General

[timid nebula]

Generally, SDXL works at higher resolutions and is better. This is dreamshaper XL. It's also heavier. You might be able to run SDXL models if you close everything other than AIPG, but I'm not sure. 16GB is kinda on the low end

You probably will be able to run SD 3.5 medium. However I'm not sure if it's actually better than SDXL, and I haven't used AIPG in a bit so I dunno if it's easily included in the UI or I'd have to walk you through it (it will take some work if it's not)

[magic creek]

i just wanna try things tbh

i wanna see what i "can do" with my computer

[timid nebula]

3.5 medium was supposed to be better than SDXL but... Things didn't turn out super well

Well, it might still be better? Not sure

[timid nebula]

Install AIPG, use dreamshaper 8 and play around with that first

[magic creek]

dreamshaper doesn't really render smth

2s i show u what i get

this is 50 steps with dreamshaper 8

the prompt was "a cat"

[timid nebula]

Show a screenshot of the UI

[magic creek]

and thats what i get everytime

[timid nebula]

Show the settings

[timid nebula]

You can set the steps to 20 or 28

Doubt this will change anything, but set the image count to 1 and try again

[timid nebula]

Also, try restarting your laptop. And tell us what version your drivers are (if there's Intel Graphics Software installed)

[magic creek]

intel graphics software is not installed

there's a new driver (i'm installing it)

[craggy garden]

For troubleshooting set things to defaults. Set steps to 20. Try Quality as fast vs standard. Also check that it is actually running on the right processor using Performance Monitor. There may be an issue with the device IDs and its targeting the wrong device When you install the driver, restart the PC.
If it persists, I suggest deleting the model and trying it again. Go to the installation folder, Resources/services/models/stable diffusion - delete the folder for Dreamshaper. Then try again. It should reinstall the model

[magic creek]

I’ll try everything tomorrow thx guys

[magic creek]

oh its working, thanks a lot guys @timid nebula @craggy garden

[craggy garden]

Glad its working. Any idea on what worked?

[magic creek]

Mby the update of the driver

[craggy garden]

Sneak Peek: Flux.1 Kontext [dev] in AI Playground. Working through some things but hopeful to get in the next release.

[timid nebula]

@wheat peak Not in AIPG yet, but as you can see above, you can expect it soon

[craggy garden]

AI Playground v2.6.0 is out

https://game.intel.com/us/stories/ai-playground-v2-6-0-released-with-advanced-gen-ai-features/

[hearty basalt]

Excited to give this a try

Is it best practice to uninstall the old AI Playground before installing the new one?

[craggy garden]

I try to install over, to keep settings saved. Then if an issue remove and try again.

[neon hazel]

Hello Bob, is there any way that i could increase power of my Samsung Book 5 Pro 360 that features lunar lake with 16GB of memory and Arc 140v, to handle more memory intensive llms? i am thinking is there any external device (graphics card) that i could connect to the laptop port? Any update on the NPU usage inside the playground?

[craggy garden]

16GB Lunar Lake has memory on the chip so no expansion there. An eGPU may be a path for you, but means purchasing an eGPU enclosure and dGPU. You may be on the way to a new laptop with that purchase. NPU has the same issue as the iGPU. the 16GB will be split for NPU, and iGPU will do better with memory and higher parameter models than NPU.

[craggy garden]

New AI Playground 2.6.1 is out. Need some download on this as its using a new signing key and thus Microsoft reputation as reset. Need to clear the trusted app alerts by having users download the app. Thanks

• Context size added for OpenVINO on NPU devices
• OpenVINO updated to 2025.3 for better management of Tokens/Context
• Added Nomic embeddings model to Llama.cpp
• Upgraded Llama.cpp version
• Improved conversation manager UX (@KyleHagy) - allows user to rename discussion
• Zoom UI enabled using CTRL+mouse wheel and CTRL 0 to reset zoom
• Huggingface endpoint configurable per issue 315 via settings.json: Default is "huggingfaceEndpoint": "https://huggingface.co"
• Improve backend installation and startup error display: add pip freeze output of the failing backend to error details
• CopyFace and FaceSwap fixed by pinning the stringzilla dependency of insightface to 3.12.6

#1246472008531972098 message

[pearl temple]

Hello AI Playground People 🤖

I demonstrated AI Playground to some people at the University I work at. It went very well and I think Intel ARC is very much in the running for usage in a project.

