🟡: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 (97)
WOW! Thanks!
I'm working on something, and now you upload this, i guess we could be friends 🤣.
I'm exploring how high definition can be in Flux dev, and I got awesome results. That will be used as a LoRa dataset, so on tests I guess your merge will outperform standard Flux dev. Excited to see what happens.
But where is the AiO version? is that the pre alpha one ?
there is a box below the details called "2 files" its in there.
got issue with last alpha version:
assert isinstance(state_dict, dict) and len(state_dict) > 16, 'You do not have VAE state dict!'
AssertionError: You do not have VAE state dict!
while pre-alpha working for me. wonder why.
I'm GTX1660Ti 6Gb GDDR6
Clips, t5 and VAE are not included in this model.
@Haircut66 I have t5 and clip - it's still same error. what vae to use with it???
@aeronliru I have "ae.safetensors" https://huggingface.co/black-forest-labs/FLUX.1-dev/tree/main the file in this huggingface, also if you go to the "2 Files", and click there will be an AIO file (All-In-One, the larger download of the 2) that will have a model with all the files included, just like the Pre-Alpha version.
if you can't download from that Huggingface, try this one, https://huggingface.co/black-forest-labs/FLUX.1-schnell/tree/main
@Haircut66 downloaded your ae.safetensor and it worked now!
Thank You very much for your work!
thanks for help @Haircut66 .
@aeronliru in the files sidebar you can find also the AIO model with t5 and vae included like the previous pre-alpha version.
@ALIENHAZE aha, I see now. Thank you!
i was able to create pov of looking up or down in XL but its not working in flux, any suggestion?
Flux limitation
@Askeladd_ NAaah, other model did it easily
The "FP16" file is in reality FP8 !
Even the "BF16" I still see "torch.float8_e4m3fn" !
Even though it's double the size!
something is really wrong here, please double check your training workflow
I also opened an issue on ComfyUI-GGUF extension github repo, hopefully they have more insight on this
it's not a mistake, i still have to figure out how to upload them to civitai, then i'll add a more accurate description of the models, and how they work. i can't upload two identical extensions for the same model, and i should therefore create a separate page for the AIO version and i don't like this, since it's nothing more than the 'unet' with t5, clip and vae included of my choice. another thing that can be misleading is the fact that when you read fp8 or fp16 in the other models it doesn't mean the weight of the model but the type of t5 used. since the unet doesn't have any clip you can decide whether to use it alongside the t5 fp16 or the t5 fp8 or even test it with other types of clips. Thanks anyway for the feedback, it will be useful to make the model more understandable and accessible.
Hello,
Thanks for the model!. But I do have one issue I would like to bring to your attention please.
Whenever I try to prompt two people in the same image, each being of a different race it always creates both of the same race. Example, If I prompt white male and chinese male, it will output both asian looking. Similarly, If I prompt white female black female, it will output both as african looking. And finally, the age prompt seems to mismatch, If I prompt for example a 24 year old person it will output a person looking much younger or noticeably older.
Otherwise a fantastic flux model : )
thanks a lot for the feedback, can you also share some examples or some prompts used? I will try to investigate for future versions.
@ALIENHAZE Hello! You're most welcome. Yes I can share some example promptst, they are simple and straight forward:-
- standing in backyard, a gorgeous american while girl aged 20 years old wearing a summer dress
- standing in backyard, a handsome american while boy aged 14 years old wearing a suit
In the above example, it will either display a human female/male whos face often(not always) appears +/- 5 to 6 years either older or younger. Another example,
- standing in backyard, a gorgeous american while girl aged 20 years old wearing a summer dress with indian man aged 20 years old
In the above example, it will often(again not always) make both subjects either white or indian race instead of keeping them individual as prompted. And finally,
- standing in backyard, a gorgeous american while girl aged 20 years old wearing a summer dress with japanese man aged 20 years old
When it comes to asian races, the moment one human is prompted to be japanese/chinese etc. it will 100% make all humans of asian race.
I am also sharing my settings used incase it is helpful to you:
Euler + Beta
Steps - 23
Distilled CFG Scale - 3.5
CFG Scale - 1
Resolution - 1024x1024
I hope that helps. Thanks for your efforts and I look forward to updates : )
How should I apply LORA...?
