🟡:Flux Models 🟢:SD 3.5 Models 🔵:SD XL Models 🟣:SD 1.5 Models 🔴:Expired Models
🟡STOIQO NewReality is a cutting-edge model designed to generate photographic content. It empowers users to capture fine details in portraits, explore breathtaking landscapes, and bring mythical creatures to life with unparalleled variety and precision. This model is tailored for artists and creators who demand high-quality results, featuring dynamic light management and intricate textures, making it the ultimate tool for top-tier creative work.
Versions and Recommended Settings:
MAIN SAMPLER: dpmpp_2m+ sgm_uniform
CFG : 3+ | STEPS: 30+
(Recommended): CFG: 5 | STEPS: 40
Finally here is a first finetuning test on SD3.5 for NewReality. As always the first goal of the Alpha version is to give a focus to the photographic aspect and textures by expanding the range of subjects that it can reproduce. Future finetuning will focus more on individual components and details.
The Main Download (Full 26GB) is the AIO version with t5xxl_fp16, clip_l, clip_g and VAE already included.
In the file section of the sidebar you can instead find the Secondary Download (Pruned 16GB) of the version containing only the Model, so you can use it accompanied by the clips of your choice, depending on whether you prefer the fp8, the fp16 or want to experiment with new clip_l or t5.
MAIN SAMPLER: Euler+ Beta
GUIDANCE (In Forge:'Distilled CFG Scale'): 2+ | STEPS: 15+
(Recommended): GUIDANCE: 2.5 | STEPS: 25
NewReality FLUX.1 Dev has reached its Alpha version, but I'm still in an experimental phase with the Flux component. Working on the Unet architecture has proven to be more time-consuming and challenging than initially expected. As a result, I will release smaller, more frequent updates during this initial period to gather valuable feedback. I’m fully aware of the existing issues within the model, and your input will be instrumental in addressing and refining them.
The Main Download (Full 11GB) is the version containing only the Unet, so you can use it accompanied by the clips of your choice, depending on whether you prefer the fp8, the fp16 or want to experiment with new clip_l or t5.
In the file section of the sidebar you can instead find the Secondary Download (Pruned 20GB) of the AIO version with t5xxl_fp16, clip_l and VAE already included.
ATTENTION: The Full/Pruned nomenclature is just to be able to keep the Unet as the main download. All the others change with the AIO as the main. The two models are composed as mentioned above.
Refiner is Unnecessary and Clip and VAE are Included
MAIN SAMPLER: dpmpp_sde + Karras
CFG: 2+ | STEPS: 15+
(Recommended): CFG: 4 | STEPS: 30
NewReality XL PRO is a fine-tuned model designed with a specific focus on enhancing prompt adherence, earning it the name "PRO" (PRO-mpt). This version excels in managing lighting and image composition, offering a more refined and precise visual output based on the given instructions.
Refiner is Unnecessary and Clip and VAE are Included
MAIN SAMPLER: dpmpp_3m_sde + Exponential
CFG: 2+ | STEPS: 15+
(Recommended): CFG: 4 | STEPS: 20
Refiner is Unnecessary and Clip and VAE are Included
MAIN SAMPLER: dpmpp__sde + karras
CFG: 1+ | STEPS: 4+
(Recommended): CFG: 2 | STEPS: 4
Clip and VAE are Included
MAIN SAMPLER: dpmpp__sde + karras
CFG: 2+ | STEPS: 15+
(Recommended): CFG: 5 | STEPS: 25
Credits:
NewReality is the result of countless hours of work, involving several hundred iterations through training, merging of models and LoRA (Low-Rank Adaptation), fine-tuning specific blocks or captions, as well as performing additive and subtractive merges. Due to the complexity of this process, it is difficult to provide a precise, step-by-step account of every decision and experiment. However, each phase has played a significant role in shaping the final product, whether by direct influence or through valuable lessons learned along the way.
Given the intricacies and challenges involved, I want to extend my heartfelt gratitude to the creators of the models that have contributed to this journey, either directly or indirectly. Regardless of whether they were incorporated into the final iteration, their work provided inspiration, insight, and progress throughout the project.
