Check my exclusive models on Mage: ParagonXL / NovaXL / NovaXL Lightning / NovaXL V2 / NovaXL Pony / NovaXL Pony Lightning / RealDreamXL / RealDreamXL Lightning
Recommendations for using the Hyper model:
Sampler = DPM SDE++ Karras or another / 4-6+ steps
CFG Scale = 1.5-2.0 (the lower the value, the more mutations, but the less contrast)
I also recommend using ADetailer for generation (some examples were generated with ADetailer, this will be noted in the image comments).
This model is available on Mage.Space (main sponsor).
You can also support me directly on Boosty.
Realistic Vision V6.0 (B2 - Full Re-train) Status (Updated: Apr. 4, 2024):
- Training Images: +3400 (B1: 3000)
- Training Steps: +724k (B1: 664k)
- Approximate percentage of completion: ~30%All models, including Realistic Vision (VAE / noVAE) are also on Hugging Face
ᅠ
Please read this! How to remove strong contrast.
To make the image less contrasty you can use LoRA [Detail Tweaker LoRA] in a negative value.
ᅠ
Orange Color = Optional
ᅠ
I use this template to get good generation results:
ᅠ
Prompt:
RAW photo, subject, 8k uhd, dslr, soft lighting, high quality, film grain, Fujifilm XT3
ᅠ
Negative Prompt:
(deformed iris, deformed pupils, semi-realistic, cgi, 3d, render, sketch, cartoon, drawing, anime), text, cropped, out of frame, worst quality, low quality, jpeg artifacts, ugly, duplicate, morbid, mutilated, extra fingers, mutated hands, poorly drawn hands, poorly drawn face, mutation, deformed, blurry, dehydrated, bad anatomy, bad proportions, extra limbs, cloned face, disfigured, gross proportions, malformed limbs, missing arms, missing legs, extra arms, extra legs, fused fingers, too many fingers, long neck, UnrealisticDream
(deformed iris, deformed pupils, semi-realistic, cgi, 3d, render, sketch, cartoon, drawing, anime, mutated hands and fingers:1.4), (deformed, distorted, disfigured:1.3), poorly drawn, bad anatomy, wrong anatomy, extra limb, missing limb, floating limbs, disconnected limbs, mutation, mutated, ugly, disgusting, amputation, UnrealisticDream
ᅠ
Euler A or DPM++ SDE Karras
CFG Scale 3,5 - 7
Hires. fix with 4x-UltraSharp upscaler
Denoising strength 0.25-0.45
Upscale by 1.1-2.0
Clip Skip 1-2
ENSD 31337
ᅠ
Thanks to the creators of these models for their work. Without them it would not have been possible to create this model.
HassanBlend 1.5.1.2 by sdhassan
Uber Realistic Porn Merge (URPM) by saftle
Protogen x3.4 (Photorealism) + Protogen x5.3 (Photorealism) by darkstorm2150
Art & Eros (aEros) + RealEldenApocalypse by aine_captain
Dreamlike Photoreal 2.0 by sviasem
HASDX by bestjammer
Analog Diffusion by wavymulder
Life Like Diffusion by lutherjonna409
Analog Madness by CornmeisterNL
ICBINP - "I Can't Believe It's Not Photography" by residentchiefnz
Description
NOT FOR TRAINING!
Model for inpainting and outpainting.
Make sure that the model name has the ending -inpainting.
vae-ft-mse-840000-ema-pruned.ckpt included
FAQ
Comments (241)
Where is the VAE? Is it integrated in the model and gets selected automatically? There is no option to select it in automatic1111 webinterface settings...
Hi. In version 1.3, VAE is integrated into the model and works automatically.
when I run 1.3-inpainting...
RuntimeError: Error(s) in loading state_dict for LatentDiffusion:
size mismatch for model.diffusion_model.input_blocks.0.0.weight: copying a param with shape torch.Size([320, 9, 3, 3]) from checkpoint, the shape in current model is torch.Size([320, 4, 3, 3]).
I will try to solve this problem.
Try downloading the config file and put it in the folder with the model, make sure that the model has the name ending -inpainting and that the configuration file has the same name as the model.
@SG_161222 I can't get the models to load in A1111. I've downloaded the model and the configuration file, and I put them in the correct file location.
I named the files the following:
realisticVisionV13_v13Inpainting-inpainting.safetensors
realisticVisionV13_v13Inpainting-inpainting.yaml
What am I doing wrong? If the files are named in correctly, can you type it out correcly for me here? Thanks!
@donald_diffusion Hi. The files are named correctly. Try running A1111 with webui.bat
Once you have tried it, let me know.
@donald_diffusion You can also shorten the name a bit to "Realistic_Vision_V1.3-inpainting.safetensors" and "Realistic_Vision_V1.3-inpainting.yaml"
@SG_161222 got it working! Thanks
@SG_161222 i just changed the name to realisticVisionV13_v13-Inpainting.ckpt . But ... File "D:\AI\stable-diffusion-webui\repositories\stable-diffusion\ldm\models\diffusion\ddpm.py", line 448, in init
super().__init__(conditioning_key=conditioning_key, args, *kwargs)
TypeError: DDPM.__init__() got an unexpected keyword argument 'finetune_keys'
@Layscream Did you update Automatic1111?
@Layscream You can also try to perform these actions:
Try running A1111 with webui.bat
You can also shorten the name a bit to "Realistic_Vision_V1.3-inpainting.safetensors" and "Realistic_Vision_V1.3-inpainting.yaml"
@SG_161222 thank u! i will update and try again~
@SG_161222 Where do you get the config file??