Can I get a workflow?
I'm getting some really nice results. Especially considering it's still in alpha stage. Great job!
thnx, Best Flux Model!
Thank you for the compliment. I'm happy that it is so appreciated even if it is still in an experimental phase. what are the things that have struck you most compared to other models and what are the critical issues encountered?
Hello, I downloaded Pruned Model fp16 (11.08 GB) and I get an error saying "GatedRepoError: 401 Client Error. (Request ID: Root=1-66f24574-4e60390f021d75a74c9f311f;8c197de6-8ea0-4fe5-8964-fde112548ede) Cannot access gated repo for url https://huggingface.co/black-forest-labs/FLUX.1-dev/resolve/0ef5fff789c832c5c7f4e127f94c8b54bbcced44/model_index.json. Access to model black-forest-labs/FLUX.1-dev is restricted. You must have access to it and be authenticated to access it. Please log in."
is this as simple as making an account on huggingface?
That issue has wide range solutions. Usually if you are using lora, that cause it. Wrong Vae might cause that issue as well. Graphics driver, gradio update or downgrade, some api issue etc.
@Reelai Thank you. I'll see
Screenshot please... Which version of Flux?
@datedman353 Thanks but I solved it. I updated forge after a week. Now it works!
thanks!! may need GGUF versions, maybe Q4/Q5 for low vrams
They will also arrive as soon as I finish experimenting with the Dev version, just a little more patience 🤯
@ALIENHAZE just curious in my install dev models are having much more difficulty running out of vram when the shnell model doesnt (im particularly using it with the cog video sampler which i feel like shouldnt matter but it is so idk. are there any model differences between shnell and dev??
This fp16 model is the only model that actually produces on par or better quality than the stock flux model on a 4090, all the other flux models appear to focus more flexibility in prompts like nsfw or merging works but at the cost of quality. this seems to allow both prompt flex and quality. that being said i wish this quality could be replicated in a gguf model for faster loading and ability to run on normal graphics cards.
May be a silly question but why is the 'pruned' version larger than the 'full' version? Isn't that backwards?
In the description:
"The main Download (11GB file) is for the Unet only. In the sidebar, in the 'file' section you can also find the Download of the AIO version (for now is under '.pruned model' _ 20GB file) with t5, clip and VAE included."
great model - please make a schnell version(!)
i get how dev is so attractive and pretty, but the licensing for schnell is optimal for a lot of people. schnell been sitting there for a month+ now waiting for it's duly deserved attention from the community✨😬
As I said - wrote new report about last alpha version.
Works veeery fine. just a little noise on skin, but I think it's about model learning issues...
Increase the guidance.
Thank you very much @aeronliru. I'll take a look and give you some new feedback too. Little by little I hope to get back to the textures achieved by XL but with the generation quality of Flux. I'm experimenting with different methods, let's see which one works best. We'll talk in response to the review! 😀
You can also try as @anonlinuxmint962 said to use a guidance closer to 3.5 than 2 if you haven't done so already.
@ALIENHAZE I'm using 3.5 usually with FLUX, but last time using 2.0 - it gives very interesting results, that are veeery nice for me.
@anonlinuxmint962 all time have 3.5, I'm using FORGE Ai engine.
Best checkpoint for Flux so far, thank you <3
what is the difference between the 20gb and 11gb one?
@ducky66 FP16&FP8?
which one should I use if I have a 3080 with 12gb ram?
so impressed
F1.D ALPHA - this model is insane
Maybe that's just me, but in comparison to commonly known Flux models these days - this one provides the best quality and accuracy. Works like a charm in ForgeUI. Great work!
Absolutely. I've tried a few and this one seems to outdo them all. Great work by the author. Their other models are also top notch. I'm still using their SD 1.5 model with great results. One of my favs, up there with the Photon model.
do you need to reate a unet folder, or put it in the stable diffusion folder in models?
@puma0000 I just put model file right into "Models\StableDiffusion".