To all the creators whose models supported this endeavor, I extend my deepest thanks. I encourage everyone to explore and support their work, as it deserves recognition for the incredible value it brings to the community.
EauDeNoire humblemikey socalguitarist ZyloO xlabs_ai aramintastudio Seeker70 Ai_Art_Vision PromptoAI alvdansen aip0pp SG_161222
WORK IN PROGRESS...
Description
FAQ
Comments (140)
I'm guessing the 20GB FP16 Flux file is just the diffusion model? No T5/CLIP or VAE? It's not a size one can quickly download to check.
yeah just tried it, it has the clip, T5, and the vae
@SenkiAI I see, it's not FP16 then as the diffusion model alone would be the whole 20GB
I downloaded and checked it out, no VAE needed! I'm using Forge, I have a rtx 3060 12gb, generation time is 4 minutes.
sorry everyone, i actually haven't inserted flux in the description yet and there's a bit of confusion, the model is completely stand alone, with t5, clip and vae already included and finetuned. The pre-alpa is in fp16 version (i don't think of making an fp8 for now). in the description also the parameters that i recommend for use. @SenkiAI @abomonov thanks for the replies. @mrneon as said it's already all in the checkpoint. I don't understand the reason for the weight decrease, I'm still investigating, but if you go and see my versions of 'Flux compact' the same thing happens, but I can assure you that the results are identical to the 24GB version being able to test them separately.
@ALIENHAZE >the model is completely stand alone, with t5, clip and vae already included and finetuned
If the finetune didn't touch the T5 weights please also share the unet/diffusion model by itself. Otherwise people will be downloading the same 5/10GB of weights over and over with each new finetune and each new version. For some users that's a few extra seconds of downloading, for others it is hours.
If the finetune did touch T5 please don't, just don't touch T5....
Thanks bro! In the Flux models I found a bug, if you write “President Idiot” in almost all cases you will see Trump. LOL!
I wonder, is that a bug, or is because Flux is realistic and depicts reality?
this is just a pre-alpha, a first experiment, I will fix it in the next versions, it's a serious flaw, all US presidents are idiots. 😅🤦♂️
@ALIENHAZE This bug is in the base model too, it's not your fault!
TDS is a sad disease.
"It's not a bug, It's a feature"
It's just that SD 1.5 is obsessed with Trump by design! No less than 9 tokens were dedicated to capturing his likeness: trum, trump, trumps, donaldtrump, realdonaldtrump, impeachtrump, trumprussia, trumptrain, nevertrump! All are single-token words!
Meanwhile, common words like penis, vagina, pussy, dick, cock, fuck, and blowjob didn't qualify to get a token of their own, and must be pieced together with multiple tokens. The priorities of those who designed the CLIP model seem a bit skewed to me, but what do I know, maybe Trump is vastly more important than human genitals to some people.
It's a great model, however where other model create an image in 30-60 seconds for me, here it takes a couple of minutes (I only have 8GB VRAM)
oooooh. I cant wait for the Afrodite Flux version!
I think for that, around mid-September. they really streamlined the model for hardcore content, even if they did it very well, so it will take longer to get to XL standards, but we are already on the right track, I think that already the next version of NewRealityF.1 will make great strides in that direction.
Any chance of a Flux NF4 version?
I think so, I have a couple of jobs in progress and little time, and the Shnell will surely arrive first, and I think the fp8. or I could as a light directly charge the NF4 and not the fp8. I'm still a bit confused! 😅😅
@ALIENHAZE That's good news - it needs a bit of loving attention from you! - although in a lot of aspects it seems to be better than fp8...
NF4 is a bit more friendly versus the straight F8 in terms of running on Forge
Is the F.1 Dev Version supposed to work with A1111? If so, any tips for usage? I am getting errors.
I don't think A1111 support Flux yet, have to use ComfyUI
or Forge
or Swarm
You'll need to use ComfyUI or Forge
@low_rank thanks. I tried Forge and did get it working but it's insanely slow. Like 5m per image somehow.
@Baaaaassssed Did you fix the slow loading times?
This Flex model produces really nice images but they are way too noisy. It's like there is faint grid layer over top the image.
yes, I'm already working on fixing it in future releases. A 2x upscale with 0.2 denoise without prompt is the most effective for now, regardless of background noise, I still recommend it on other generations too.