I have the same exact error and can't seem to get it to work. Using Easy Diffusion.
@jimmyjazzjazz https://postimg.cc/NyrqPJB8
I'm having serius problem with this model, it always generate ugly malformed faces, even if i put all my input efforts in prompting nice faces
ive found having strange dimensions to your photos influences this heavily. stick to standard widths and heights
Hi. What GUI are you using? What size image are you trying to generate?
@SG_161222 AUTOMATIC1111 512x512
@Sirh You can also try version 1.4
The link is in the first post: https://boosty.to/evgk.132
Can you help me please, did a clean install and added the model, I keep getting this error when trying to generate an image, also made sure everything was named correctly. Much appreciated!
File "C:\stable-diffusion-ui\stable-diffusion\env\Lib\site-packages\torch\nn\modules\module.py", line 1497, in load_state_dict
raise RuntimeError('Error(s) in loading state_dict for {}:\n\t{}'.format(
RuntimeError: Error(s) in loading state_dict for UNet:
size mismatch for model1.diffusion_model.input_blocks.0.0.weight: copying a param with shape torch.Size([320, 9, 3, 3]) from checkpoint, the shape in current model is torch.Size([320, 4, 3, 3]).
Hi. This is a very strange situation. Some people have my model working, and some people don't have it working. Then we will try to solve this problem. What version of Python do you have installed? Have you updated the Automatic1111 GUI?
i am pretty sure ive done everything correctly (downloaded the model, and put the model in the right folder) however whenever i try to generate images, it is "in queue" and never generates. is it because its not ckpt file? or could there be another issue? I'm sorry, im very new to this
hi i created a model with the previous version just fine, however with this version im getting an error. this version was working fine with the right yaml, but when creating a model i get this error : Exception training model: 'Given groups=1, weight of size [320, 9, 3, 3], expected input[2, 4, 56, 64] to have 9 channels, but got 4 channels instead'.
any ideas ?
my model spent all days training 4000 images, and when it came time to do the last quick swoop of everything before generating the ckpt i got this error. is there anyway i can fix that issue, or at least use the images generated with another model ? so i didnt waste all this time for nothing
thanks :D
Hi. Sorry for the inconvenience. What model did you use for training?
i used your model -inpainting and trained it with my custom pictures (which ive done yesterday) with the non inpainting model and it worked fine. i did the exact same thing and got this error
@chrissy22 You can make an inpaint version of the model from your model, I can send you a post with information on how to do it.
@SG_161222 i dont have a model, i trained my images in dreambooth, i would love to be able to somehow pick back up from the generated images i was able to make and not restart the whole generation progress. also when crashing it generated a ckpt file which appears to be bigger than realisticvision13-inpainting, which should mean my data is there, but i cant do anything with it
@chrissy22 Can you upload the ckpt file to Google Drive?
@SG_161222 uploading
@chrissy22 Started downloading. So you can't open the model?
@SG_161222 no it doesnt open since it only went thru the first part and longest part of generation. its only missing the last part before it generates a ckpt. which is why its confusing that i even have one.
@chrissy22 I downloaded the model and just opened it in Automatic1111. I did not see any errors. Have you updated Automatic1111? :)
@chrissy22 I can also say that inpainting models do not lend themselves to training.
@chrissy22 So you have to train the regular version of the model and then combine it with the inpaint version to get the inpaint model with your new data.
@SG_161222 does it mean i spent the whole day training my images for nothing ? :/ im updating now and trying
@chrissy22 Yes, it was a waste of time :(
Sorry for the delayed answers. I use a translator and it doesn't always translate the way I want it to.
@chrissy22 It's my fault for not marking that the inpainting model is untrainable.
@SG_161222 but how did you manage to open it ? i can not open it and have the latest version :/
@chrissy22 Try deleting the .yaml file and running Automatic1111 via webui.bat
@chrissy22 But it is still useless, because the inpainting model has a different structure and most likely all your data is not in it.
@SG_161222 alright thank you, also one more question, i trained a model on the non -inpainting version and it worked great, however i can only get closeup pictures, i have added 280 pictures to train, all different angles, some full body, some face only, why is it that is only gives close up results (for example boobs to face, sometimes waist to face ) i can not get a proper full body result, what kind of pictues should i add to the training ? if i crop the picture to 512-512 i can not get the full body in the pictures that i want to train, this is complicated. thanks
@chrissy22 Perhaps for this you should make a text description in .txt format for photos. txt file must have the same name as the photo.
It's a pretty broad topic.
@SG_161222 i have no idea how to do it easily, i type my prompt in dreamboot and forget about it, im new to this lol now i have to spend all night retraining :( thanks for your help tho
@chrissy22 Try with a few images :)
Also don't forget to specify in "Class Prompt" the value [filewords] so that the model is trained based on the text description.
Hi dev just letting you know i am getting error message when trying to load your model in to invokeai
the error message is-
raise RuntimeError('Error(s) in loading state_dict for {}:\n\t{}'.format(
RuntimeError: Error(s) in loading state_dict for LatentDiffusion:
size mismatch for model.diffusion_model.input_blocks.0.0.weight: copying a param with shape torch.Size([320, 9, 3, 3]) from checkpoint, the shape in current model is torch.Size([320, 4, 3, 3]).
it won't let me even change to your model in the model manager, I am a noob so don't really understand any of this.
thanks
What version? I was able to load the 1.3 .ckpt version in InvokeAI version 2.5.0 without any issues.