@puma0000 what interface do you use? if you use ForgeUI better if you use the AIO version and insert it as said by @kapec512 in "Models\StableDiffusion". If instead you use ComfyUI you can use the AIO version in "models/checkpoint" as for SDXL, or use the Unet version and load separately T5, clips and VAE and insert the model in "models/unet"
@kapec512 @mmdd2543 thank you very much for the positive feedback. I'm glad my work is useful and appreciated! outperforming Flux is not easy, there are still many things to fine-tune, but we are on the right track.
If you have anything to point out, positive or negative, it is very welcome to be able to intervene in improving or maintaining it. with the vastness of possibilities offered by the model it is difficult to test them all alone 😅😉
Hello!
Is it possible to run the model (F.1 Dev) on SwarmUI? I placed the All in One model in the Unet section. When trying to load the model I get the errors:
“backend #0 failed to load model with error: ComfyUI execution error: The text encoders (CLIP) failed to load”.
“Backend request #3 failed: All available backends failed to load the model”
Unfortunately I don't know, I don't use SwarmUI. I'll try to find out or try it directly and if so I'll let you know.
@ALIENHAZE Thanks for the reply! I think I found a solution - I put it in the “Stable-Diffusion” folder. It's strange, because Flux works from the Unet folder, and the NewReality model is a fork based on Flux (if I understood correctly).
Nevertheless, the model works perfectly! Thanks for your hard work!
@kot1vertolet Wait, what is the fix? I didn't catch it.
@kot1vertolet ah ok, so 'Swarm' like 'Forge' prefers the use of the Checkpoint version over the Unet. Be careful that both versions of all models exist, simply StableDiffusion was light enough to not be necessary to split clip and model. In Flux you can instead find only the model (Unet) or the checkpoint version (AIO). It's a bit like what happened with the VAE for XL that some models already had it included, for others instead it had to be added separately.
question? are you using the AIO version right? the 20GB one?
I absolutely love this model. It handles everything I throw at it with ease.
I don't know what is going on, this and the realvis beta model appear worse than the basic FLUX models when trying the same prompts/setups in them and experimenting with samplers etc, the OG model is just more realistic, am I missing something? I'm not seeing anything about the prompting being different for these? the XL fine tunes are amazing but these Flux fine tunes just aren't working for well for me, any ideas, am I missing something?
What are your settings when running prompts?
RealFlux? can't speak for the dev model, but the schnell one really likes Shift lower than 1. for me, it affects both form (slightly less overfit) and skin details (a lot better). conversely, ArtFusion (also a schnell model) is the opposite; it likes using that 'formula' based on image dimensions like dev does, so shift goes up as target size does.
what's my point? this stuff is wild and crazy and model creators haven't found a universal approach to it - and may end up with workflows that totally betray what you know from using other models. u just gotta try experimenting, break some common sense 'rules' along the way. it's frustrating but also makes discoveries more.. dopamine-infused lol
Hey! I can't even imagine how much work and knowledge it takes to make flux better. It works like a charm!
I am wondering if there is any point of using additional reality enhancing lora like those iphone photography or others since there is a lora merged into this checkpoint.
It depends a lot on how well the lora is trained and how well this model is able to generate the desired result. If a lora is well trained it will greatly affect the final result and outperform this model, especially if you are looking for a specific photographic detail. On the other hand, NewReality is much more flexible and versatile, allowing you to generate different styles and high-quality subjects without having to access hundreds of different resources.
@ALIENHAZE actually i am using Forge and it Has this Realistic Sampler built in, with it and no other loras the output is very realistic already and i cant really find a difference on comparison when using that realism loras. You did nice job and i cant wait for more updates.
Where is SD 1.5? Can only see 1
cant get the flux model to work it shows the image towards end and then goes black
That usually means you ran out of VRAM. If you're on ComfyUI you can enable Split mode in the settings if you have more than 32GB of system RAM (you can get away with 32GB if you have a fast NVMe SSD as your pagefile drive).
Automatic1111 / Flux : I had this problem with SD1.5 because I selected a wrong or corrupt VAE file. Maybe it's a similar problem since the actual rendering process has taken place according to your description (?). Is there no error message in the console?
what are the settings in forge? I always end up with a gray image (during generation I can see some image being generated)
I have a 4090
I don't use forge so I've never had this problem. I'll try to help you intuitively.