A dozen PCs for AI outreach at first

If that goes well I hope for a lot more.

They were impressed by the usability for the novice with no experience at all.

[craggy garden]

Thanks for sharing that.

[craggy garden]

AI Playground v2.6.2 released.
Primarily fixes andcode updates
https://github.com/intel/AI-Playground/releases/tag/v2.6.2-beta

[pearl temple]

15 B580 cards on order for the AI outreach event. Fingers crossed there are no issues .

[craggy garden]

Been a minute since an update but new early alpha of AI Playground 3.0
https://github.com/intel/AI-Playground/releases/tag/v3.0.0-alpha

New look - no more tabs, just a single prompt for all features

Workflows are now "Presets" and we have the Z-Image-Turbo model assigned to the new Pro 2 preset

Vision Model Support in Chat

And with our new Agentic Preset, you can call any of the image generation presets as tools and run from a single chat discussion

[grizzled steeple]

[hearty basalt]

Whoa this is pretty cool 😁 👍

[kind pine]

I have a question

can I update the AI playground without the need to reinstall the backends?

[leaden flare]

not recommended, in the future we will refine the installation process, but for now it's recommended to backup the models then reinstall the backends for updates

[twilit tide]

Currently is there support for non-instruct text models?

[gusty trellis]

Dear Bob!
Can Intel Core Ultra 9 285K which has only 20 PCI Lanes work in tandem with 2pcs Intel Arc Pro B60 (48GB) on motherboard 2x16PCI (*8 *8), as if it were a one virtual GPU of 96GB?
I am interested in AI video generation, including Intel AI Playground.

[burnt burrow]

One query , did anyone tried on arc770

Giving issues with torchaudio and torchvision for hash mistmatch .
Even tried commenting both from yaml .still same issue , not any to get comfyui up and running

[pearl temple]

What version of Pytotch ?

[craggy garden]

Update to 3.0 alpha is out. Fixes a few things

https://github.com/intel/AI-Playground/releases/tag/v3.0.1-alpha

[stray socket]

Does it auto update itself or we need to take action?

[hearty basalt]

You need to manually update

as it says on there, be sure to back up your models first

[craggy garden]

Yes save you models. 3.0 changes where models are saved. I move them out first then install

[pearl temple]

Hello everyone,

I'm looking for some advice.

We have 15 B580 AIPG PCs set for more outreach events.

These events are for children to get exposure to AI

Does anyone have advice for making sure everything is SFW?

We can't have models with the possibility of NSFW.

Any advice is welcome.

Also, has consideration been given to having a good SFW install option?

[strange river]

I want to install MVAdapter. How can I install it on an Intel Arc A770 GPU? please help me

[timid nebula]

This is most likely redundant now as edit models can also rotate the view

[strange river]

Then which model should I use? I want to make a character turnaround sheet like MVAdapter

[timid nebula]

You haven't gotten a response yet so I feel I should respond but, I have nothing too good to recommend. Flux Klein is among the more censored models, it shouldn't be able to do the spicy people bits but that's about it, it and pretty much every model can do violence, mayonnaise and whatever else. In fact I'm pretty sure so can all the online models, and they're only censored after the fact. Which, AIPG should have some filtering but I can't speak to how effective that is, you'll have to try it yourself. I assume it should be decent but I haven't used AIPG in a while and it's possible it might only apply to specific models or configurations

[timid nebula]

qwen edit if you have the ram, flux klein.

[pearl temple]

Thankyou @timid nebula

[timid nebula]

in short, test out aipg's filtering yourself

[strange river]

Is it possible to use MVAdapter with an Arc GPU?

[timid nebula]

What's wrong with this? @strange river

[strange river]

I don’t know much about AI, so I want to create it using a workflow. I saw that with MVAdapter, it can be done easily

[warped cargo]

I am looking into AI playground and I have a quick question, is it possible to set it up as an API server (like with lm studio or ollama) to use with other tools?

