@karldonitz28599 gpt wrote this didn't it

The styles could be due to the low CFG used normally for Lightning, so it doesn't adhere that much to the prompting style

@LiteSoulHD THX~

Use higher resolution like 1024 x 1024.

@vapochill I use them already, but when upscaling, the result is not as good as this at all

@Helpmepls Use a good upscaler like 4x ultrasharp. Max upscale 1.5-2x for best results, steps the same as sampler or half.

Thats an issue with Forge though so you have to report it there

@HanaShiina thx for your answer! I guess so! so sad :(:(

@HanaShiina how do I report this case?

If you are using png info to extract a prompt you were using in A1111, it will transfer a lot of information over, including things not present in Forge. I recently encountered the same error, and it was because one of the ADetailer models it was trying to load through png info wasn't in my Forge install directory. You can verify this by simply copying the text you used in the prompt, negative and positive, and try running it without moving all the other settings.

you don't have the upscaler that is being pulled over from the prompt in the high res fix. change to an upscale you have and that error should go away

buenas SG_161222, Podrías contactarme por privado nos gastaría comentarle algo sobre sus modelos XL.

o email. [email protected].

gracias

@SG_161222 sorry i am getting this error
PS D:\stable-diffusion-portable-main> git apply C:\Users\s\OneDrive\Desktop\14390.diff

error: configs/sd_xl_inpaint.yaml: already exists in working directory

error: patch failed: modules/processing.py:106

error: modules/processing.py: patch does not apply

error: patch failed: modules/sd_models_config.py:15

error: modules/sd_models_config.py: patch does not apply

error: patch failed: modules/sd_models_xl.py:34

error: modules/sd_models_xl.py: patch does not apply

Ur gpu is bad.

Turn off live previews if you have not already. I still get the error sometimes, but its less frequent.

@UMCEKO atleast comment something useful ,we can troll some other day , i don't think an rtx 4050 6gb is that bad , even if that's the case then they should've atleast mentioned the required VRAM size

@lukesnell ok i'll try it , or should i try to update torch version , also i am thinking of switching forged version as i am still using automatic1111

forgot to mention i am using replacer , with groundingDino_SwinT , and SAM model SAM_vit_h , is this causing problem ? cause i am changing wardrobe or removing it i wanna see what it's results are ? previously i was using epicphotogasm_zinpainting.safetensors , was making great results but sometimes , poor results too

@akshaygodofwar413 definitely update torch version, you can also try using these in the command line if you have not already,

Nvidia (12GB+ VRAM): Use --opt-sdp-attention or --xformers.

Nvidia (8GB VRAM): Combine --opt-sdp-attention or --xformers with --medvram

Nvidia (4GB VRAM): Combine --opt-sdp-attention or --xformers with --lowvram.

@lukesnell Can you please detail what those do ?

@gladstonepro562979 https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Optimizations

@akshaygodofwar413 Sorry mate I was high when I wrote that.

@UMCEKO it's okay , understandable , no offense

most of the time I use 832*1216

You need to use a program like a1111, comfyui, sdnext, or sd forge (or similar).

@MysticDaedra Thanks a lot!

If you're unsure about how to install various software, Pinokio is an excellent resource. It has one-click installations for a wide range of new AI tools, simplifying the process. Just make sure you have a PC equipped with a pretty decent graphics card. A1111 and forge are great for newbies and if you really want to get into it then you can use comfyui which in my opinion has the highest learning curve

Getting the lighting down is first thing before thinking about details/texture. Have you played with that?

D/L the latest version, 4.0. You can choose between Lightning, Turbo, or normal. The first 2 have faster inference, but perhaps the quality lacks a bit. Either would do, imo.

there's some good short hair loras. or start with a SD15 image and /remix to SDXL

@MorganFreeman Thanks, I found a lora for my needs

@niper53 can you tell Wich one please

@kd231002573 This one: https://civitai.com/models/78661/ashortcut

blonde pixie cut，try it

@akuwa666 Do you mean a lora named 'blonde pixie cut' or do you mean just use the keyword 'blonde pixie cut'?

@niper53 keywords，gpt give to me

@SG_161222 Thank you very much for your answer

@simpson33 Did you use it commercially?

how do you even set a negative clip skip? Im in stable diffusion forge and it can only go down to 1.

@wesadgveagea I don't know 'forge' but I think I read somewhere that by default all models (checkpoints and loras) are using Clip Skip: -1 by default and in ComfyUI it uses Negative -1 and it's the minimum while in Forge can use it as a positive 1 - which would be still the same thing but represented by the software differently (I think). But some Models require Clip Skip: -2 to work properly. Also from my experience anime/cartoon models are using Clip Skip: -2.

This is the error that I get if I try to change Clip Skip in ComfyUI to a Positive Integer:

"Prompt outputs failed validation CLIPSetLastLayer: - Value 1 bigger than max of -1: stop_at_clip_layer"

I think in ComfyUI -2 translates as +2 in Forge.

