The main goal was to make more 'live' images with livelier emotions and more dynamic poses and slightly amateurish quality (also can produce high quality images).
P.S: Thanks to everyone for the feedback! I've noticed the comments (not only here) about anatomy issues, and I’ve been collecting reports and examples from all of you. I'm considering a couple of options to address this:
Training a full checkpoint, which might be the more rational approach for consistent improvements.
Expanding the dataset with more photos that cover diverse poses to help refine anatomical accuracy.
I appreciate the input - it really helps me shape the future updates
P.S.2: I’ve moved training to RunPod for consistent, high-quality results. If you'd like to support my work and future updates, you can find me on Ko-fi. Cause i plan to fine-tune a model, not just a LoRa. https://ko-fi.com/danrisi
Setting i use in ComfyUI for Flux:
CFG=1, Guidance =2.5, Scheduler=Beta, Sampler=dpmpp_2m, Steps=40, Strength= from 0.8 to 1.0 works good, but i'm usually use with 1, but if hands getting worse then i set 0.87
Setting i use in ComfyUI for SD3.5:
CFG=1, Guidance =3.5, Scheduler=sgm_uniform, Sampler=dpmpp_2m, Steps=40, Strength= from 0.5 to 1.0 works good, but i'm usually use with 0.7
V2 - Flux
Brings even more realism and versatility to your creations, with significant improvements in stability, anatomy, and overall quality. This update makes the LoRA more adaptive, allowing you to achieve various quality levels based on your prompts—from high-definition realism to intentionally lower-quality aesthetics.
Trained on 1048 images.
What's New:
Stability Improvements: The new version is more stable and works better with text-based prompts, providing a smoother and more predictable output.
Enhanced Hands & Anatomy: Hands and body anatomy are more refined, enhancing lifelike quality.
Quality Flexibility: With the right prompts, you can adjust for both high-quality and lower-quality aesthetics (examples available).
V1.2 for SD3.5 - Large
Decided to make a version for sd3.5 almost with same settings. Imo looks good, but i noticed some problems with anatomy (in some moments even worse then flux), but aesthetic (colors, contrast and other stuff looks even better then flux). I see that sd3.5 have potential, maybe new versions will be on sd3.5 too.
V1.2
Changed half of images in dataset, changed prompting style, improved hands, less 'flashlight effects' at night scenes and overall LoRa quality improvement (i hope).
In this version it's unnecessarily to use tons of 'trigger words' like in V1. Just add some of them that i mentioned in trigger words
V1
Trained another amateur lora (that's already been done a lot around here).
Still, I took some pictures from my dataset on the 2000s and added another 700 pictures. It came out pretty good so far, but there is a controversial point about quality optimization, I was hoping it would help control the quality, but as I see it only confused the model. In the dataset there are many different gradations of quality, like:
1) High-resolution photo, shot on a mobile phone, no visible artifacts, clear and sharp
2) Low-resolution, amateur photo shot on digital camera, no visible jpeg artifacts, slightly noisy
3) Medium-resolution photo, shot on a mobile phone, slight graininess due to low light conditions, no significant digital artifacts
And other their combinations and variations. I think i'll remove such in next version. But this one working the best for me:
Low-resolution, amateur photo shot on digital camera, no visible jpeg artifacts, slightly noisy
Description
FAQ
Comments (104)
A 2GB Lora?!?
Yes, I used a high rank for the LoRA to achieve better quality, which is why I'm aiming to train a full checkpoint as the next step
It's a 22GB Model and you want to change of lot of it with the lora ... so yes, 2GB is not too much.
@J77gonzales thanx for understanding =) Yeah, this LoRA isn’t just a detail enhancer - it's a comprehensive collection of real-life concepts that significantly enhance not only details but also poses, hands, feet (though results can vary), and general anatomy (roughly speaking). The higher file size is definitely worth it for the level of versatility and realism it brings
Does it work well with SD 3.5 M ?
Good question. I trained this LoRA specifically for SD 3.5 Large, but I didn't test too much it with SD 3.5 Medium (cause thought it wont work). But for some reasons It work, but only partially - some prompts look good, while others show distortion in details. So, it’s not fully optimized for SD 3.5 Medium, but you may still get decent results with the right prompts
sd3.5 loras are starting to pop in :D hell yeah
Train an 18MB LoRA alongside it and market it as a lite version in addition to the base variant. Let the people decide if this 2GB model is truly necessary. There are so many mediocre multi-GB LoRAs out there, yet some of the best models are only 18MB. 2GB? That's a whole fucking SD 1.5 model. Did you really train it on millions of pictures as well?..
nice. any chance to train it on pixelwave model ?