Super noob question. I know how to train and create a .cpkt file and then launch Automatic 1111 with that. How do I use this file? And how is it different than a .cpkt file? (Or what am I searching for in google to find out?) Thank you!
Use it the exact same way as .ckpt files. SafeTensor is just safer for the end user,
hello sir what how can i use the vae from huggingface i didnt use vae included bc it woudnt let me train using it
Hi. You need to download the VAE file then put it in this path *\models\VAE if you are using Automatic1111. Then start Automatic1111 and go to Settings -> Stable Diffusion -> in SD VAE select vae-ft-mse-840000-ema-pruned.ckpt -> Apply Settings
Any way to train this on the fast-dreambooth colab? I keep getting errors when trying to convert to diffusers.
You can find diffusers on Hugging Face: https://huggingface.co/SG161222/Realistic_Vision_V1.3
@SG_161222 thanks! Amazing model btw
Can someone please explain for dummies that how you can use this model with dreambooth? Tried the version also from Huggingface but still errors. Other models work well so what should be done for this model prior to dreambooth?
It seems when i make an image with two people it applies the same description to both, even when they are described separate and very different from each other in the prompt.
i.e if i say one is japanese, they both will be, or if one is supposed to be a policeman, the others outfit will take on police-like properties. how do i weigh against that in the prompt so that images with two figures can have different looks.
I think this is a problem with many models.
You can use the inpainting version to put a mask on the person you want to change (not a 100% way).
@SG_161222 does this model use parenthesis and brackets to weigh positively and negatively? or another way perhaps?
@bbackagainn Yes, brackets can be used with this model.
That's quite common. I think it happens with all the models.
Dreambooth:
I got error while trying to train with this model: https://huggingface.co/SG161222/Realistic_Vision_V1.3
Error:
Traceback (most recent call last):
File "main.py", line 907, in <module>
if trainer.global_rank == 0:
NameError: name 'trainer' is not defined
Any ideas why I got this error?
@SG_161222 Thank you for your reply! I'm using JoePenna Repo on Runpod and everything mentioned in that post seems to be ok. I can run 1.2 with no issues: https://huggingface.co/ykurilov/realistic_vision_diff
Is there some changes from 1.2 to 1.3 affecting to this?
@Physical_Arugula_115 Most likely a bug in the model, I'll try to fix it in a future update.
The new version will be finished soon and after the tests will be uploaded here and on Hugging Face.
@SG_161222 Thank you! You have done great work with the model!
I am new to this, can I create a custom model and use it in tons of different scenarios?
Yes, you can do that, for example using Dreambooth or some of its counterparts.
I love it but idkw it keeps adding "analog style, modelshoot style, nsfw, nudity" to the beginning of my prompts. anyway to avoid this?
Can you tell me more about the problem? :)
@SG_161222 when i generate something it doesn't add to the prompt perse it add to the metadata in the image i have one that starts "analog style, modelshoot style, nsfw, nudity analog style, modelshoot style, nsfw, nudity (Masterpiece photograph:1.4),of (Ultrarealistic:1.3), High Detail RAW photo Professional full body shot of...." let me know where i can upload images and they conserve metadata to show you some examples of the problem
@efrax What GUI do you use?
@SG_161222 automatic1111 web ui. i thought it was maybe the version but i re download it and same problem
@efrax Very strange, I have now checked the metadata and there is no such thing, only the prompt, the negative prompt and the generation parameters. You can upload pictures to https://postimages.org/
@SG_161222 https://postimg.cc/gallery/fSkmCrS
@efrax In the metadata I found that you use some template that includes these very words (analog style, modelshoot style, nsfw, nudity analog style, modelshoot style, nsfw, nudity), maybe that's the reason?
@SG_161222 no, here some more without template https://postimg.cc/gallery/jrRV2mF
@efrax I found the source of the problem. It's the model-keyword extension.
@SG_161222 how can i fix it? i tested with realisticVisionV13_v13 and realisticVisionV13_v13VAEIncluded both safetensors
@efrax Try disabling this extension in the Extensions tab of your Automatic1111, then apply these settings using the Apply button and reload the UI
@SG_161222 it looks like its solved. You are a god. how did you find out without knowing i have that extension?
@efrax Just assumed, downloaded this extension and checked it myself and it gave the same result as yours :)
inpaint doesn't work with controlnet in automatic1111. Has anyone experienced this?
RuntimeError: Given groups=1, weight of size [320, 4, 3, 3], expected input[2, 9, 80, 120] to have 4 channels, but got 9 channels instead
This may be a problem with version 1.3. I am now preparing a new version for release. After successful tests I will post it together with the inpainting version.
How do I train this model with my photos?
You will be able to train the model in your photos with Dreambooth. There are many videos on Youtube on this topic. If you have any problems, feel free to contact me :)
Hey I want to make it a live model, how can I do it ? Like I want to generate images as your model does on mage space ? How do I do that ? I want to make a website like mage with your model only, please guide. Would love to support you as well.