@EROS1CA yes

May I ask how you found out about CLIP skip? I compared -1 & -2, but -1 seems a little better.

@CHNtentes With the ComfyUI 'default' workflow I gen an image, then load a "CLIP Set Last Layer" node set it to :-1 make new generation (same fixed settings, steps: 4, cfg: 2, dpmpp_sde karras, same seed) and I get different result but when I change the CLIP to :-2 I get the same 1:1 image as the 'default' workflow (with no 'CLIP Set Last Layer' node at all). I also find -1 gens to have issues in the image while in -2 I don't see those. Correct me if I'm wrong. But doesn't the lack of "CLIP Set Last Layer" implies you're using the default CLIP Skip by the model/checkpoint?

It's an XL model, that makes no sense

@diffusionfanatic1173 what "makes no sense"?

Thanksssss

BakedVAE means it doesn't need extra SD VAE. Therfore "None" is best. In this case "Automatic" would also work because there is no VAE file with the same name and it would do nothing, which is the same as "none".

change your sampler to ddpm with cfg 2 and it will generate nice pictures

@hamid29365494 already did that but to no avail, i redownloaded and now it works. dont know what went wrong

@jAC3KMdz3M5H Most of the results my prompts produce are full body views. I think, it is because I usually specify the shoes, that the girl in my image wears. I don't know, what kind of NSFW you want to achieve, but my images are all NSFW. But I want erotic, not sexual acts.

@niper53  okay, but "full body shot", "wide shot", etc. works in other models

@niper53 and regarding NSFW, this model produces scary anatomy. I've solved it by using a NSFW LORA at a low level without using any of the trigger words. Apparently this fills in the gaps of this model

I was able to generate NSFW using in the positive prompt (naked, nipples:1.5)

@nicholasmccabe304 I generate all my images with this model. They are all full body and photorealistic. I think they have realistig faces, see my profile (https://civitai.com/user/niper53 ).

I create for each prompt and seed with 10, 13, 17 and 20 steps, cfg-value 1.5 and sampler is euler_a.

If you are interested in the prompt, you need to download the image and look to Exif-Data (e.g. with ExifTool.exe) because civita.com does not realise generation-param from NMKD SD-Gui.

You can also send me a message, and I send you the prompt of the image

You may consider to use ADetailer. It hugely improves the results for full body arts

@SG_161222 you means Realistic Vision 6.0 B2 is the last update? but RealVisXL & SD3 update continue in future?

@SG_161222 Many thanks, your models are always my choice

@SG_161222 Many thanks for the info. I'm looking forward to the next release of RealVisXL, and SD3 - just have to wait and see when they release it. By the way, need any help testing? (On Automatic1111). Rob

We see that on a lot of SDXL. You'll notice it even more when they are wearing clothing, specially bras or corset tops. I think it's leftover from something in the base model, since it's a somewhat common clothing hemline. It bleeds into unclothed images too. Saw it a lot in the early days across many models. Nowadays, with so many extensively retrained models out there, it's not very common, but I suspect there's a lot of the "early days" still under the hood of this model.

@shapeshifter83 I didn’t know that. I have tried in negative scars or something like that. But no success. Thanks anyway

This is a limitation of diffusers, a library used for generating images with SD models. Other methods use kdiffusion which parses the tokens differently and applies them at different steps, as far as I know.

@MachineMinded thank!

100%

I think it´s quite underrated

is lora needed?

pick v4.0 (it should be already automatically be selected) and then click on the download icon. model should be around 6. something GB

I believe as long as your generator software includes metadata about which checkpoint you used, as long as you make a post anywhere on the site it will cross-post it to the Resources that you used (like LORAs and Checkpoints). Alternatively, you can click the Add Post button on a resource and it will definitely post to that resource.

@RedAstronaut interesting. Thank you. Much of my metadata doesn't auto populate because I change my png files to jpg for the most part.

@MysticMindAi You can configure A1111 to save a JPG automatically and I think that will include the metadata.

imo, dont use restore faces, use ADetailer for webui. if you using fooocus try use paint for detailed face. Its easy to try. There is lot of tutorial on youtube. GL.

More speed for generate. Try sd forge

yes you're right..

I gotta ask... do you need the lora with the lightning?

@zGenMedia Nope, no LORA. The lightning is baked in somehow. Just load up the model and use the recommended settings (DPM++ SDE Karras, 6 steps, CFG 2)

@RedAstronaut Is clipskip -2 on this? Thank you for that information btw. No where does it say anything about that. Thanks.

@zGenMedia I use clip skip 1. The CFG is the Guidance which I set to 1.5 or 2.

@SG_161222 Sounds great! Looking forward to it very much :-)

@SG_161222 Same man cannot wait for this. Made the jump from V2 to V4 and the changes were dramatic, huge improvement. So assuming you will merge all the other checkpoints update the latest versions and improve your own training. V5 has a lot of potential, I am sure.

same here (not everyday) but this model is the best i could find for realism !!! Love it !!!