Thanks, glad you like it! I’ve actually been planning to train my own full checkpoint with a more realistic style soon. In theory, there’s a chance I could train a new version of the LoRA on PixelWave to take advantage of its quality, but I’d need to get in touch with the PixelWave author to see if it’s possible. Still thinking it through, but it’s definitely on my mind!
@Danrisi Hi there, thanks for your answer ! Good new for your checkpoint ! For train on pixelwave, you can now with Ostris AI toolkit by adding "mikeyandfriends/PixelWave_FLUX.1-dev_03" as model path ( transformers files has been added just today! )
Great Lora! But can you please release SD 3.5 medium version as well? 🙏🏻 a lot of people can't run Large model 😞
Thanks, glad you like it! Yeah, I'm planning to train a version for SD 3.5 Medium as well, not just for the Large model. Stay tuned!
@Danrisi nice!!
sd35 large lora will work with sd35 large turbo ?
didn't test, but should work well
@Danrisi you are going to develop this more ?
@amazingbeauty you mean version for 3.5?
@Danrisi about developing this lora (sd 3.5 large - ultrarealistic lora) more ,you're going to do this ?
@amazingbeauty yeah, ofc. i just change a little bit settings and dataset and then train again for flux, sd3.5L and 3.5M. Also i already opened Ko-Fi to earn some money on it, cause 1 training almost 50usd =)
It's a great LoRA with caveat. The hands are really bad, even when enabling hand group, the hands still come out bad.
Thanks for the feedback! I wouldn’t say the hands are always bad, but I get what you mean - certain hand poses can come out distorted. I’m aware that adding more hand-focused images to the dataset could help improve this. If you could share some examples of the hands you’re seeing issues with, it would be super helpful for fine-tuning in future versions
https://github.com/rgthree/rgthree-comfy
The "Lora Loader Stack" node has been deprecated. Used the "Power Lora Loader" node instad. In case somebody encounter some issue.
good to know thanks
Hi! Yep, I've been using the Power Lora Loader too - it's pretty convenient. But, out of old habit, I still reach for the Lora Loader Stack sometimes. As long as it gets the job done, that’s what matters, right? 😄
Please considering fine tuning the dedistilled flux dev model at fp16. Or even use the Verus Vision 1.0b fp16 as your fine tune base model (it's based on flux dev dedistilled). I'm getting some hand issues with this LoRa, but if that won't be a problem when you release a checkpoint.
Works great!
Thanks
Bro! Thanks for sending some love to SD 3.5!!!
Crazy, is there any way to train a lora model on a person and use it with this? (I'm a noob sorry if its a stupid question)
I tried on previous version of LoRa and seems like working good https://civitai.com/images/34591863
yes, you just have to add the loras you want to use, depending on your UI ithere are different methods, but it is totally possible
@Danrisi what I am doing wrong? I added LORA with my character but then Images loose this ultra realism and looks like normal flux images ;/
@GrzesMis this means your lora have much more influence on style. Can you tell me settings with which did u train lora?
