⚠️New updates to Flux Reality will be posted to Patreon ⚠️
V11 - What's new:
This is a softcore only version with no hardcore concepts trained into it.
Boobs got a big upgrade, lots of new shapes and sizes added to the dataset, this improved the nipples compared to the previous softcore version. I will expand it even further.
Pretty major stability improvements with this version
Some of the example images were made using my amateur photography LoRA trained on v11: https://www.patreon.com/posts/amateur-style-133959021
Concepts:
Big ass - Tag: big ass
G-strings - Tag: g-string
POV Cowgirl - Tags: vaginal sex, cowgirl, breasts, man, pov, spread legs, male pubic haie, penis, girl on top, straddling, from below, solo focus
*Example prompt: A 21 year old asian woman with long blonde pigtails and medium breasts looking at viewer with an open mouth nude having vaginal sex with a man in the cowgirl position with spread legs squatting from POV from below with male pubic hair penis, pussy. girl on top, she is adorned with a black choker and she is straddling outdoors. female pubic hair, realistic, cowgirl position, solo focus, upper teeth only, she is riding on top of him, translucent white cum dripping out of her pussy and down his penis, crotch focus, full body shot
POV Reverse Cowgirl - Tags: vaginal sex, reverse cowgirl, ass, pussy, anus, male pov, man, penis, male pubic hair, from behind
*Optional tag: squatting
*Example prompt: Woman, 21, with long hair, large breasts, and black hair, adorned with jewelry and nail polish on her fingernails, looks back at the viewer with an open mouth while engaging in vaginal sex from behind indoors on a bed. Her ass is prominently displayed, as are her pussy and anus, while the mans penis male pubic hair and testicles are evident. Elements like a bracelet, lipstick, and a mole accent her look, with the scene set by a backlit window with blinds, rendered in a realistic reverse cowgirl position from a male pov.
POV Boobjob - Tags: male pov, paizuri, breasts squeezed together, nipples, penis, solo focus
*Example prompt: Caucasian woman, age 24, medium breasts, looking at viewer, long wavy platinum blonde hair, large breasts, male POV, brown eyes, nipples, penis, solo focus, nail polish with matte sapphire finish, long blue nails, eyelashes, paizuri, freckles, breasts squeezed together, realistic, majestic mountain range, cum on her breasts and dripping down his penis
Anime Style - Tags: Anime woman, anime style
*Not always the cleanest output but it can be fun to play around with. Sometimes it makes anime/realism hybrids, other times the anime style is more dominant. I didn't train this into the model this was already trained into Flux Dev but it wasn't very usable until this version.
Blowjob / Deepthroat from side - Tags: nude woman performing fellatio to a man view from side, male pubic hair, testicles, solo focus, deepthroat, penis, tongue out, fellatio, close-up
* Optional tag: looking up
* Example prompt: A nude woman with eyes closed while performing fellatio to a man view from side, she has long, lustrous copper hair flowing in an elaborate style with twisted, ornate braids, wearing jewelry with earrings and a choker in a realistic outdoors raw photo 8k, male pubic hair, testicles, solo focus, deepthroat, penis, tongue out, fellatio, close-up
Facial - Tags: open mouth, tongue, facial, penis, cum, solo focus, cum on tongue,
* Example prompt: A woman with long hair, looking directly at the viewer with brown eyes and adorned with jewelry including a necklace and earrings, shows an open mouth and tongue as she is engaged in a realistic facial moment; her brown hair frames her face while a nude man's penis provides cum, cum in mouth, and emphasizes his solo focus with a cum on tongue detail, the mans penis is ejaculating cum on her face
Ahegao face - You can start the prompt with "a woman making ahegao face with her tongue sticking out and her eyes crossed" or "A woman making an exaggerated ahegao expression—with her tongue playfully sticking out and her eyes crossed." Or use some variation of this in your prompt.
Innie pussy - No tags needed, just use "nude" or "pussy".