First thing check the CFG. The 'CFG SCALE' on my models keep it onlu at 1 and only use the 'DISTILLED CFG SCALE' at 2+.
second, if you are using the AIO version check that the 'VAE/TextEncoder' entry does not have Clips and VAE loaded which could conflict. If on the contrary you are using the Unet version check that you have loaded the Clips and VAE correctly.
Will and which version work with RTX 3060 12GB? Also, could you guys recommend best settings for Forge WebUI?
from my experience, DDPM, DDM, or just DDP, there's different versions and variations of each, but no matter the choice (with the same settings otherwise) all of "Diffusion Denoising ....xyz" previously referred to in their abbreviated formats, will always give me blurry blobs as images in forge. I was trying avoid using Euler all the time because of its simplicity but surprisingly its one of the best (generally (though there are specific scenarios which are anomalous)) and is definitely the fastest of any, so i tend to stick to it, or Euler A. Scheduler is just preference of desired output, i'm always working with some desire of complexity and abnormality so i usually use karras, exponential, or if not needed or wanting anything special, linear or simple is just fine.
This is a beautiful checkpoint. Thank you. Is there any chance you could help with my VRAM shortage by releasing this in a Q_6K GGUF format? That would save us at least another 4-5GB and still produce great results.
as long as i'm in testing, i'll continue to work on the dev version. my goal is to outperform flux, and i prefer to do it to the full potential. quantize versions will come in the future.
This is my favorite checkpoint. About to come back to it, I love it!
It won't work in A1111. I keep getting this crap:
----- begin copy/paste -----
Loading weights [f488e9fa0a] from C:\AI\A1111\webui\models\Stable-diffusion\STOIQONewrealityFLUXSD_F1DAlpha.safetensors
changing setting sd_model_checkpoint to STOIQONewrealityFLUXSD_F1DAlpha.safetensors [f488e9fa0a]: AttributeError
Traceback (most recent call last):
File "C:\AI\A1111\webui\modules\options.py", line 165, in set
option.onchange()
File "C:\AI\A1111\webui\modules\call_queue.py", line 14, in f
res = func(*args, **kwargs)
File "C:\AI\A1111\webui\modules\initialize_util.py", line 181, in <lambda>
shared.opts.onchange("sd_model_checkpoint", wrap_queued_call(lambda: sd_models.reload_model_weights()), call=False)
File "C:\AI\A1111\webui\modules\sd_models.py", line 967, in reload_model_weights
state_dict = get_checkpoint_state_dict(checkpoint_info, timer)
File "C:\AI\A1111\webui\modules\sd_models.py", line 344, in get_checkpoint_state_dict
res = read_state_dict(checkpoint_info.filename)
File "C:\AI\A1111\webui\modules\sd_models.py", line 318, in read_state_dict
pl_sd = safetensors.torch.load_file(checkpoint_file, device=device)
File "C:\AI\A1111\system\python\lib\site-packages\safetensors\torch.py", line 310, in load_file
result[k] = f.get_tensor(k)
AttributeError: module 'torch' has no attribute 'float8_e4m3fn'
----- end copy/paste -----
In Fooocus it just won't do anything, sad, it looks like a great model based on demo pics here.
Flux models do not currently work in A1111, you need to use forge for that.
as @hestiacummings said automatic1111 doesn't support Flux yet. Forge (https://github.com/lllyasviel/stable-diffusion-webui-forge?tab=readme-ov-file) is the most similar interface to automatic, or you can use ComfyUI (https://github.com/comfyanonymous/ComfyUI), scary at first glance, but more versatile and memory friendly! Even Fooocus hasn't updated the compatibility yet!
@ALIENHAZE In my experience, Forge is very memory friendly too.
@Akalabeth yes of course, even Forge is very light, I said it quickly, I implied a detail. in comfyui you have the possibility to load into memory only the nodes necessary for your generation which in simple operations does not differ much from other interfaces, but in case we start talking about ipadapter, controlnet, video gen and much more complex things, both for a question of memory efficiency and for a question of concatenations of customized actions, a much wider range of possibilities begins to open up.