Can't train this model on fast-dreambooth
conversion error
You can try to do it with the new version of the model (it should have less problems, but is still being tested): https://boosty.to/evgk.132
@SG_161222 Thank you, i will try it for you <3
@SG_161222 Just want to tell you, your new version of your model is amaaaaaaaaazinnnnnnnnggggggg!!!!!!!!!! keep doing great mate! <3
@lysella Thank you so much for the review :)
@lysella could you train the new version? I know these errors, so many models have them :(
@omenizer I haven't try it yet
Do the following: Checkpoint merger > Primary model A set to v1.5pruned > Secondary model B set to RealisticVision > Set multiplier to 0.99.
It will now identify the model as v1.5 and not have conversion errors, Works to convert all models if they need it. You can also use other models as the base A model if you know they work without error in dreambooth, I'm just using 1.5 as an example. When multiplier is set to 0.99 there will be practically no effect to model B visually.
I'm getting odd results, tan or wood like colors, with just rand text or just photo like pictures that have nothing to do with the prompt. Anyone know what kind of error that would be?
Hi, what GUI are you using?
I am using NMKD GUI and getting similar results.
@girlswithafros What version are you using? The model should run stable on version 1.9.1
I don't think the model is working, I'm trying to write a prompt and I'm only getting portraits that don't have anything to do with the prompt, I'm using automatic111
What version of the model are you using?
@SG_161222 1.3 the newest one
@reddot You can try the new version of the model, which is still being tested and will be improved in the future (should work stable and without errors): https://boosty.to/evgk.132
If I have the VAE already selected in A1111 do I need to turn it off for this model? If so, is there any way we could get a version without the VAE?
Hi, you can test the new version of the model, you can find the download link on this page (this model does not have VAE included): https://boosty.to/evgk.132
Important note: "RAW photo" in the prompt may degrade the result.
"If I have the VAE already selected in A1111 do I need to turn it off for this model?" - it does not matter.
Any idea why i always get multiple copies of the same person , i only want 1 person in the picture , not 3 of the same person
Which graphical interface do you use? In which resolution do you generate the image?
Bigger resolution usually clones the subject in all models. Maybe that's your problem.
i installed and use ControlNet to control the framing and pose, and it save a LOT of trial and error getting it to have the subject the way I want.
@Blacksmith yeah, I think these base models are trained on 512 or 768 resolution images, so going above 768 will start to "tile" the contents it seems.
@westingtyler can these image be upscaled to higher definition like 1920pixels with good details
fantastic model and great with hands, but it wants to default to people wearing shorts or skirts, showing their legs and knees. even "49-year-old man school teacher" puts him in a mini-skirt unless I clearly specifi (jeans pants:1.3) and even then it's hit or miss. apart from that, great model.
You can also try version 1.4 :)
The link is in the first post: https://boosty.to/evgk.132
Still my favorite model ever. I saw a link to test the 1.4 version but CKPT only, can you make it safetensors? Thank you
Wait a while, as soon as I upload the model, I'll tell you.
Done. I updated the link, you can download the model :)
https://drive.google.com/drive/folders/10Xet_fDOIM9DBE9BYoQS2VwnWVA5dAOw?usp=share_link
@SG_161222 Lovely. Thank you
@SG_161222 I've been testing it for a few hours and for now I notice that the 1.4 tend to produce empty/white background more often than the 1.3. It's not really a problem but I have to write a prompt about the background too if I use the 1.4.
@Tacite Good, thanks for letting me know about it :)
Dear Sire, im with diffusionbee on macbook. and every time i tried to import the model(1.3 or 1.4test ckpt), it gives me the following error info: Error Traceback (most recent call last):
File "convert_model.py", line 28, in <module>
KeyError: 'state_dict'
[50607] Failed to execute script 'convert_model' due to unhandled exception!
Perhaps this information can help you solve the problem: https://github.com/divamgupta/diffusionbee-stable-diffusion-ui/issues/372
Try using a icognito mode
I also used diffusionbee first, performance was not great so I switched to google colab, just follow this tutorial - https://saashed.com/how-to-run-stable-diffusion-for-free-on-google-colab-in-less-than-10-minutes/ I don't understand why people still continue to use stable diffusion on their own GPU when clearly it is not well optimized to run these. With colab, you get free decent GPU access for few hours every day. That should be good for most casual users.
How do I post an image here?
You can attach it when you want to leave a review :)
Thank you! Can you do multiple reviews?
@theonewhodiffuses Perhaps I misunderstood. But you can leave multiple reviews.
Really a great blend, one of the best !
Can you upload a smaller pruned fp16 version, without baked vae, for even greater flexibility ?
Hi! Unfortunately I won't be able to cut VAE from version 1.3, but you will be able to test version 1.4 beta (without VAE). https://huggingface.co/SG161222/Realistic_Vision_V1.4
That's bad. I guess I'll wait for the 1.4 final version: will it be vae baked too ?
Can you make the actual v1.3 smaller, a 2gb pruned fp16 safetensor version ?
Anyway, I remember of a script that can remove the vae from vae baked models:
https://github.com/Akegarasu/sd-webui-model-converter
"convert/copy/delete any parts of model: unet, text encoder(clip), vae"
@ritcher1 Version 1.4 will be without VAE.
I will try to remove VAE from 1.3 and reduce the model size.
@SG_161222 Thank you very much for your efforts. Waiting for any news.
@ritcher1 I uploaded the pruned model to the website. I tried to remove the VAE, but I don't think it worked.
@SG_161222 Thanks, just downloaded. I'd change the filename in realisticVision_v13_pruned to remind the users what they got.
@ritcher1 Renaming the file does not help in this case :)
@SG_161222 It's said that the pruned versions aren't so good in creating new mixes, so knowing if a model is full or pruned could help.
