count. Looking forward to seeing what it’s capable of!
@c0ur4ge Yeah, unfortunately any amount of Dev dna in the model is going to trigger the non-commercial license requirement. At least that's my understanding at the moment. Just wanted to make sure I understood what this model was based on. In any case, I look forward to trying it out (as soon as it's done downloading, which might be a while). Thanks!
@Grumblebutt You’re welcome! Please share some of your creations.
Schnell varient is up :)
@c0ur4ge Just tried it and it works good, nice job. The only critical feedback I could give is that, by default, I would say that the color and contrast are a bit muted, if you know what I mean. The images come out a bit soft. However, once I bumped the max_shift up to 1.50 that seemed to raise the color saturation some. Haven't figured out the right settings yet that work best for me but I'll probably play around with it a bit more.
@Grumblebutt Thank you! I really appreciate the feedback and (I think) I know exactly why its doing that but short answer is a large part of the content is shot with DSLR/Analog cameras. The current follow-up in training has some more "traditional" diffusion dataset aesthetics. For posterity, I'll release the DEV variant and the LORA extracts before focusing on v1.
@c0ur4ge Awesome, can't wait to test drive the next version too
When using the mixed Dev/Schnell version it usually takes 14 steps to reliably get a clear image every time. With the all-Schnell, it does fine with 8. In most cases it takes less, but I want to make sure every image has time to complete.
@lesjo Thanks for confirming. I’ve been sticking to 6-8 mostly for schnell but I’m not usually displeased with the results of 4 either!
a HyperFlux version would be nice
Sure thing! I'm planning on putting out the Dev version and a bonus LORA in the coming days. I'll look into it. I've actually not even really had a chance to play with it yet so I'll need to look into the process. :)
Looking into this, btw - If the quality doesn't completely tank (or experience concept loss), I'll upload the dev variant!
I was having some problems with their LoRa - it seems like it only works with purely Flux dev and must deeply integrate with specific layers/weights. For now, I'd advise just making use of the FUX LoRa (within these checkpoints proper)
Hyper has arrived. :)
Interesting model, am I asking too much for an fp16 version too? Thank you.
Sure. This should be easy enough but I’ll need to make sure everything looks good.
Just an FYI - all is looking fine here so I'll be starting the (massive) upload shortly. This will be v0's final edition - I'll holler when done. I like your workflows and look forward to seeing what C4P can do with 'em!
FP16 of Dev is up @Daedalus_7 ! Let me know what you think - its one of my favorites so far! :)
@c0ur4ge Thank you! I'll have to test it for a bit :)
@c0ur4ge I'll make sure to post an image. Thank you for the upload!
i created gguf models for others to download ;)
for people with 8gb of vram
Wild! How do they perform? Happy to upload them here and credit you if they’re looking up to to the challenge!
@c0ur4ge Q8 is 99% indistinguishable from the real fp8 model and works with 8gb vram. and Q4 has a slight quality decrease but is between 15-20% faster than fpq/q8
@pkmngotrnr Excellent - I've unhid this comment for now. I'll run these through the QA cycle and get them added in the coming days. Thank you @pkmngotrnr - did you just do a standard GGUF quant with llama cpp? This is something I should probably get more familiar with doing!
@c0ur4ge yes i did exactly that ;)
I use DrawThings, only way to run SDXL or Flux on an 8GB M2 mac. Unfortunately, Draw Things doesn't support gguf, yet.
Amazing, Thanks, going to try those asap.
What is gguf?
Could you upload just the UNET without VAE or text encoders? Nearly 23 gigs seems unnecessarily huge.
This is a normal FP16 C4PACITOR D (dev) UNET release, same size as the official. The FP8 version at half the size will be up later this evening or tomorrow! For now, this checkpoint can be run in FP8 mode if you can load it initially! I appreciate your patience.
Had some spare time so managed to get the FP8 upload done. Enjoy and please share your creations!
Great, thanks! Downloading now...
Schnell version Works on DrawThings, M2/8gb, Euler A/4 steps = 3 minutes
Great news! I'll do my best to get a Schnell version of v1 out after taking care of the dev variants. Please share your generations and thanks for testing!
New Schnell RC is up for v1 :)
what is optimal value in webui forge? i get only crappy outputs.
@cornix9 For Schnell? They’re listed on the resources below but essentially 4-14 steps. Nsfw concepts often require trying varied step amounts with the RC of s_v1 as it’s undertrained on those concepts.
s_v1 proper is being tested and will be out shortly so it may be worth it to check back tonight or tomorrow if you’re primarily using it for “anatomical correctness”.
Q5 please!
Will fp16 works on gf 4080 and forge?
It might! I’ve not tried it in forge yet but it for sure works in comfy. Please let me know!
