Which Should You Choose? Stable Diffusion 1.5 or SDXL?

Are you curious about the differences between Stable Diffusion 1.5 and SDXL? Look no further! Join us on Playground as we unravel these two versions. From the significance of Stable Diffusion as a foundational model to the unique features of the newer XL release, I'll cover some main points in this video. Don't forget to subscribe for more exciting content and updates!
0:00 Stable diffusion models and resolutions
1:10 SD1.5 examples
2:22 SDXL examples
3:00 Negative prompts and filters can help SD1.5
4:25 SDXL examples 2
5:00 SDXL refiner
5:51 How do I know what filters do I use?
6:30 Which model do I use?
6:53 Leave me some questions in the comments!
Find us on
Playground AI playground.com
Discord / discord
Twitter / playground_ai

[@playgroundai]

Which model (or filters) do you use most? SD1.5 or SDXL?

[@playgroundai]

@LouisGedo Yo! 🙌👋

[@yongenvincentsoon7471]

SDXL

[@bmazing8987]

SDXL with RealVis or StockPhoto, sometimes Dreamshaper. But maybe that's because of the target Pictures I'm Just working on.

[@GryphonDes]

SDXL is my goto

[@playgroundai]

@@yongenvincentsoon7471 pure prompt or you like to use filters?

[@scooterdooter]

No questions, just want to thank you for letting us play in the Playground! Amazing work!

[@playgroundai]

You're welcome! Glad you are enjoying playground and we appreciate the support!

[@FosTv23]

Thank you for the valuable information always. I love playground ai

[@playgroundai]

Our pleasure! Really appreciate the feedback and support!

[@dah-veed]

Always good videos. But I would love more content. I love playground AI ❤

[@playgroundai]

More to come my friend! Anything in particular you want me to cover?

[@hansscharroo5228]

I use 1.5 at this moment but am wondering if you can train models and styles with SDXL as easy as with 1.5.

[@theg4925]

Nice info, was enough deep to get started om the topic for any non-tech savvy! Mt personal fav is SDXL, for it requires less prompting to give good results. One more thing is that the XL images really look like they are animated or virtual, while 1.6 tries to male it real life like (idk, thoughts are welcomed). Since I use PLAYGROUND for my AI painting hobby, I prefer using SDXL!

[@playgroundai]

I'm glad to hear that! I didn't want to go to deep but give you enough info to understand and get going. You're correct SDXL tends to have more of a hyper-realistic look, on if it's slight negatives but can easily be corrected with good prompting and negative prompts. Great eye!

[@wayneout]

What do you think is the best SDXL model checkpoint to use? I use automatic 1111 and comfyui. I am not familiar with program you are using in the video. Could you tell me what it is? Thanks

[@playgroundai]

I can't say there is a best one, Juggernaut and NightVision are great overall models, Starlight is very popular for 2.5 and 3D and so is Zavy Chroma, there are a few we have on playgroundai.com/ that cater to different needs. We're an online service for text to image ai at the moment.

[@0A01amir]

I need both, photo editing + img2img + ipadapter + NSFW stuff + Krita plugin SD 1.5 is the king, high quality generation out of the box + turbo SDXL is the king.

[@yongenvincentsoon7471]

I hope to update more videos. I also want a more in-depth understanding, or is there any place that can be used as an article or other reference for me to deepen my learning about playground ai?

[@playgroundai]

What else would you like me to go in depth on? We're working on updating our FAQ so people could have something in a document form. 👍

[@yongenvincentsoon7471]

Specify a specific angle for a specific item

[@btascalon]

How do i prompt a portrait photo but still retain the look of the person

[@mdtron]

Hey playground AI, how would I use control trait say like the edge to generate based on all of the edges except the face should be generated independently? I could use the inpaint feature to draw frame over the face and generate but how do I keep it consistent? so I can use another photo and retain the same face.