I'm new around here. What folder inside of model folder do I place the download file? I am using stable-diffusion-webui. Thanks and sorry for the n00b question
Where can I find R-ESRGAN General WDN 4xV3 please?
Any tips on getting better eyes/pupils with this model? Most of the time, even if I follow the high-res fix tips, I just get very low detailed ones
You can use this information: How to Fix Face and Hands with Inpaint
Я тебя добавил VK, хочу тебе сообщение отправить.
Curious: Do most people on here generate images locally on their computers or do they use a generator website?
some on colab or similar websites, others on personal computers.
@sdsdcs12qw I run codeformer and gfpgan on colab - not the fastest, but it works. You can run models on colab? Is that entirely pc dependent? I guess I'm not sure what resources colab is using or if they're on some server in google-land. I've used openart, dreamlikeart, getimg, dall-e....not even sure what other sites are good. But I would like to know how the colab thing works for these models if you know...thanks so much for replying!
@kleerp2018 Yeah, you can easily run this or any custom models on google colab. A free colab account is fine too. Very decent performance and can use it for roughly 3-4 hours daily max before google takes away the GPU. Just follow this guide to use colab with stable diffusion https://saashed.com/how-to-run-stable-diffusion-for-free-on-google-colab-in-less-than-10-minutes and this one to download civitai models to colab https://saashed.com/how-to-easily-download-use-custom-stable-diffusion-models-from-civitai-in-google-colab/
@sdsdcs12qw Thanks so much! I assumed it was running off one's own gpu so I never tried anything more than gfpgan and codeformer. I'll give it a try! Thanks again!
@kleerp2018 I run on my own pc, I have a 2060 with 6gb vram you can run stable diffusion webui on your pc if you have at least 4gb vram with the --medvram parameter.
https://github.com/AUTOMATIC1111/stable-diffusion-webui
local on my laptop with 16gb vram
I generate images locally on a 3080 TI with 12GB of VRAM. It bogs down the video card when generating, but it only takes a few seconds or a couple minutes to generate a single image depending on the resolution and number of steps.
Thanks for this Great Model !!!
and please，why does the my character always have to be facing the front and looking at the camera（viewer）？What can i do to change that？should i use some specific words in the prompt？like what？
thanks again ！！！peace
Hi, thank you. I will try to fix this problem in a future update. For now you can experiment with the following words in the prompt: [front view|side view]
Try not to use "close-up" and "portrait", use just a photo and maybe these, inbrackets:
Wide-Angle Shot - Pull the camera back from subject
Ultra-Wide Angle - Pull the camera very far back from subject
Satellite View - Pull the camera into the stratosphere
Eye-Level Shot - Camera even with head of figure
Far-Shot Angle - Figure looks very small/distant
Medium-Shot Angle - Looking up at figure from knee level (child gazes up at parent)
Ground-Shot Angle - Looking down at figure from knee level (parent gazes down at child)
Low-Angle Shot - On ground, looking up at figure (puppy gazes up at human)
Full-Shot Angle - Your target in full view
Full-Body Shot - Head-to-toe view, like full-shot angle
Glamour Shot - Different zooms, but always flattering to figure
Cinematic Still Shot - Emphasis on context/background
Also you can use controlnet with depth/hed/openpose models to catch up a pose that you like from any photo
@SG_161222 thanks pal！looking forward to your update！
@RoninRex Really helpful ！thanks man！
OpenPose Editor + ControlNet can give you a bit more control on poses and head positioning. Something to look into.
You can also try giving the prompt specific features to focus on. I've used (back, butt:1.3) before. A bit crude lol but it works. The 1.3 weight helps to focus the output.
You could also try adding these as the first few words in your positive prompt: (1/4 profile shot) or (profile shot) or even (editorial photograph)
@ovalshrimp 👍👍👍
@ovalshrimp cool
I can't get it to make a photo where it's a dark bedroom. No matter what I do there's lots of sunlight. Trying to make it look like a selfie taken at night. Any help?
This LoRA can help you with that: https://civitai.com/models/13941/epinoiseoffset
One trick you could use to darken a picture is controlnet. It's a little complicated but it will work. First, generate your image then send it to img2img. Within img2img, open controlnet and import that image you just generated into the controlnet panel. Set the preprocessor to depth and the model to depth. Leave all the other settings alone in controlnet. Now, grab a picture of a random black background from online and add that to img2img pane at the top of the webui. You'll be removing your picture that was added there and replacing it with the black background. You still want your original picture in controlnet, though. So, black picture up top, your generate image down below in controlnet. Now, Click generate. You should see a dark version of your image. You can then adjust the denoising strength up or down to fine tune it. Lowering it will make it darker. This method actually works great in tandem with the offset noise lora.
"Commercially licensed" for fantasy ai but you can't even credit all the model makers who did the actual work for you, how do you know if you can even license this in the first place considering that? I suppose if I make a merge but forget to credit you and then commercially license it in turn, that's fine? lmfao
The model is still free and under the creativeml-openrail-m license.
About the models I listed in the description: I merged my model for myself and did not plan to release it to the public, so I can not remember which models were used and which were not.
If you have any more questions, I'll try to answer them.
@SG_161222 If it falls under creativeml-openrail-m license that means anyone can still commercially use the model. So what's with fantasy AI's claim about commercial use exclusivity? Do they only mean image generation sites but not individuals?