@ALIENHAZE Can't disagree with that!
@ALIENHAZE Can't disagree with that!
So yeah, it does work with ComfyUI, and I'm amazed by the image quality! Cheers! :)
use comfyui，one day work only
@StargateMax great that you managed to use it, now prepare yourself for a degree course to discover the whole world behind ComfyUI! Super happy that you like the model too! whatever you ask, I read everything and sooner or later I will also answer, and as you have seen there are a lot of people ready to help. See you next time!
This is now my main checkpoint for realistic images. Thanks :)
I've had quite a bit of fun throwing everything I can at it.
I'm happy to hear these beautiful words. Of all that you were able to test, what did you like the most and what would you improve? Are there things that you were unable to generate?
@ALIENHAZE I've been using the full model (I was going to downloaded the pruned one but I had all the encoders + vae anyway)
I've been comparing to Flux dev and I can say that at 3.5 distilled + 25 sampling in forge (just my preference for my style of prompting + loras), this has better composition and more realistic saturation. I feel like my loras do not affect the checkpoint as much as in dev, which in my opinion is good as it allows me to better finetune. The shadowy tones are also not as harsh on higher sampling steps.
It does pull back a bit on certain aspects, in order to keep with a more realistic output (which is not necessarily a bad thing). In my video here: https://civitai.com/images/34471423, the same prompts on flux dev had more of the exagurated skull features more often whereas this checkpoint had more painted/high fashion look, as a quick example. I generated about 100 images for each checkpoint (outside prompt matrix just to get a feel). In the end, I more preferred this over flux dev, regardless.
When forcing perspective on things (ie: such as giant animals being walked by humans), without the use of extra loras (or a specific lora I created for said purpose), I found that deformities begin to occur much easier. However, flux dev was much worse in my tests anyway so this checkpoint does better. That and it is such low priority that I don't think it's really an issue much, as loras exist for that on flux.
Overall, I am just fairly happy with this one. While I downloaded afrodite as well, I don't generally generate those types of images/videos so I can't really review or comment on that one.
Great CP! This is best I used so far. Thanks for making this available to try!
Works great w/4080:16g and Forge.
For now I use the XL PRO version, it's become my favorite, it can create amazing fit girls with perfect body and face!
Wow, I love it! May I ask a stupid question (I'm new to AI)? Why is the pruned model double the size of the normal model?
The 'normal' Flux.1 Dev model is more like ~24GB. We get these great re-releases that are less than half the size.
Generally speaking there's three generations, SD1.5 models are ~4GB, SDXL can be up to 8 or 10GB, and FLUX is 10GB++, but it's not exactly linear, and there are lots of exceptions and interleaving, with multiple 'foundational' models being released by different businesses/organizations/institutions.
No such thing as stupid questions.
https://huggingface.co/black-forest-labs/FLUX.1-dev/tree/main
This article explains the difference between full and pruned models quite well:
https://techtactician.com/pruned-vs-unpruned-models-stable-diffusion/
TL;DR: Pros of the pruned models are smaller size, faster image generation, cons are lower accuracy, images may be (but not in all cases) worse in quality than in the full model.
Details
Files
stoiqoNewrealityFLUXSD35_F1DAlpha.safetensors
Mirrors
STOIQONewrealityFLUXSD_F1DAlpha.safetensors
STOIQONewrealityFLUXSD_F1DAlpha.safetensors
stoiqoNewrealityFLUXSD35_f1DAlpha.safetensors
stoiqoNewrealityFLUXSD35_F1DAlpha.safetensors
stoiqoNewrealityFLUXSD35_F1DAlpha.safetensors
stoiqoNewrealityFLUXSD35_F1DAlpha.safetensors
STOIQONewrealityFLUXSD_F1DAlpha.safetensors
STOIQONewrealityFLUXSD_F1DAlpha.safetensors
STNRFLUXSD_F1DAlpha.safetensors
STOIQONewrealityFLUXSD_F1DAlpha.safetensors
STOIQONewrealityFLUXSD_F1DAlpha.safetensors
flux1.safetensors