@c0ur4ge Wanted to use in Forge but error:
AssertionError: You do not have CLIP state dict!
You do not have CLIP state dict!
@Tozi_White So I believe this is where you need the clip_l and T5 checkpoints also (just like most flux models).
@c0ur4ge Any tutorial for flux-newbie?
can this model run on forge?
Yes! If you're not familiar with how to use FLUX on Forge in general, I believe it was a recent change and you can now provide the three checkpoints (C4PACITOR, CLIP-L, ae and T5 of whatever variant) and off you go!
Just remember to use DEIS as the sampler. I took way too long to figure out that Euler (which is default for Flux) was the reason it was spitting out garbage. And even with DEIS (as advised) it was quite unpredictable with many undesirable results.
Also - I don't know if I was doing anything wrong, but using the C4PACITOR model had Forge using something like 50GB of RAM and all 24GB VRAM.
is loras work on it?
For sure!
Yes, but you have to be very careful with the weights of your LoRAs because Flux in general (not just C4PACITOR) has a marked tendency to go crazy if you try to push it too hard in a given direction or if there are conflicts between the way your LoRA wants to go and the way the checkpoint wants to go.
I've found that a weight of 0.69 for your most important LoRA and slightly less (0.68 is fine, it just has to be less) for all subsequent ones should keep from running off the rails too often.
@lesjo i tried the same sittings many time on this model and original 8 model and with original was loras work very well and with this model it gives me something else like example it's given me someone looks like my character like same eye colour or big lips but not the original character i don't know is something I did or that's what the model do with loras in general 😔
@NoobFromEgypt This is almost certainly because our captions/tokens overlap a bit - future versions are taking this into account and reintroducing some dev default layers to help combat overfitting of character Lora’s in later versions. Stay tuned!
@c0ur4ge Thanks and good luck 🤞
Loras not working
@tupu Loras are absolutely tested prior to release - can you say more about how and what you are trying to do?
@c0ur4ge Yes, im on forge, flux1-dev-bnb-nf4-v2 works quite well with lora, today i tried same setup with c4acitor and the result is not good at all, also takes to long, i have a rtx 3060 12gb and 16vram...any adivice?...thanks for answering.
@c0ur4ge character loras dont seem to work at all for me. i tried a bunch of different characters. Of the 23 random images i tried to do, only 1 image came out ok-ish. the rest of the images was completely incoherent, bodies deformed to smithereens, etc. my prompt is as basic as " character_trigger_word<lora:character_name_here:1> nude in a serene environment <lora:flux_realism_lora:1> " . Also using Forge, sampler euler simple 20 steps 3.5 distilled guidance. nothing crazy. as soon as i remove the character lora and do "girl nude in a serene environment <lora:flux_realism_lora:1> " it works perfectly. If i switch to base flux dev, the character lora's work fine.
here are the GGUF q4_0 and q8_0 versions of the c4pacitor_dV1BetaRC1Fp16 model:
q4_0:https://pixeldrain.com/u/VYxqvBH9 6gb instead of 22
q8_0: https://pixeldrain.com/u/oA25EhYy 12 gb instead of 22
can you upload these somewhere else ?
@Mick0 I’d prefer we don’t have a bunch of bootleg quants advertised here. If preferred, I can upload them to the page itself but is there a reason you’re asking?
@c0ur4ge pixeldrain limits download speed after 4gb i guess.
@c0ur4ge If you could add them here, that'd be great, especially q8. Yes, there's a limit for pixeldrain, 5GB per day.
@c0ur4ge 4gb limit is all
@green_anger Sure thing - if @pkmngotrnr could be so kind to provide them there, I can upload them here! I just won’t have time to quant them myself in the coming days.
@kritzonly3 ooof. Understood. I’ll Download the existing and upload them to their respective tabs too then.
Can this model be used to train LoRA?
It COULD but I’m not sure what the outcome would be! I’d naturally make sure you use a FP16 if you try! That said, you can use Lora trained off FLUX.Dev just fine as well!
The dev full model seems to make really awkward bodies with extra limbs and stuff for me. I think the beta was actually better.
I’d suggest trying different resolutions/step counts - the Beta is likely aesthetically a bit better, v1 has better NSFW concept knowledge but is, admittedly undertrained still. v2 is in the works as a full checkpoint tune vs Lora pipelines.
If you are using Forge, update your copy and then try Forge Realistic (slow) or Forge Realistic for the sampling method, or you can use DEIS, although I don't find that to be the best on Forge. Also on Forge, use Normal rather than Simple schedule type, and consider 25 steps for the render. Those settings should get you a good render. Afterward, you can dial the settings back if you want it to go more quickly. I hope this helps.
@arnesacknesson553 Thanks for this - I really like DEIS in comfy but this is not the first time I've heard of some "performance differences" between the two. Thanks for the info/ helping out!