@twistedellie Hi! This will only affect image generation services. You can still generate images that you can sell, and it doesn't limit the use of the model in any way.
If you have any more questions, feel free to ask :)
@SG_161222 Ah, that clears up a lot. You probably want to put that in your description to dissolve many people's worries. Thanks for answering. ^^
Can anyone advise me. I've downloaded all model of 1.3. I load them into InvokeAI successfully. However rendering and image always produces something from an entirely different genre. Example, prompt = portrait of a woman. Results are images of money, or a piece of furniture, a vase or anything from the 1800 time period. Help!
I am not sure about how it's called in invoke AI, but in Automatic1111 the option is called CFG and you have to put it higher number so it generates things closer to your prompt. If you use low CFG your prompt may get completely ignored.
So exactly is fantasy.ai claiming rights to a model merged so many times which models weresed aren't even remembered?
The model is still free and under the creativeml-openrail-m license.
About the models I listed in the description: I merged my model for myself and did not plan to release it to the public, so I can not remember which models were used and which were not.
If you have any more questions, I'll try to answer them.
There are models on fantasy.ai that don't even disclose what they merged, let's not punish @SG_161222 for being honest. Fantasy know what they are buying, and they are the ones that should deal with it. It's impossible to check the license on this merges of merges of merges.
If we do everything by the book, none of the popular models can be hosted commercially by anyone.
IANAL, but my suggestion to fantasy.ai and anyone thinking about doing this is use trademarks instead of copyright. You should probably avoid trained copyrighted models like dreamlike and novel ai, or consider getting a license from them. If you can't see if the stuff you merged has it, they probably can't either. But test how your model react to their trigger words to be sure.
@fbmac I’m sorry, but if a company like fantasy.ai is going to charge anything at all, they should be training their own models from scratch. I’ve also never understood the Dreamlike license thing, seems like an honor code thing and not actually enforceable after a simple PNG info removal.
Just trying to host free models and showcasing the names of them on their homepage is kinda a joke. The ONLY people that have any clue what those model names are, are not the target market for paid Stable Diffusion. This shit is already free, and totally untraceable when it comes to the model used. Licenses on any of these models are an unenforceable joke, ESPECIALLY merges.
As a photographer though, I could see training a model on my own work and charging for that with a license, but first it’d have to be good enough to warrant payment over the endless free options, and then I’d have to distribute somewhere other than Civitai so that I could actually sell it initially. And even then, the first person to buy it would just upload it here. And next it’s 10 layers deep in a merge someone else is trying to sell.
People are just desperately trying to figure out how they can monetize this before someone else and throwing horrible websites up. Like, we get it Fantasy.ai, you know how to change fonts.
I saw the latest update today (March 9th), but they are all v1.3, so what is updated? Where can I see update records？
Hi. The description has been updated.
@SG_161222 emmm... where is it? I don't see anything different in the description than before. :(
@wen834 New information in the text about the "Official Partner" :)
@SG_161222 publish your v1.4 here. community will be happy
@ManuManu I'll do it tomorrow :)
@SG_161222 Still doing it?
@skullzy77 I won't be able to do it yet until my weekend. Version 1.4 is available on Hugging Face: https://huggingface.co/SG161222/Realistic_Vision_V1.4_Fantasy.ai
@SG_161222 Oh nice! Thanks
Lmao. “Licenses.”
The second you put a model on Civitai or HF for download, you can kiss that license goodbye.
The ONLY way anyone is ever going to monetize their model, is if it’s exclusively locked down within their own webui environment that keeps record of generations.
The second you give that model out to the public, it’s open season. You cannot prove what I used to make my own images locally, especially once scrubbed of any PNG info.
Unless you’re training models from scratch/only merging with other models YOU made from scratch, again, “licensing” it is a joke. You also need to be the exclusive owner of any content you train the model on to truly be the license owner. And even then, you better keep that source ckpt on lockdown before it gets merged a million times without evidence.
Training a model from scratch on someone else’s work doesn’t count either. Which is what all of the “big” models are.
I laughed when I uploaded an embedding of a person here and it asked me about licensing. Imagine owning a commercial license for an embedding of someone else’s face. This is all legal *make-believe*.
Please stop putting license talk in these uploads until they’re actually enforceable.
Same thought and if these people keep doing these licenses jokes, don't worry, more people will jump with better models and will wipe them out, period.
In fact a lot of "licensed" model are using protogen models (among others in the mix), this is a fucking joke. I will only respect licenses for those who really took time to drew these images and train them as a model, if not, they're just training over MJ/SD generations so please, as an artist in the industry, this is total lame and you should start looking more at the community instead of trying to make as muas as $$$ possible.
People are desperate to make money, smh.
"Look guys I merged 10 different models in civitAi but I forgot the rest of them, so now if you use the my checkpoint please pay three fiddy $$".
It's like he's owning all of the images that all of the models that were trained upon.
@ManuManu Now show me where I demand money to use the model?
Problem is the training data is a legal quagmire. For this same reason, industry (film, entertainment) is unable to use AI generated content for commericial production until these legal issues are sorted out (if ever).
@Fyyre already hitting commercial markets actually https://www.engadget.com/levis-will-supplement-human-models-with-ai-generated-fakes-190011557.html
The AI company they’re using looks to be creating their own training data at a quick glance
exclusive 1 star for exclusive fatnasy.ai
why many peopel use SD1.5 instead of 2.0?