@Danrisi I generated 2x2 net of images (4images) with your UltraRealistic Lora plus flux dev fp8 with upscaling and face detailer, then I trained it on civitai, tensor art and flux gym on my pc
here settings from tensor (,2 types of settings) I dont remember settings from civitai and fluxgym. But every setting from tensor and civitai makes LORA similar, it works great but it doesn't aloud for your level of realism
🔧 Parameter Settings number 1
Network Module: LoRA
Use Base Model: FLUX.1 - dev-fp8
Trigger words: modelkav2
🖼️ Image Processing Parameters
Repeat: 20
Epoch: 10
Save Every N Epochs: 1
🧠 Training Parameters
Seed: -
Clip Skip: 2
Text Encoder learning rate: 0.000005
Unet learning rate: 0.0001
LR Scheduler: constant
Optimizer: AdamW8bit
Network Dim: 16
Network Alpha: 15
Gradient Accumulation Steps: -
🏷️ Label Parameters
Shuffle caption: true
Keep n tokens: -
⚙️ Advanced Parameters
Noise offset: 0.05
Multires noise discount: 0.1
Multires noise iterations: 10
conv_dim: 16
conv_alpha: 8
Batch Size: -
📸 Sample Image Settings
Prompt: portrait of a modelkav1 girl
Sampler: euler
🔧 Parameter Settings number 2
Network Module: LoRA
Use Base Model: FLUX.1 - dev-fp8
Trigger words: modelkav1
🖼️ Image Processing Parameters
Repeat: 20
Epoch: 10
Save Every N Epochs: 1
🧠 Training Parameters
Seed: -
Clip Skip: -
Text Encoder learning rate: 0.00001
Unet learning rate: 0.0001
LR Scheduler: constant
Optimizer: AdamW8bit
Network Dim: 64
Network Alpha: 32
Gradient Accumulation Steps: -
🏷️ Label Parameters
Shuffle caption: -
Keep n tokens: -
⚙️ Advanced Parameters
Noise offset: 0.03
Multires noise discount: 0.1
Multires noise iterations: 10
conv_dim: -
conv_alpha: -
Batch Size: -
📸 Sample Image Settings
Prompt: portrait of a modelkav1 girl
Sampler: euler
What checkpoint are people using with this Lora ? It doesn't seem to play nicely with Pixelwave, goes ok with AcornIsSpinningHyper8step but can't seem to get great results like many seem to..
Yeah, it doesn’t work as well with custom checkpoints as it does with the default Flux model
Pixelwave doesn't work well with any loras, sadly.
what setting do you suggest for ForgeUI,I have very over bright results.
with Flux Realistic sampler and 0.8 strenght i had cool results but it doesnt mix well with character lora.
@magmirant0011 Thanks I will try it , Iuse faceswaplab for the characters so shouldn't be a problem
@magmirant0011 dunno. i don't have problems with characters https://civitai.com/images/39390829
LoRA très prometteur, bonne continuation pour le finaliser en modèle
Thanx a lot =)
After generating nearly 250 images, here is my feedback:
1. The model requires 30-40 steps to get a GOOD image at all. (Amateur photo lora requires 15-20 only to get AMAZING images)
2. Potential is higher, but the lora gives too much "amateur" photos. I mean I love amateur photos. But the images are not good.
3. I heard that you have a checkpoint in training. My recommendation is that apart from selecting the amateur photos for dataset, also include some "professional photos" in your dataset, that are realistic, sharp and cinematic (Like in verus vision). You can collect images from sources like LAION (their aesthetic dataset is vast and great and free as well)
Hey, really appreciate you going in-depth with the feedback =)
P.S. I do have a variety of good-quality images in the dataset, but yeah, there’s a bit of randomness in there, which can impact the results. I’ll definitely consider refining the mix to balance that amateur vibe with more professional shots. Thanks for sharing the tip about the LAION dataset, too - that’s super helpful
@Danrisi really happy to see that could share something useful. Definitely excited about the finetune
Hi everyone, i'm trying to use this lora together with another lora trained on a specific person but i'm struggling in achieve good results. My subject tends to no longer resemble the person I trained on when i mix the 2 loras, even when i use low weigth for the realism. Can anyone give me a tip?
Hi! I believe this is the main problem with Loras in general. When mixed with others it's always change the main appearance of the trained Lora person. Only solution I found is to increase the strength of the main Lora to 1.2 to 1.4 and the other lora to 0.8 . It's not perfect solution but it helps. Maybe someone else has another solution but these are the settings I use.
@nontas I think it depends on how the character LoRA was trained. I’ve tested with some celebrities, and it worked well. If you could share the LoRA and what you're trying to generate, I'd appreciate it
you can try to set my lora to 0.87 and your character lora to 1. it's usually helped me on pony checkpoints
I am also trying to understand how to train loras that don't interfere with faces. From what I have seen so far, it seem to work really well when the non-character lora is trained without ANY faces at all; for instance one Lora I used seem to have zero impact on the face, and the author told me it was trained by cropping off all faces from the training set. I also heard that it's possible to train with "masked" images but haven't found how to do that yet. FluxGym has an option for masked images, but I don't know exactly how to build those masked images so ... still investigating.