@c0ur4ge Hey, thanks for the kind words, and the buzz, but really, I should be thanking you. I like your model, and I appreciate your posting it. I just thought it was a shame that some of the newer folks weren't getting the good results I know it can render!
Please give the new rebalanced v1.1 a shot! :)
@c0ur4ge Ok, will do, thanks for making it.
@c0ur4ge I tried v1.1 just now - did batches of 4 images using the same prompts I've been doing today with other models. Mostly I am using Ipmdm+beta but I tried euler+simple, euler+beta, deis+simple, deis+beta too. I'm getting the same pretty faces with mangled bodies as before and on all 4 images on each batch. I tried with and without lora. Mangled bodies happens sometimes with all the flux ckpts but it's not the norm. I'm using comfyui but I don't know why it would be different than forge. Sorry I don't know what's up, but it's not improved for me. I'll leave the model on my hard drive for a few days. If there's something else I should try let me know.
@EricRollei21 I would suggest trying with different resolutions then. What are you presently trying? And with what step counts? If you’re asking for something it “doesn’t know” from a nsfw/anatomy perspective this can happen. Please ensure you are using natural language to prompt C4PACITOR for such concepts as well - while the traditional "1girl, solo, nsfw" etc work, they tend to have more issues like this.
@c0ur4ge I'm using natural language prompts for T5 clip, usually flux guidance 3.5 (real guidance 1) with 40 steps. My prefered combo is ipmdm+beta. This works very well for many other ckpts that I use. Can't figure out why C4pacitor is mangling things. I'd post comparisons but they are ugly.
@EricRollei21 Send me a DM - would be useful for me to see if I can find any potential culprits in the prompts and ultimately, the dataset to prevent them from making anything worse later.
What resolutions (and if raising the CFG to 4.5 helps) as well would be helpful! Thanks for the feedback!
I think your NSFW training is too strong and decreases the quality of anything else.
Most often results are glitchy or uncanny weirdness, but often NSFW triggers with innocent prompts. For example, I was trying to render "close up photograph of granny stealing oranges" and she was having BJ instead of OJ.
Beta is slightly safer, but still lowers quality too much.
Thanks for this feedback (and laugh) - v1.1 takes this to heart. We hope you’ll find it balanced but still capable of these … unique abilities at higher guidance settings.
The unfortunate truth is that so-called 'training' of LLM or Stable Diffusion models actually ruins them. OR these 'modified' models should only be used for their 'new' abilities. With image generation, it is almost always better to use LORAs.
@blobby99 This is only somewhat true - in the case of these initial versions of C4PACITOR, its the result of merging trained lora layers between various checkpoints/back to the base model, etc. That said, if the "new ability" is half NSFW (but HQ) content and half stylistic tendencies you want picked up (essential reg images to a degree) it doesn't inherently wreck things.
Where v1 "comes on strong" is due to the strength at which certain layers were merged with a focus on anatomy and not the preservation of other things. 1.1 is based completely on Dev/Schnell with less potency on the NSFW concepts but enough to wear they can be coaxed out at different step counts (Schnell) and Guidance (dev).
To the point, V2 is a full checkpoint tuned on LOTS of stuff and has no such encumbrances thus far.
lol this made me download this
Can you put out an NF4?
Hey there! I actually do plan to do a bump release along with another variant of v1 while ironing out some "unforeseen" issues with our distributed training setup. I'll see about getting this handled in a few days if I can get the time!
@c0ur4ge Nice!
I have a 4080 Super so unfortunately I can't use any of the current ones without running out of VRAM; but I've been able to use NF4 FLUX models.
I'd appreciate it!
@SickMoonDoe Another suggestion would be to load either of them as FP8 - they support it if you have the system memory to load the checkpoint first. Should put it in range (tested in Comfy)
@SickMoonDoe You can absolutely load the fp8 with a 4080 super and even use Loras with comfy, I have tested myself as I have both a 4080 super and 4090, part of it also is how your launching try with --force-channels-last --dont-upcast-attention --normalvram and also try PyTorch attention with version 2.4.1 cuda 12.4 setting expandable segments (it’s a environmental variable) instead of xformers and using any special cuda variables. Edit: though yea it means allowing offloading and sharing resources with cpu but I don’t know what you have paired with your 4080 super but my 5950x and higher system memory seem to pick up the slack with my 4080 super far better than I expected. I got a 4090 from work and was surprised when I realized I didn’t have it as bad as I thought i did
@joehorse it might be time for me to switch from InvokeAI to ComfyUI. It crashes for me in InvokeAI with a VRAM OOM error :'(
@SickMoonDoe honestly I don’t always like comfy but man do I like quality and you can always rip workflows from this site including mine and not have to learn much. It’s much better with vram.
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.