1.5 is way better than 2.0
Short answer: people proompt for coom and SD 2.x has no nudity, so people proompt in 1.5 instead.
Long answer: SD 1.5, being a 512x512 base model, has smaller VRAM requirements compared to 2.x, so naturally more people can use it and therefore why it is so popular. SD 2.1768 produces photo-realistic outputs with far better fidelity, but it's a 768x768 base model, meaning it has more robust VRAM requirements. Never mind the fact that StabilityAI took some... questionable decisions in the direction of SD 2.0: they removed the ability to prompt for artists art styles, celebrities and nudity, which the community did not take lightly. They later tried to circumvent the backlash by reintroducing nudity to SD 2.1, but that was too little; too late. All in all, I find 1.5 amazing for illustrations, paintings, drawings, anime, because you can prompt for an artist and have the art in their style (like the infamous Greg Rutkowski), but for photo-realistic outputs, everything looks kind of fake, and this mix is no exception. Meanwhile, I find 2.1 awful for art/illustration, but infinitely better for photo-realism, and its finetunes (SciFi v2.1, Illuminati Diffusion) are far more photo-realistic (although these finetunes are all SFW-only).
@opengls thanks for your reply.
@opengls Even 2.1 literally plasts random splotches of color to censor nipples, and good luck having actually nude humans, so nope. And camera models, art styles AND famous people likenesses removed, what's left, landscapes? Nobody cares for AI-created paintings of random grass fields, AI exists to make things you can't do otherwise, fun stuff like picture of US presidents as Warhammer marines or Oscar nominees drawn by Van Gogh, this kind of stuff...
@anonyvpn001962 It is not just about nudity and artist styles, SD 2.x models appear to be way less steerable when using unusually complex prompts (such as verses, songs etc.). Overall, less degrees of (artistic) freedom, and probably more fawning to corporate interests.
@vuzz I've read about it and the reason is simple, they used OpenClip as a base, which works better BUT is limited to open source/non-copyrighted work so say good bye to stuff while trying to clear reputation of being zOMG evil copyright infringers. The weirder choice is the ridiculous NSFW limit. But that is avoided in custom models, the lack of things like modern artist, photographer and device names in dataset is not. And the 1.4/1.5 models of SD, DALL-E 2, Midjourney etc. all use regular Clip thingy, which is suckier but since it's closed source they don't actually know what's inside though peopel figured out by trial and error.
Easy way to check is inputting "photo by [photographer name]" and see if it shows photo typical for them, or photographic equipment, i.e. Steve McCurry will show women in gritty Afghanistan photos. He's in dataset/ "photo by Kim Anderson" shows a photography camera. So the name isn't in the data. Interesting case is "by Annie Leibovitz" shows her face in 1.5, meaning dataset knows her as celebrity and not a photographer.
They ALSO have different "weight" to them so some artists, styles, names, whatever are pulling the picture heavier. There's a huge database of artists people made with how they influence images. But all of that is reverse engineering, we don't have the data CLIP has, we wing it and guess. For 2.0, OpenClip data IS known so devs find it easier to train and to know what's up in general. It SHOULD have been upgrade, but it wasn't.
From what I gather, 2.0 DOES train on artists and photographers, celebrities etc., it just doesn't label them so doesn't "know" that by prompts. I don't even know if it can be fixed with custom datasets.
What I tried to explain in this long ass post is that it's more of a case of "good intentions" paving road to hell than StabilityAI just randomly middle-fingering customers to please corporate overlords. TBH, none of this should be even possible to copyright at all, except for code of programs/methods themselves and stuff like sampling algorithms that can be patented (and has academic documentation for them). After all, AI is learning on things, not copying them. You can't really ban learning, that would open a bad can of worms nobody wants to go into, like can you now persecute artists from learning art they seen in gallery? The argument "but AI fast" doesn't hold legal candle here.
But hey, we've seen worse abuse of law, and I can't blame devs of Stable Diffusion for wanting to be more transparent and placate the idiots in parliaments and various judges, especially after those keep proving themselves being incompetent boomers who don't understand business and copyright 101. See Zuckerberg in American parliament thingy being asked dumbest questions since Rob Halford was trying to exhale backwards in court to prove rock doesn't make kids off selves smh. It's PMRC and Napster stuff all over again.
@Rin_chan Thank you for the elaborate and concise development, much appreciated.
I wasn't thinking about "copyright" when mentioning corporate needs, actually, it transcends AI, we could call it "political" or "ideological", even "philisophical" as in neoplatonism : What I see is a more general pattern of data suppression / misrepresentation / distorsion, repeated each time it doesn't fit some preset narrative. Getting into the details, and examples, would bring us far from the specific question in this thread.
Just to answer your point, Rin_chan, learning suppression, as well as its corollary, burning books, (and sometimes people) have been attempted many times in history, and the worst is that as tactical devices they have worked efficiently in delaying or derailing societal processes.
As a recent example of un-learning, just consider the slow (3 generations) eviction of the study of the ancient authors and "dead languages" from the western curriculum. A whole database of political, religious and behavioral experiences got slowly out of view of the commoner, pushed into the realm of "specialists". The consequences are both tragic... and quite lucrative, so as a rule of thumb you'll meet some people very willing to go that way and open that can of worms.
@Rin_chan According to this paper, OpenCLIP aint that bad at all. https://arxiv.org/pdf/2209.06103v1.pdf
Dwells on the importance of dialing in class information, more than details, this is an useful hint.