hello, great lora & i'm loving it btw! i was wondering tho if you could release v2 on tensorart. it would do wonders for me & others too. thank you for this awesome work of art :)
Hi, thanx =)
I uploaded to tensorart, but someone uploaded it even before me
@Danrisi thanks for uploading on tensorart, i wish everyone uploaded there simultaneously, its an amazing site that's cheap to use
great lora but i noticed sometimes when used with character lora or alone, it messes up the text words and affects the Prompt adherence too, in ur next release you should also focus on improving the text quality and the prompting in general!
but great job on ur work nonetheless buddy! 😁👌
Can I use this lora with turbo lora : https://huggingface.co/alimama-creative/FLUX.1-Turbo-Alpha ?
Thank you
Hey can you make a lore for low-tied long hair hairstyle? When given the prompt flux messes up the hairstyle so bad, literally every other model can recreate it. I wish there was a flux lora for this hairstyle
cool
Hey, I really like what you got here. I'm very new, this is only my 2nd day here and this is my 2nd model I'm trying. I'm keep getting the following error. How can I resolve? Thanks!
ERROR: Could not detect model type of: C:\Users\******\Desktop\New folder\ComfyUI_windows_portable\ComfyUI\models\checkpoints\UltraRealPhoto.safetensors
Hi. do you use my workflow or your own? cause i think u placed Lora in checkpoint folder. If you will provide your workflow it will help a lot
@Danrisi, hey again! Thanks so much for your response, apologies for the delay --here on my side I've been experiencing a lot of issues with the website not loading; it'll take like 15 minutes just to load one page.
Honestly, I have no idea how I would share a workflow on here. I followed a Tutorial that was teaching me a lot and the guy in the video said when I add a new Model, I could just change the Model to my new download and it'll work. However, when I tried it with this, it kept giving me errors.
I also hear that we can save any image to drag it into our ComfyUI to obtain the workflows. Is that how I'd use yours? I just tried with the image of the woman on the swing, but it said there's no workflow data found.
I appreciate your feedback and hope to get this working soon, you've done an amazing job! :)
@Danrisi, oooh... I think I know what's causing it. My computer is compatible to run Flux, but every time I try to use it I keep getting errors. So, maybe I'm experiencing issues because it's trying to use Flux which seems broken on my end. I got Flux 1 [Dev], Flux Text Decoder and the AE. I tried downloading these online but they kept failing so I installed them from the ComfyUI Manager. After, it still gave errors saying it couldn't find the Model so copied the file to the /Models folder but it still errors it can't be found. I tried another one of your images to load your workspace and it still says no workspace found. /:
Any idea?
Based on error text you put Lora in checkpoints folder and probably trying to load it like checkpoint which is mistake.
You should put UltraRealPhoto in loras folder and load it with LoraLoader node and connect it with checkpoint/Unet loader
is there an img to img example? with lora
I don't have samples here, but I used lora in some cases: usual img2img, upscale and inpaint, worked good
@Danrisi ReActor makes the face pixelated and blurry, and PulID kills the effects of a real photo
@dimasystem didn't test with reactor. what about pulid, it also depends on values that you used, if using medium values, like start steps 0.2, end steps 0.9, also max and min fusion 0.2 and 0.8 then quality will not decrease too much
Thanks. I have a question. Is this model regarded as 'StabilityAI's model'? so I wanna know, can I use this as a commercial/noncommercial purpose while following the stability's instruction?
works so so well with the NF4 model too and in my humble opinion this is probably one of the greatest if not the greatest realism lora that i used from flux release till this day, like WOW!! i didn't even expect to work with a character lora that well and it actually improves some of the loras too, like how did you even do this and how did u work to achieve this kind of realism?? i can't imagine how the V3 will turn out to be if you plan to work on it or something!!
mate u really outdone yourself, hats off to you really and keep up the amazing work!! 👌👏💙
I -REALLY- liked the quality this brought to my images, but it converted all my guys to female, no matter what I did. Please could you add more male faces to the training? Because that would make this perfect. <3
Thank you!
... I even tried weighting the word "male" at 1.5, adding additional words like "masculine/manly", and using a male-faces lora on top of it. Nothing worked! lol. It is the ultimate girly LoRA. Which is cool because it's trained that way.