@SG_161222 BEST. NEWS. EVER! Thanks!

@SG_161222 What do you use for Captioning? are you using GPT-4o captioning or something? I want to make a lora myself and want to know from one of the best models on civitai

You do realize this is a photography-centric checkpoint right?

@penguinsrsocool 
Do you realize this is a photography-centric checkpoint, right?
yes. I know that. But I want to create an image that is not what you think.

@karldonitz28599 use models that cater to hand drawn styles then? why is anyone even agreeing with this comment..? "make my milk more like apple juice, please"

This is an sdxl checkpoint

Make sure you are using the Model: ipadapter faceid plusv2 sdxl

Any 10/10?

@maykerev1812 I suppose each model has its own strength & weakness...

@maykerev1812 nothing model cant be 10/10 friend, because there is no perfect in this world

Yeah they posted info about it but no download.

@zGenMedia where?

@SG_161222 awesome can't wait

@SG_161222 I assume the one below is 5? If so, good eyes, I hope you can improve the entire body from a certain distance but it looks great

@SG_161222 the skin textures looks amazing!

@SG_161222 love the look of v5, do you have an ETA on the model?

@SG_161222 Take your time man, I'm sure it's going to be great, thanks for your work.

https://tenor.com/bcvWF.gif

DO IT!!

CHeck this
https://huggingface.co/SG161222/RealVisXL_V5.0

@ai_dreamer seems very close

@ai_dreamer thanks for that. so v5.0 will not be based on Stable Diffusion 3, and that there will be no SD3 version of RealVisXl as such?

@makaveli_313 sd3 totally sucked people gave up on it a while ago if anything id expect to start seeing FLUX fine tunes soon but the current v5 of RealVisXL is still based on SDXL-1.0

@pigeon_ thank you for answering : )

You are right, but I will still be using the SDXL model, like RealVisXL, since I am not capable of training a LoRA for the Flux AI model. I’m sure most people do not own 24GB of VRAM to train a LoRA for Flux. I have only about 6GB of VRAM on my RTX card, and I was able to train a LoRA using one trainer in just 3 hours. Flux is great, but SDXL models will still be very usable for most people like me, especially those who do not own 24GB of VRAM to train their own LoRA.

@PredictAnything I fully agree. I don't care about SD3 or Flux. Most users don't have a high-end graphics card. And then there's the censorship, even on Flux. I work about 90% with RealVisXL4.0 and can't wait for v5.0.

@favutuqe well yes of course for people who cant afford to run FLUX V5 is gonna be the best option but I can run Flux so I specifically am not excited. I'm not saying that V5 will suck or something RealVisXL is easily the best SDXL fine-tune but since I can run Flux I wish he would make one for Flux that's all I'm saying

Flux is unworkable without a 4090. The time sink is real. If you value time as a commodity and unless you are satisfied with lower quality Schnell than that's on you. But for vast majority of people that actually use Gen AI in a time sensitive setting with lower end GPU, SDXL is still the go to. And like the majority we looking forward to V5 on SDXL.

Looking forward to RealVis V5 on SDXL NOT on Flux. Thanks.

Flux is better than Stable Diffusion for composition and prompt adherence but I'm getting much better realism out of RealVis and it's much faster. My current workflow is to generate in Flux then img2img with RealVis to get something decent. Even though I use Flux, I'm looking forward to RealVis 5. Perhaps RealVis 6 can be a Flux model, but RealVis 5 is still a great idea.

Flux creates plastic skins on people, also sort of too clean and perfect for me.

@3r3de again so am i looking forward to V5 i just want to see a FLUX version developed aswell because I do have the GPU to be able to run FLUX dev

@Spoonman2002 um its called realism lora... fixes that issue easily built it doesn't really have it to begin with 

@pigeon_ Have you actually tried the realism LoRA for Flux? It's nowhere near as good as RealVis.

@jbboncemore2 yes I think its almost as good as RealViz however a full blown RealVis version of Flux would crush easily

Yup it's better overall but RealVisXL makes images in two seconds meanwhile Flux 1 Schnell takes 1-2 minutes

@zayx only if your computer sucks I can make FLUX images in seconds so I personally am wanting a FLUX fine tune of RealVis but that doesn't mean I'm not still excited for SDXL Rewwalvis v5

I think the Flux Hypetrain is overrated. It is annoying to write extremely detailed prompts, and most of the custom nodes are not compatible. Therefore, I prefer SDXL. The quality of Flux is not bad, but with a bit of thought and a good workflow, you can generate images with this model that outperform Flux in terms of realism

Flux is a vastly superior base model, but it currently lacks a ton of knowledge that SDXL checkpoints have. Practically no real people, few copyright characters, few artists, bad at animals, lacks a lot of clothing styles, tends to ignore facial feature and body shape descriptions, etc.
It will be awesome once we get some good Flux finetunes, but that will take time.