Hello sir,
Please make a model for low end devices at least 8-12gb graphic processor. it is really great for all those who don't have expansive gpu . Thank you sir.
at the moment I still prefer the output of the normal dev model.
I thought it wont work in my pinokio forge. But it does and it makes excellent dreamworks.
Please do not include the basic VAE and encoders in future models. These checkpoints are killing our SSDs. We all have the base VAE and encoders already downloaded. Let's save a couple of GBs for everyone.
+1
There is a lot of visual noise in the generated images
ye i noticed that too any ideas to fix that?
@magmirant0011 sure go a higher resolution - its only that models have their trained sweetspots and this one is no exception. usually you would upscale and there are workflows doing this - that being said i havent seen a good one this far. its common to upscale with a XL model but in my opinion this ruins fine details. i tried upscaling with flux itself but with rather mixed results. for the moment its something we have to live with. patience
@androsynth7610 when i try 1024x1024 usually there is less noise than default res. And that was kinda dumb question because i have Waifu2x with PhotoConservative Ersgan model and it is amazing asf with deleting noise and enhancing resolution on images.
@magmirant0011 @androsynth7610 @AneleMelchiorre yes, I'm already working on fixing it in future releases. A 2x upscale with 0.2 denoise without prompt is the most effective for now, regardless of background noise, I still recommend it on other generations too.
what's the FLUX training data like? Gender balanced or just women?
The goal of NewReality is to be as complete and versatile as possible in generating photographic content. This is just a first test, so it is still full of biases and shortcomings, but I will try to make it increasingly broader in possibilities.
I don't suppose anyone could tell me how to get this working could they?
I'm currently using Fooocus but not sure if it will work within that or not?
If not then does anyone know of a simple workflow for ComfyUI or something else that this will work with easily?
I'm on a Mac so anything that will work on that would be greatly appreciated. Many thanks.
it's been a long time since I've used Fooocus, but I don't remember having to do anything special to get NewReality working. It should be as simple as dropping the download into the right folder. So: Fooocus>Models>Checkpoints is where you need place this file. If you are not seeing certain settings that you think should be available, then you probably need to select the Advanced check box that is below the prompt window.
Fooocus supports SDXL exclusively. Flux is not supported and never will. I'm assuming you are talking about the Flux version because it was the most recent and hyped variant but for the XL versions of this checkpoint you can follow the comment of @Model911 and you should be all set.
@bratzmeister Aha there are XL versions? I must have completely missed that. Thank you so much!
hi, what is the optimum image size for this model?
you can safely use SDXL dimensions, approximately 1Megapixel images (base x height = 1,000,000)
I dont get it, whats different from flux dev base model?
it is slightly fine-tuned to generate more photorealistic content. it is a first testing experiment, the differences are minimal. The new more fine-tuned version will arrive soon.
@ALIENHAZE thats totally good to hear <3
Can you released Unet only version? I downloaded this but have deleted it because it was ~38 seconds/it due to overflow of VRAM. Loved SDXL version!
I second this request. UNET version please. As of now this takes twice as long as fp8 flux dev on my system.
GOD
TOP TIER
MIND-BLOWING
THE BEST CHECKPOINT AVAILABLE TO DATE
🫡
Nice details, but how do you get any other body type and tiny with no curves and tiny breasts? Would love some advice from a more experienced prompter :)
trigger more detail from model, recommand 2 ext
Auto prompt by LLM and LLM-Vision (Trigger more details out inside model)
sd-web-ui: https://github.com/xlinx/sd-webui-decadetw-auto-prompt-llm
ComfyUI: https://github.com/xlinx/ComfyUI-decadetw-auto-prompt-llm
Automatic1111 | Forge extension (LINE | Telegram | Discord)
sd-web-ui: https://github.com/xlinx/sd-webui-decadetw-auto-messaging-realtime
ComfyUI: https://github.com/xlinx/ComfyUI-decadetw-auto-messaging-realtime
I cannot get this (Flux version) to work on Forge UI. I got a 4090 with 16 GB RAM but no matter whether i increase GPU memory allocation in the app or enable\disable malloc or pin shared memory - It Just Does not Work!!