Missionary position - Tags: missionary, vaginal sex, male POV, spread legs, lying, pussy, penis, hetero, solo focus
*Example prompt: Woman lying on her back on a bed in a missionary pose having vaginal sex, bathed in light shining through a window, captured as a raw photo 8k. From a male POV, she looks directly at the viewer with an open mouth and eye liner. Her hair is auburn and she has large breasts with visible nipples. Completely nude except for emerald striped thighhigh socks, she presents a sensual scene with spread legs that reveal explicit details including her pussy. The realistic composition also features a penis, emphasizing the hetero encounter with a solo focus on every intimate detail
*You can make the penis bigger by prompting "big veiny penis", you can also add further description to the man such as "male pubic hair", etc.Doggystyle position - Tags: doggystyle, vaginal sex, male pov, pussy, anus, penis, male pubic hair, solo focus, hetero
*Example prompt: Woman on all fours on a bed doggystyle position bent over top down bottom up having vaginal sex with a man, bathed in light shining through a window, captured as a raw photo 8k. From a male POV, she is facing away with head turned to to the side with an open mouth and eyeliner. Her hair is sandy blonde, long and wavy, and she is completely nude except for rainbow striped thighhigh socks. She is skinny with a big ass. Pussy, anus. The realistic composition also features a big veiny penis and male pubic hair, emphasizing the hetero encounter with a solo focus on every intimate detail, view from behind and above.
Cum - Tag: "cum". You can specify where and how much of it you want. For example: "excessive amount of cum dripping off down penis, cum on her face and dripping off her tongue" etc.
POV Blowjobs - Tags: "woman, looking at viewer, man, hetero, nude, penis, solo focus, lips, pov, erection, oral, fellatio, male pubic hair, realistic, pov crotch, male pov" It can help to also include how you want your model posed like kneeling, lying on her stomach, on all fours etc. You can also include the view angle like 'from above' and 'close up' if you want.
Pubic hair - Tags: "pubic hair," "female pubic hair," "landing strip pubic hair," "male pubic hair," or "excessive pubic hair" for even more.
📌 Check the pinned image posts for prompts to get started. Another useful technique is using a LLM to incorporate danbooru tags into natural language prompts. Many of my example images were generated that way.
LoRA's:
LoRA's are compatible with this checkpoint if they are trained using this model (Flux Reality) as the base model.
LoRA's trained on this model work with other Flux Dreambooth fine tune checkpoints.
LoRA's trained on Flux Dev are not compatible they will burn the image!!!
LoRA's trained on previous versions of Flux Reality work ok with this version but retraining is recommended to take full advantage of the new features.
LoRA's trained on softcore versions of the base model tend to work better with other softcore versions, and LoRA's trained on hardcore versions of the base model tend to work better with other hardcore versions.
Sampler:
Preferred sampler is dpmpp_2m with sgm_uniform scheduler with 35 steps.
Prompting:
The model was trained using danbooru tags (except '1girl' was replaced with 'woman') but that doesn't mean you have to solely use them. They can be incorporated into natural language prompts which works best. You don't even have to use them if you don't want to. You can prompt in an entirely different language if you want. I tested French and Italian.
Use "realistic" in the prompt for a DSLR photograph effect. Use "amateur photo" to get a more amateur quality. You can also use my Amateur Photography LoRA for Flux Reality which is hosted on Patreon.
Poses:
To get more complex poses that weren't trained into the model use Canny ControlNet!
Workflow:
If the workflow isn't attached to the images you can get it from the #faq channel in my discord server https://discord.gg/UFkPTfTDWs
Description
FP8 Version of V4 - What's new:
The dataset was expanded by 300%. Several new concepts were trained into the model. The common artifacts found in V3 have been eliminated. The model is finally what Flux Dev should have always been and then some!
Concepts:
Cum can now used in the prompt using the tag "cum". You can specify where and how much of it you want. For example: "cum raining down from the sky and covering her body, waterfall of cum, cum on her face and dripping off her tongue" etc.
Spreading pussy and ass from behind can be prompted with tags like "woman, barefoot, pussy, from behind, feet, kneeling, anus, toes, back, soles, realistic, facing away, spread ass, hands on own ass"
POV Blowjobs can be prompted with tags like "woman, looking at viewer, man, hetero, nude, penis, solo focus, lips, pov, erection, oral, fellatio, male pubic hair, realistic, pov crotch, male pov" It can help to also include in the prompt how you want your model posed like kneeling, lying on her stomach, etc. You can also include the view angle and 'close up' if you want. This hasn't been perfected yet but with the right seeds and prompts you can get very nice highly realistic outputs.
Pubic hair - Tags: "pubic hair," "female pubic hair," "landing strip pubic hair," "excessive pubic hair," "male pubic hair."
Check the pinned image posts for prompts to get started. Another useful technique is using a LLM to incorporate danbooru tags into natural language prompts. Many of my example images were generated that way.
LoRA's:
LoRA's are compatible with this checkpoint if they are trained using this model (Flux Reality) as the base model.
LoRA's trained on this model as a base model work with other Flux Dreambooth fine tune checkpoints.
