Join The Tinkerer on Whop to get early releases, private pages, and the Tinkerer Discord role - all in one place. 👉 Join on Whop
After a ton of requests, I’m finally rolling out CyberRealistic Flux (FLUX.1 dev)! It’s designed to make realistic images, both safe-for-work and not-so-safe-for-work. It’s not perfect yet, but it’s a solid start and sets things up for what’s coming next.
Heads up: The (non AIO release) doesn’t include clip models or VAE. You’ll need to grab those separately here.
(clip_lorg.safetensors = orginal clip_l file. clip_l.safetensors is the CLIP-SAE-ViT-L-14 one)
Settings I Use:
Sampling method: DPM++ 2M / Euler
Sampling steps: 20/30 Steps
Distilled CFG Scale: 2–3.5 (Samples were done at 3.5, but lower values often give better results)⚠️ A Quick Warning:
This model can make mature or sensitive stuff. You’re in charge of what you create, so use it wisely (or at least get creative with it).
đź’¬ Want Better Prompts?
There’s a custom ChatGPT that’ll whip up great prompts for CyberRealistic Flux:
đź”— [Try it now on ChatGPT]
Description
CyberRealistic Flux V2.1 AIO (All-in-One) is your go-to if you want everything set up and ready from the start. You get the VAE and both CLIP versions included, so no need to track down any extra files. Just load it up and start creating.
FAQ
Comments (27)
Very very very good work, amazing!!!!
Does anyone have a workflow for using this model in comfyui?
drag and drop this in comfyui GUI
I have also a Workflow on Patreon: https://www.patreon.com/posts/cyberrealistic-138349516
Friends, I need help. If I write the word breasts in the hint (it doesn't matter, big breasts, small breasts), then nipples are necessarily drawn or the breasts are completely naked. How to deal with this? Or if I specify cleavage in the hint, then again the breasts will be naked. How to avoid excessive nudity?
(((Candid))), breast,(((wearing dress))) or other. no boobs,simple cleavage.
@EMYSTARIELÂ Flux does not understand parentheses or weights on individual words or expressions.
@joeboeye In forge, expressions priority is effective even with flux. I've done a bunch of testing on this and I confirm.
@EMYSTARIELÂ Priority, as in position in the prompt, sure. As far as parentheses go, Flux itself has no use for them. It understand natural, descriptive language.
@joeboeyeÂ
In image generation with Stable Diffusion:
Double or triple parentheses are an inherited convention for weighting or prioritizing certain words/phrases in the prompt.
For example:
((blue eyes)) = makes "blue eyes" a little more important.
(((blue eyes))) = prioritized even more.
Actually, this is a "trick" of the SD prompt parser, not a universal rule of the language.
Why doesn't Flux recognize this?
Flux uses a different tokenizer and generation engine.
It doesn't "translate" multiple parentheses into weights like the SD parser does.
As a result, (((blue eyes))) is read literally as additional text, which has no effect.
This model favors a clean and descriptive natural language, possibly enriched with explicit weights if the UI/Forge supports them (example: "blue eyes:1.4").
but with forge It works for me with flux! Incredible.
@EMYSTARIELÂ If your results are truly due to this syntax in Forge, then that means Forge rewrites your prompt to turn them into something weird to try to mimic the effect weights have on SD and SDXL, by altering the order and introducing duplication.
That could be annoying, if that's the default, for someone coming from another UI.
That would explain why there are so many examples here that are totally useless in ComfyUI. But lots of people seem to think they are getting results with comfy with this. They are not. Just luck.
@joeboeye hopefully....
If you drop a CyberRealistic for Qwen you'll become my first buzz purchase. :D
Looking at the picture feed from most models, this one included, it's FAR too obvious that an insane amount of learning has been carried out defining Ana De Armas as the one face to rule them all.
I can't seem to find any models that don't just inherently turn all "good looking woman" into Ana De Armas's distant relation... sighs...
Punk girls model gave me some non-standard attractive girls, but I agree with your statement that most of these flux models pick one or two super attractive female face styles and just run with them.
it's the chin. there are loras to fix it. https://civitai.com/models/775002/chin-fixer-2000
@pursuit_of_beauty could you link?Â
any thoughts on adding a little Krea? feeling like it's that missing piece Flux has always seemed to lack. immediate skin improvement, like, no contest
I am using the model in Python and Kaggle Notebook. I am facing problems with the models:
Pruned Model fp16
Exception has occurred: TypeError
unsupported operand type(s) for -: 'set' and 'list'
File "E:\python practice\Deep Learning\Loding CivitAI model to pipeline.py", line 15, in <module> pipeline=StableDiffusionPipeline.from_single_file( TypeError: unsupported operand type(s) for -: 'set' and 'list'
@EMYSTARIELÂ How much CUDA memory is required? Kaggle offers 15-16GB GPU free to use
@kskushagrasaxena233Â Kaggle notebooks have limited resources (GPU, VRAM, CPU time). For a model with 12 billion parameters, this can quickly become cumbersome. If the model requires a lot of GPU memory or computational power, it may not be usable in a free Kaggle notebook or may be very slow.
The base model (flux.1-dev, ~12B parameters) is very heavy.
In FP16, it requires 24–30 GB of VRAM for comfortable inference.In FP32, it's completely impossible on Kaggle (it can go up to 50–60 GB). Kaggle T4 or L4 (15–16 GB VRAM), too small to load the entire model at once.
Possible solutions with 15–16 GB
Use flux.1-schnell (or lighter versions), This is a variant optimized for fast inference, which runs on ~12–16 GB of VRAM. → So yes, it works on Kaggle on a T4/L4 GPU.
Please, one GGUF Q8 version for 8gb VRAM cards. Thanks in advance.
I will upload this version tomorrow (hopefully)
Done!
@Cyberdelia Perhaps a dumb question, but why Q4_1 or Q5_1 and not Q4_K_M or Q5_K_L ?
@Brewce Not a dumb question at all! Q4_K_M and Q5_K_L are newer grouped quant formats - they usually offer better quality at the same or smaller size compared to Q4_1 and Q5_1. I used the classic formats for broader compatibility.
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.