But I'd love to use it on my fellas. :)
im unable to run this.. could anyone help me out? im new to this stuff but im fascinated by this model
Hey! Can you please clarify on which step you are and what problem you have?
This is the first time i've seen a 2GB LORA
This is why i made checkpoint, cause dataset pretty big for LoRA. But i plan to retrain this LoRA with less images and rank, so it must weight 128mb (i hope)
BTW, you can check early sdxl loras. I remember I saw some loras that weight 6-8gb
Best quality comes from high DIM - (imo) 64/32 works best for me. no complaints here, keep up the good work!
Maybe I'm dumb, but is there a comfyUI workflow to go with this? Where can I download it?
👋 You just need to go under the image where the prompt is located and find Workflow label, so near will be button button "nodes". Press on it and then ctrl+v on comfyUI screen
@Danrisi Thank you so much!
will this work with Web Forge? Thanks
Will this work with FORGE
Hmm, am I understanding correctly that this LoRa only works in the Comfy UI? I just can't do anything on Forge at all: either in good quality, or everything is smeared. I took the prompts and basic settings from the examples on this page
Amazing and quick results with comfyUI 8_steps workflow, just chain another LORA loader just after the lora turbo... Enjoy!
despite good quality i frequently encounter artifacts
I like what this does to realism in terms of making generated images look less like covers of a marketing brochure but more real-life pictures.
I feel like it needs more knowledge of smaller faces and body parts. When you have your main element of the image surrounded by people in the background, their faces and hands quickly become messy.
for better small details i can recommend generate image in 2.0MP (in 1408 x 1408 instead of 1024 x 1024). but yes, i want to update this lora (cause it uses lots of people who can't use my custom checkpoint)
awesome, it fix feet and hands issue, and give best result, love it
Hope you don't mind my bid to see it boosted, I use this LoRA a ton and send a lot of praise and wanted to show something tangible for that.
Hi. Thanx a lot ❤ I have in plan to update this lora, just need some understanding how to optimize it as much as it possible (I already trained some versions but I'm not satisfied with a result)
bonjour comment utliser vous ce lora
there is a AIO workflow?
Nice lora works very well
I can't make consistent models with the same face using this LoRA. Are there any tricks?
pulid?
also very curious about it
Yeah, this LoRA changes faces a bit, so even some LoRAs can't guarantee consistency. For better face consistency, it's better to use my UltraReal checkpoint + Samsung LoRA + character LoRA
Danrisi do you have any recommendation on weights? using 2 loras at the same time doesn't always work great..
ncbzmdtstdwooyiawg472 yes. if your lora with character has a big amount of steps (smth around 3k+), then i recommend other loras to lower down significantly (u need to test it manually). but yes, that's a problem of using many loras at the same time
I'm trying to get this to work on my 5090 and it does work, however it takes forever like 15 min, and the vram is always at 50% is there a change I can make so it goes balls deep on my 5090's v-card
I was wondering is there was a change I can make to maximize speed with a 5090, its like 15 min @ 50% vram usage
To make a single image? I have a 5090 and making 4090 x 5140 size image takes about 30 secs?
@pursuit_of_beauty lol, yeah dude I've evolved, this was 3 months ago, I'm at about 30 sec now too.
This is amazing for doing groups of people. It gives you lot's of variation in the faces instead of making a group of twins.
holy shit this thing is amazing man good job
Details
Files
UltraRealPhoto.safetensors
Mirrors
UltraRealPhoto.safetensors
UltraRealPhoto.safetensors
UltraRealPhoto.safetensors
UltraRealPhoto.safetensors
UltraRealPhoto.safetensors
UltraRealPhoto.safetensors
UltraRealPhoto.safetensors
ravzofluxultra_.safetensors
UltraRealPhoto.safetensors
UltraRealPhoto.safetensors
realistic-amateurishphoto-lowlighting-inmotion-overexposed-underexposed-goprolens-eerieatmosphere-smearedbackground-smearedforeground-flux.safetensors
ULP.safetensors
flux_train_replicate.safetensors
UltraRealPhoto.safetensors
UltraRealPhoto.safetensors
UltraRealPhoto.safetensors
UltraRealPhoto.safetensors
UltraRealPhoto.safetensors
UltraRealPhoto.safetensors
FLUX_UltraRealPhoto_amateurish_photo.safetensors
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.