LoRA's trained on Flux Dev are not compatible they will burn the image!!!
LoRA's trained on previous versions of Flux Reality work ok with this version but retraining is recommended to take full advantage of the new features.
Sampler: Preferred sampler is dpmpp_2m with sgm_uniform scheduler with 35 steps
Prompting:
The model was trained using danbooru tags (except '1girl' was replaced with 'woman') but that doesn't mean you have to solely use them. They can be incorporated into natural language prompts. You don't even have to use them if you don't want to. You can prompt in an entirely different language if you want. I tested French and Italian.
Use "realistic" in the prompt for a DSLR photograph effect. Use "amateur photo" to get a more amateur quality. To get an even more amateur feel to the outputs all you have to do is train a LoRA with the dataset consisting of amateur photo quality images. Send me a message on discord or patreon if you would like me to make one for you.
Poses:
To get more complex poses that weren't trained into the model use Canny ControlNet!
FAQ
Comments (26)
Anyone know how to traine a lora for this checkpoints, some link of youtube?
fluxgym is the easiest way but tbh I havent found the need to retrain my already made loras. Playing around with block weight until you find where it wants to be applied and not applying it to every block indiscriminately results in perfectly usable images despite the author's warning.
@RuggedPineapple I think thats great you've found a workaround, this is the first I've heard of this
but without me actually testing it I can't change the warning, and block weights are something I've never messed with, and have no real desire to mess with.
If you ever want to make some kind of tutorial I will take a look at it and consider adding it to the description after verifying it myself
@getphat I use this workflow for testing: https://civitai.com/models/886270
theres also an explainer here: https://civitai.com/articles/8362
for a full run of 768x768 examples for each block takes about 7 minutes on my 4070 ti super, nothing near high end hardware.
if I use a lora often, I make a metadata note about what blocks are most important and what they do. Helps me avoid lora clash mostly, or apply a character lora without even touching pose/composition, or in this case making it play nice with a model it doesn't really know :)
@RuggedPineapple well im glad you've found a fix, but honestly, that is way too complicated for me
You can use flux gym, modify/add custom model of your choice, so you can download this model and add it to flux gym's model.yaml. If you aren't aware on how to modify the file, you can simply ask ChatGPT to do it for you, that's what I did myself. Training your loras on any fine tune gives great results TBH.
@kakkkarot where is fluxgym's model.yaml? Can't find it.
Can this model draw pubic hair without LORA?
v4 which will be uploaded tomorrow does
I strongly suspect this will be used for naughty images of naked pretty ladies doing something disgustingly pleasing with unmentionable things. I am thus looking forward to downloading it merely to look into its capabilities more closely.
Excellent work. Many Thanks.
Thank you, make sure to check out v4, as thats where all the fun stuff is. CivitAI is currently bugging out and not showing the updated model cover properly or linking to the new version.
Would it be possible to add V4 to tensor.art?
no I dont have an account there and my stuff thats uploaded there wasnt uploaded by me
I have trained a LORA on Civitai's LORA trainer using Flux Dev, and it works out of the box with this model. Absolutely fantastic!
If it works for you then great 👍
I still dont recommend it, training on Flux Reality will work way better
@getphat Where can you download Flux Reality, haven't found it so far.
Hey first of all great work brother cant wait for the update to be released. Any tips on how to train a lora with this model as a base? Does it work in FluxGym or do you have another recomendation?
Thanks, never used fluxgym, but kohya ss works very well
It is not possible to train LoRA here on site?
If yes, how?
I don't see that option.
Go to --> Create, then click --> Train LORA
its not up to me, its up to civitai
How can I train my Lora character on this checkpoint?
@Kastro13 I have tried it in FluxGym it works. You just have to upload the Full Model and edit the models.yaml File that it will show up in the drop down menu.
@mclovinofm417 cool. What exatly I need to edit in yaml? and as I remember I need to add huggingface link for this model?
@Kastro13 You can make a symlink. This will make a file called flux1-dev.sft that actually is just a shortcut to this model. That way when fluxgym tries to use flux1-dev.sft it will actually be using this model.
Use this in a terminal in the unet folder.
ln -sf /workspace/fluxgym/models/unet/getphatFLUXReality_v5Hardcore.safetensors /workspace/fluxgym/models/unet/flux1-dev.sft
You will still have to download this model to your unet folder using wget commant.
@mclovinofm417 Where is the yaml file? Could not find it in the fluxgym folder. I assume its not the same as the one in comfyUI
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.