"Our analysis shows that: (i) most of the classes in popular zero-shot benchmarks are observed (a
lot) during pre-training; (ii) zero-shot performance mainly comes out of models’ capability of recognizing class labels, whenever they are present in the text, and a significantly lower performing capability of attribute-based zeroshot learning is only observed when class labels are not used; (iii) the number of the attributes used can have a significant effect on performance, and can easily cause a significant performance decrease."
why many peopel use SD1.5 instead of 2.0?
2.0 purged quite a lot of keywords especially those related to specific artists and NSFW content, so while it might be "better" if you're creating what Stability creator wants you to do, it's less than ideal otherwise.
because not all models and extensions can work on 2.0
What the fuck happened to the Publisher? Reddit got deleted.1.4 Version is no longer existant.The Images here are gone also
i create a perfect character but now how do i make more pictures with the same character? i train the machine?
You can either grab the seed from the generation with png analyzer, and continue tweaking (although that will not be completely replicable always, since if you change enough keywords, it will change), or the better option, which is to take it into img2img.
I know you are mad because of fantasy.ai, but rating this model 1/5 stars is just more work for us moderators. Can you please find a different way of venting? We have enough to moderate and don´t need that kind of nonsense on top, thank you!
let people express their mind like a free society.
perhaps you should stop moderating because it's a valid concern. We're not hating personally on the creator and in fact want them to get paid so I don't see why you're blocking all bad reviews. If users are unhappy with the model or its conditions, users are unhappy and by moderating this hard, you're undermining your own credibility.
maybe you can add a warning sign behind the true score for this model saying that the average may not be representative of the model's generation quality as it is also judged by its licensing conditions.
@denisl just take a look at the comments of the 1/5 ratings, it's not related to the model. Nothing gets removed or moderated, all the reviews and comments are still there. The rating is just not counting.
Reddit gonna Reddit. At least have the dignity of leaving commentary in a Comment, not a Review. No need to tank a rating, which should be the quality of a model, not its usage.
@RalFinger it is related to the model, more specifically to the license conditions behind it. By the same logic, in the future, you shouldn't allow to take into account downvotes on some model with some ridiculous license that makes the use of the model tricky. So yeah, I think you should let the rating count.
@denisl What license do you keep talking about?
And how does this license (creativeml-openrail-m) prevent you from using the model?
Or are you talking about some obscure license that doesn't exist here? :)
@denisl It very much depends on what they want here. Allow ratings just to have them? No matter what gets rated on? Or ratings regarding only the quality of the model....I very much would also want to have ratings for the modelquality but on the other hand ratings dont matter that much anyway
@Godbearmax the rating system will get an overhaul within the next weeks. If you like to leave feedback or how to improve the site, please use https://civitai.canny.io/feature-requests
Thank you for keeping this model free, please don't change! Love you!
Thank you :)
This model have been and will be free.
I already have version 2.0, but it's in the process of testing now. I will be able to release this model here soon.
Have a great day.
@SG_161222 А есть в ней nsfw фильтр или нет ?
@RoninRex Realistic Vision 2.0 основана на SD1.5, так что nsfw фильтра нет :)
@SG_161222 а, я думал основа 2.0) из улучшений просто больше база фото на которых была тренировка?
@RoninRex Я пока не могу позволить себе тренировку. Realistic Vision V2.0 это смесь Realistic Vision V1.3 и Realistic Vision V1.4 :)
fuck fantasy.ai
Thank you for the pruned version of the model! Now I can keep more ControlNet models installed on Paperspace.
What is the difference between the normal model and the in-painting model? I am very interested in knowing how to train an optimized model for in-painting.
Hi! The inpainting version of the model handles inpaint and outpaint operations much better. The inpainting model is not trainable as far as I know.
I think you need to train a normal model and then make an inpainting version out of it. There is a guide on this topic: https://www.reddit.com/r/sdforall/comments/zyieht/how_to_turn_any_model_into_an_inpainting_model/
I'm confused.. license declared here says one thing.. and then there's a link to the stable-diffusion-license in there too... which one is the real one? is there a downloadable pdf with the license to study?
Thanks!
The real license is "creativeml-openrail-m"
Why when I use this model to train my face or another image the file gets 1.99gb instead of getting bigger? Also, I can't train two faces in my model at different stages. why would it be? I use FastDreambooth
What's the goal/improvements of V1.4?
The goal was to improve skin texture. But version 2.0 is coming out soon (this Sunday). There are more improvements compared to version 1.3. You can see it here.
@SG_161222 show samples of hands & fingers please
@SG_161222 Excellent work! The skin is definitely more realistic and less airbrushed
Details
Files
realisticVisionV60B1_v13-inpainting.safetensors
Mirrors
realisticVisionV60B1_v13-inpainting.safetensors
4201_realisticVisionV20_v13-inpainting.safetensors
realisticVisionV60B1_v13-inpainting.safetensors
realisticVisionV13_v13-inpainting.safetensors
realisticVisionV60B1_v13-inpainting.safetensors
Realistic_Vision_V1.3-inpainting.safetensors
realisticVisionV13_v13-inpainting.safetensors
Realistic_Vision_V1.3-inpainting.safetensors
realisticVisionV13_v13-inpainting.safetensors
realisticVision_v13-inpainting.safetensors
31_realisticVisionV20_v13-inpainting.safetensors
realisticVisionV20_v13-inpainting.safetensors
realisticVisionV13_v13-inpainting.safetensors