[@playgroundai]

For consistent faces you want to name your character, just some random name. The problem with using control traits like edge since it takes into account the outlines, it will also take those features of the reference image. It can work but it also not work. Pose is good for what you want to do, that way the face can be anything. I Will be doing content on this soon, just waiting for something that I can't really talk about just yet 😬

[@frankmorlock1403]

Very nice and informative. I'm very comfortable with SD1.5 and creating with Board. Perhaps I should be (or will be one day) but for now. I'm not looking for High degrees of Photorealism . I like to make my images look like the works of famous artists like Klimt or Chagall as examples. Certainly you have to watch out for extra fingers or missing appendages and the like but that seems to be improving on all image generators. And also I'm a little more savvy then I was 9 months ago when I started using AI. In many instances I just want a head and shoulders shot of, as an example, Elizabeth Taylor as Cleopatra in the style of Mucha. In the old days I would simply say Liz as Cleopatra, etc. I might get an upper body shot and if you looked closely you'd see she had extra fingers or deformed hands marring an otherwise beautiful image. These day, I would specify a head and shoulders portrait of Liz as Cleo, etc. Which is what I wanted in the old days, too, but I didn't know enough to limit the prompt to a head and shoulders portrait. I think both 1.5 and XL produce about the same quality in the range I am looking for. XL is for the photorealistically inclined.

[@playgroundai]

Aaahhh it's refreshing to see a comment like this Frank! Too many people are quick to judge after playing with playground a few times. Like anything else it takes time and effort to really dial in something you can be proud of. I'm glad you decided to stick with it! There is so much to discover, even for myself but for me, that's the beauty of ai images. You can truly create anything you can imagine. As for the deformities, I will be covering some content on this soon. Unfortunately, generally speaking that is one of the downsides of any ai text to image platform. Starting with a good filter helps along with playing with the prompt guidance, not to mention controlnet. There are many work arounds that can get you better quality hands but it does take some effort. There is a lot of progress compared to a year ago. Appreciate the feedback my friend!

[@frankmorlock1403]

That's true. One additional thing you can do in many instances is simply crop the figure and eliminate hands that are deformed. either at the bottom or on one side. Sometimes an oval frame would do the trick@@playgroundai

[@billjones3720]

Any improvements coming for controlnet? For example, and at the moment, control images have to be the same aspect ratio as the generated images for best results. In Tensor you can generate a square image, and not have to worry that the control image is also square. I think the ai automatically adjusts for this difference. Thanks.

[@playgroundai]

We recently made ControlNet SDXL compatible so now that is done we will be looking at other controlnet features. 👍

[@billjones3720]

@@playgroundai Thanks. 👍

[@sudumal-DeArrow]

I like most SDXL ❤ it gave us image with so many details and quilty ❤

[@playgroundai]

Appreciate the feedback, SDXL is a great model not to mention the supporting filters. 👍

[@Elwaves2925]

Both. They both have pros and cons and I swap between the two depending on exactly what I want. I often try a prompt for one model on the other (adpating it if needed), then pick the one that works best. This does tend to be SDXL first though. Also, none of this us with the base models, it's RealisticVision and RealVisXL.

[@playgroundai]

Indeed, personally I use both but lean to SDXL more these days. SD1.5 is slightly better at some artist styles.

[@sikliztailbunch]

I wonder where you can apply custom LoRA´s in this interface? Seems to be a bit behind

[@playgroundai]

Can't use Lora's at the moment unfortunately.

[@SimeonKolev]

I really miss the Analog Diffusion style in SDXL, is there an option that I can use it there now or in the future?

[@playgroundai]

Still there my friend. It’s one of the SD1.5 filters 👍🏼

[@SimeonKolev]

@@playgroundai I know it is working with SD1.5 but I do not like the 1.5 :D it makes too much artefacts even when I use lower resolution and negative prompts.

[@Elwaves2925]

@@SimeonKolev Have you tried upping the steps slightly and lowering the CFG scale? I started getting more artifacting after an update some time ago, so I went from 25 to 30 steps and sometimes drop the CFG from 7 to 5-6. It works for me so maybe trying it if you haven't. This is primarily with RealisticVision. :-)

[@BRAINTEASERS_EN]

I use SDXL. Thank you for your videos. Could you make a video about how to create a consistent character, please? Thanks in advance.

[@playgroundai]

Been meaning to tackle this and to be honest it’s not that straight forward but I will cover it soon!

[@HolisticVibeStudioG]

Hola y Aloha 🥥 I Am so new BASIC tech understanding i am learning everyday I Appreciate this is Awesome, it is Helping My Imagination ITS ON❗ Gracias Grattitude, trying to grow I Will let you know Much Mahaloz 🚀🌱

[@playgroundai]

Awesome! I made this video just for people like you! So glad it’s helpful! Enjoy playground and let us know how else we can help!