It always gives the CUDA out of memory error.
Please help, i don't want to delete a 20 GB model file since its useless and just taking up space.
Try converting it to NF4 in the checkpoint merger panel. That should reduce the model size by at least half.
@plk Unfortunately, I cannot find an option like NF4 or similar. Or I'm blind ;-) Sadly, this model does not work as delivered. Therefore, at least the documentation how to run this properly isn't simply given.
@plk Sorry but for me too. I struggled but it simply refuses to work.
Same problem for me. 4090 and 32mb of RAM. Using it with Forge UI and system always freezes because of memory issues.
@schorle1708435 I am sorry but I gave up. It does not want to work for me.
https://github.com/lllyasviel/stable-diffusion-webui-forge/discussions/1224 this looks like your answer. i'll be trying it today. it created an 11GB nf4 unet model, but now you have to manually select 'vae/text encoder' models.
Very good and fast, I imagine the full version will be even better.
I use it with
Stability Matrix + Stable diffusion WebUI Forge
Ryzen 5600 + 32RAM + nVIDIA 2060 Super 8GB RAM
ssd Fury Renegade SFYRS/1000G 1TB
Are you using the 20 GB model file?
@Light_Saber Yes
@Gabryel1960 Hey man, for the love of everything I cannot get this working at all. How did you do it? I am having Ryzen 3600 + 16 RAM + 4090 + Kingston Fury and using Forge UI.
The 20.3 GB file always gives errors. Any guide to follow here?
@Light_Saber Look, I have a Ryzen 5600x
and 32 GB of RAM
I see that you have 16 GB.
Try configuring this to see if it's a lack of memory.
Go to Tx2img
and then to the end of the list.
There is an option called..
Never OOM integrated
and check the option "Enabled for VAE UNet (always maximize offload)
This option uses the disk as RAM when the real memory runs out.
I hope this has been useful to you... I haven't done anything special to make it work
Maybe the hard drive where it is installed has little space and can't use part of the space as RAM.
I installed the Stable Diffusion WebUI Forge
with the StabilityMatrix.exe app
Which almost does the entire installation automatically.
You can download it from here.
thanks @Gabryel1960 for the help and information. how did you find my first experiment? do you have any feedback, positive or negative?
the new finetuned version will arrive in a few days, with only the unet model, so that it can be used on less performing machines, but every observation is always useful to be able to intervene and improve the model.
@Light_Saber did you manage to make it work in the end?
@ALIENHAZE No problem, it works fine, I have nothing negative to say. Congratulations. I will try that new version when it comes out. Thanks
@ALIENHAZE Thanks for getting back. Unfortunately no. I will be looking forward to your next version. Kindly provide clear instructions on how to use it, some of us including me are completely clueless.
thanks
As others had said, A UNeT model only is essential in Flux to avoid adding extra gbs to every checkpoint.
yes, thanks to all the feedback the new version will be released in a few days, with the Unet model also attached, so as to give you more freedom of choice.
How do you get anything bigger than tiny breasts? I even tried (((Huge Breasts)))::4 to no effect. Please help.
not making fun of you... but the other nsfw models i'm using i can only get large breasts. i can't prompt them smaller. ymmv :)
@tedbiv @Hieheihei the new version will arrive soon, we will see if the feedback will be positive or not on these generations. in the future also the Flux version of 'Afrodite' which being dedicated to NSFW hopefully will be able to do it well!
I cant wait for even more fine tuned towards realism version, i like the name New Reality, this is exactly what my ideas are for, the blank canvas that i have painted in with warmth and peace in new timeline ;)
nf4 version without the included clip and t5xxl? I swear to god if every model starts bundling clip-l and t5xxl downloads are gonna get hit very hard cause im not buying another ssd just for this shit
thanks for the feed, the new Dev.1 version will arrive in a few days with only the Unet model, and to follow also the various light versions such as Schnell, nf4 and quantized! This was just a first experiment, this is exactly what feedback is for!
@ALIENHAZE your pre-alpha flux model got reprinted at Tensor Art. Did you consent it?
thank you so much for the feedback! unfortunately I haven't implemented tensor.art lately. I've requested to get all the models I have back, but it will take a while. Thanks again for the info, feel free to use it in the meantime, thanks to whoever uploaded it, at least you have the possibility to use it there too, because waiting for me will make you old!😅🤣
@ALIENHAZE well, then I'll have to play with it and maybe mix it with other good models through tools, while waiting to try New Reality, and hopefully Rafaello, full Flux versions. And the good thing about the wait is that Flux will be a bit more mature then.
Where to install?
It depends on the interface you want to use, in general just download it and put it in the normal 'Checkoints' folder of the program, already has clips and vae included. I personally use mainly ComfyUI, but you can also use Forge or Fooocus which have updated the compatibility. The new version will be available in a few days with the Unet model only, as requested by many!
stable-diffusion-webui-forge-main\models\Stable-diffusion (in-case you are having problems with that still)
Hello @ALIENHAZE , i am training a lora for a realistic person on your checkpoint, i am wondering if you did trains loras on your own checkpoint (the latest SDXL PRO XL NEWREALITY) , if so it would mean a great help if you could tell me these exact training settings values to have the best success with my lora: Epoch Num Repeats Train Batch Size Steps Resolution LoRA Type Enable Bucket Shuffle Caption Keep Tokens Clip Skip Flip Augmentation Unet LR Text Encoder LR LR Scheduler LR Scheduler Cycles Min SNR Gamma Network Dim Network Alpha Noise Offset Optimizer Optimizer Args
HI @blooddragonhero, with XL I have always used the basic version for lora tuning, and I have experimented little. It is with Flux that I have increased my experimentation on lora. in general it depends a lot on the resources you use for generation. locally, on the cloud on Civitai? how big is the dataset you use? and be careful with the captions that I think are the most important thing.
@ALIENHAZE I've tried on civitAI, with a 20 pics dataset but as i ve found i will try a full fine tune with 70 images + 6000 reference images we will see how it goes :)
@ALIENHAZE What advice can you give me on keeping the same face when doing full body images, closeups can be trained well and portraits from all angles can be achieved but when doing full body shots the face deforms all the time too much. Have you been able to solve this someway? Appreciate any help
first, the model will be refined and that feature will also be improved. then you can try to do multi-step generations like: txt2img-->upscale_2x-->inpaint face (with lora, with ipadapter, or with controlnet)-->downscale_1x(original dimensions)-->reactor_faceswap(with reference face)-->upscale_2x
FP8 version when?
you have my axe for direct and precise description. the amount of illiterate creators that write anything except something related to the models they upload is too high
I'm happy to be able to wield it!! 🤣🥳 I think I would like to further improve it, lacking various information on the various versions and their capabilities, but I never find the time!
The new finetuned version and the only unet will arrive soon, so as to give you more freedom, then I will try to find the time for that too!
Is this a fp16 Version?
yes, as a first experiment I tried with the most complete version I could do! in a few days a new fine-tuned version will be released with only the Unet model so that you can then use it as you like and then the various lighter versions.
Do you have discord server?
Hi @anubhavk198 , unfortunately for now I don't have time to keep up with that too. in the future there will surely be a way to stay in touch and more content. for now you just have to enjoy my models and if you have something to ask just ask, sooner or later I'll try to answer everyone!
Could you provide a set of quantized UNet-only gguf-files for your flux model? This works very well and extremly memory efficient for other flux dev models. Thx :)
hello @cremus , there will surely be at least one quantized version (which one would you prefer?), an nf4 and the schnell, but I preferred to focus on experimenting on the Dev.1 model, of which a new finetuned version will arrive in a few days.
Unfortunately, alone and with the resources I have, time is my greatest enemy. and what's more, every day something new comes out to experiment with, it's enough to drive me crazy! ahahahah
@ALIENHAZE Please Add NF4 Dev
@ALIENHAZE I'm using Q8 models on comfyui on the last version, i think it's the best compromise for quality and speed. It takes 2 minutes generation after model load. I only have 6 GB VRAM.
compatible with Ruined Fooocus 1.55 ! thanks a lot !
is it possible to have any clue or info about what is this trained on?