[@kiksu1]

Hey I'm wondering why do you call models "filters"? Is there an automatic predefined prompt it adds with each model so the model tries to do a specific style? I've been in a world of SD for 7+ months and I've never seen anything called filters yet so this is a very weird design choice. Otherwise very informative video! 👍🏻

[@playgroundai]

Great question! Technically some of our filters are "models" however in the early days of playground we had "text filters" that created a certain style and look, sort of like presets. But these were not trained models. As the tech advanced and custom models were introduced, we just decided to keep the lingo to not confuse users. So yes "filters" only pertains to Playground, otherwise they are indeed models. 👍

[@tripleheadedmonkey6613]

I see the algorithm brought you here too :D

[@kiksu1]

@@tripleheadedmonkey6613 Oh noes! It knows! 😅

[@toshlyra]

I recently tried out SDXL Turbo and tested it versus epicphotogasm last unicorn which is a 1.5 model. SDXL with 5 steps and cfg 1.5 could generate a nice 1024x1024 image, but the skin looked auto-smoothed just like a cheap attempt to photoshop bad skin. SDXL looks somehow more artsy than anything else. The 1.5 ( epicphotogasm ) i used 30 steps and cfg of 8. It produced a very crisp 1024x1024 image with skin pores detail and skin shine reflections etc etc... i've n idea why but as for now, i like 1.5 better.

[@playgroundai]

SD1.5 models still do people better in regards to photorealism but SDXL is getting there. We'll be bringing in some other models soon that do a bit better with photorealism.

[@NEXUS-401]

Thanks

[@playgroundai]

Welcome!

[@WolfFeX]

i like them both, XL is more detailed. Would we be able to upload our own models in the future?

[@playgroundai]

I think both have their strengths for sure and I still use both as well but I do lean towards SDXL more these days. As for uploading models, can't say that we have plans but who knows in the future. Do you mean your own trained models?

[@WolfFeX]

@@playgroundai yup! trained models or ones that are already out there. Maybe could be a feature only for the users with the pro or highest paid plan?

[@playgroundai]

We have some of the popular ones out there like Juggernaugt, Starlight, Niji, Copax Timeless etc.

[@quercus3290]

if you have 12gig or more vram use XL(at a push with 8), if not use 1.5

[@playgroundai]

Doesn’t apply with playground as it’s online. 👍🏼

[@quercus3290]

@@playgroundai I mean for people who use local.

[@rionix88]

I like both but 1.5 is faster and has more customization, control net, addon, etc

[@playgroundai]

Actually SDXL is now compatible with ControlNet 👍

[@Steamrick]

Why use 'filters' as a pseudonym for community-trained custom models?

[@playgroundai]

In the early days of playground we didn't have any custom fine tuned models, instead we had text filters that were basically pre defined styled prompts. These were called filters. But as the models progressed and we started to bring in fine tuned models, we just kept the "filter term" since we still use the text based filters along with fine tunes.

[@inforamaAI]

Is really pitty to see SD1.5 go after the 28th of Feb. I find filters like RPG, Protogen and Relistic Vision giving amazing results when comes to photorealism. SDXL always tries to over enhance results by giving a more digital/plastic feeling to my art generation. I still hope that some legacy version of Playground will be preserved for users who want to keep using SD1.5

[@playgroundai]

I totally understand and personally I still use SD1.5 for more photorealism. It has been a difficult decision but in order for us to focus on better and greater things the decision has been made. I do encourage you to send your feedback to the team at [e-mail]. We've gotten a few messages regarding this.

[@inforamaAI]

@@playgroundai Thank you for the reply. I did already message there as well in regards to this plus another query, hopefully will get an answer soon. :)

[@inforamaAI]

@@playgroundai May I ask something else please. Do you know where I can find the exact filters that Playground had? For example, I was using another AI image generation website from time to time which had RPG4 filter incorporated to it, but the results were not the same as Playground. If I use the filter via any SD interface like A1111 will the results will be the same or the filters in Playground have been tweaked

[@AlexUsername]

For sd 1.5 - every model is refiner model. 😜

[@playgroundai]

Fair enough 😊

[@blender_wiki]

Why choosing? What a silly question. Use both . 🤷🏿‍♀️🤷🏿‍♀️🤷🏿‍♀️🤦🏿‍♀️

[@playgroundai]

You need to know the difference before choosing, not everyone knows the difference.