Second that....
How do you make video with flux
Luma / Kling / Runway Gen3
@magmirant0011 I am just starting out but I want to make them from comfy AI is that possible?
@Valket look into CogVideoX
@Narz @Valket @magmirant0011 @cyanocitta there is no open source Flux based video model yet, but you can use flux generated images as a starting point for other tools. there are many of them for different things. check out MimicMotion, LivePortrait, CogVideoX, DynamiCraft, ToonCratft, AnimateDiff...
i have RTX 2070 but why in 512x512 takes 12 minutes for image?
run it on colab
Because the model does not fit in to your VRAM!
you could try low vram tag on the arguments
Found an issue on flux(not tested for this model) add to queue twice wait for sampler then cancel the first one. The second one will render in less than 1 minute.
@xXLukyDevilXx How many GB of Vram does the card have? And how much ram does the PC have? What interface do you use? Because not having enough vram the model is loaded partially on the ram, acting as a bottleneck since the ram is much slower. In comyUI you can try to insert the string --lowvram in the startup .bat file. Or try to use the GGUF versions of which a NewReality one will soon arrive.
I love your SDXL checkpoints!
Can you update the description of your flux checkpoint with what you've done to flux? Not sure why I would use this instead of Flux.d/s/f
Cool..! but I want SWF Version 😅 better that for safe...
thanks for your work. Will you make a gguf version ?
Yes, as soon as I finish the experimental phase with the Dev 'Alpha' version and I have clearer ideas, I will move on to tuning the other versions.
Any news for update?
late, but a new update is out! let me know what you think, every feedback is very useful to improve the model! Even if I can't always answer everyone, I read everything!!
Not bad, but a lot of dots or some noise in picture.
generating in 675/1200 (9/16 portrait))
It's about Flux model
The training is likely broken. There was some bugs in some of the earlier trainers that did not scale/shift the latent correctly during training which led to this typical low resolution/dot matrix/blue noise pattern. Not sure if this is the same bug or something else that was broken during their training run
@civit0423 very sad, cuz it's really one of best Flux models today as for detailing amd overall picture.
Looking for next version, and using now dev-bnb_v2 and Unachained models.
@aeronliru thanks a lot for the feedback, even if I can't always answer I read everyone! A new update has come out, I tried to fix this problem with a slight improvement even if at high zoom it is still present. If you try it let me have a new feedback, they are very important to fix the problems.
thanks @civit0423 , i'll look into that too, it could give useful info. but since even the basic version of flux has this thing, i think it's mainly Flux's way of handling latent. Let's see what comes out of it with the various tunings and wait and see.
@ALIENHAZE Ofc I will test it. will post here new report about model!
What is "FLUX GUIDANCE: 2+"? in the setting, im using Forgeui
There's a new Forge version that supports Flux. You might be able to just update your forge? I'm too scared to do that with how often SD breaks so I decided to just have a new install. I just followed this tutorial.
in comfy it's called guidance, but in forge its called distilled cfg scale
@rjox Thank you so much for actually just saying what it's called in Forge lol <3
thanks a lot @rjox for the info, I will update the description for those who use Forge
Is anyone else with 16gb of vram taking 1 minute to generate each image? Do need to tweak any settings or is this normal. I'm new to Flux, ty
Seems normal.. Using a 4060Ti 16GB here.
I also have a 4060ti 16GB card and generate at about 2.5s/it. I use the Hyper Lora so an 8 step generation only takes about 20 seconds.
Details
Files
stoiqoNewrealityFLUXSD35_F1DPrealpha.safetensors
Mirrors
STOIQONewrealityFLUXSD_F1DPreAlpha.safetensors
stoiqoNewrealityFLUXSD35_F1DPrealpha.safetensors
stoiqoNewrealityFLUXSD35_F1DPrealpha.safetensors
stoiqoNewrealityFLUXSD35_F1DPrealpha.safetensors
stoiqoNewrealityFLUXSD35_F1DPrealpha.safetensors
STOIQONewrealityFLUXSD_F1DPreAlpha.safetensors
prime.safetensors
STOIQONewrealityFLUXSD_F1DPreAlpha.safetensors