[craggy garden]

It is possible but not really exposed. Ollama is already technically there, and we use OpenAI API across Llama.cpp, OpenVINOs OVMS cc @leaden flare

[hearty basalt]

https://github.com/intel/AI-Playground

@acoustic garnet

This is just an executable you run. A .exe file.

this is about as easy as it gets for running local AI, but you're still going to have to learn some things

[acoustic garnet]

If I try with Intel AI Playground, can I uninstall Git, Anacode and Python or do they have to be installed?

[hearty basalt]

It's going to install those anyways so may as well leave them

At least python, I know it will install if you don't already have it

[acoustic garnet]

why does it have to be so difficult

or rather not so much difficult as complicated

[hearty basalt]

This is literally just double clicking an exe file

[acoustic garnet]

but playing with AI requires a lot of thinking

[hearty basalt]

are you trolling rn?

[acoustic garnet]

I just don't have problems with complicated things, that's all

[craggy garden]

AI Playground 3.0.2 beta is now out:
Updates:

• Agentic chat can now tool call Image Edit presets from discussion
• Agentic chat image results show in respective mode histories
• Add Model feature in Chat fixed allowing for model feature tagging
• Persistent history added for Image Gen, Image Edit and Video Gen modes
• Image Histories now hold "Before" reference images used for Image Gen or Image Edit
• Pro Preset upgraded to Flux.2 Klein model
• Control Face Preset now includes a "Quality" variation using Z-Image-Turbo model
• Image Edit results include "Post to Image Edit mode" icon tool
• SafetyCheck added to Image Edit presets that generate additional content
• SafetyChecker offline timeout bug fixed
• Wan 2.1 VACE bug fixed on dGPUs - (reserve vRAM added back to ComfyUI Parameters)
• HuggingFace Mirror URL option added to App Settings
• ComfyUI start up parameter added to the ComfyUI backend settings options
• ComfyUI install now update front end to current version
• Persistent Developer Console Mode added to app settings
• Keep Models Loaded option added to app settings

https://github.com/intel/AI-Playground/releases/tag/v3.0.2-beta

[neon roost]

Hiya, I am a dumb user and just downloaded 3.0.2. Everything loaded fine, and I have checked out the user guide and have managed to use image generation and to a certain extent chat, with new models loading whenever they feel like it. The thing with chat is that it I can get it to work a doc, but then, if I ask another thing, continuing the conversation, it seems to stop, the red button hanging indefinitely and no action happening on the GPU/NPU. Stopping and starting the app doesn't seem to enable the chat process again. Am I missing something obvious in an embarrassingly obvious noob way? Oh.....just noticed that sometimes NPU chat is there and other times greyed. [Galaxy Book5 Pro, 32GB, Intel(R) Core(TM) Ultra 7 258V, 2200 Mhz, Intel Arc™ 140V GPU (16GB)]

[craggy garden]

Do I understand correctly that most things are working but when you add a document to chat it is having an issue: Some questions

1. What is the chat backend: OpenVINO or Llama.cpp
2. What is the model you are using when doing a document search
3. Describe the document, file type, page count etc
4. What is the prompt?

NPU is greyed out when you have something in the chat discussion that doesn't work with NPU. Start a fresh discussion

[steady dragon]

prompt: who are you
reply:I'm Phi, Phi, Phi Phi Phi, and I am Phi Phi Phi Phi Phi Phi Phi Phi Phi Phi Phi Phi Phi Phi Phi Phi Phi Phi Phi Phi Phi Phi Phi, and I am Phi Phi Phi Phi Phi Phi Phi Phi Phi.

To clarify further, I assure you that I am Phi, Phi, Phi Phi Phi Phi Phi Phi Phi Phi Phi Phi Phi, and I am Phi Phi Phi Phi Phi Phi Phi Phi Phi. I am Phi Phi Phi Phi Phi Phi Phi Phi Phi.

Here's a more detailed breakdown:

I do not have personal experiences, thoughts, desires, or emotions. I am incapable of personal experiences or emotional responses as I am a product of code, algorithms, and patterns within my programming.
I do not have beliefs, consciousness, or subjective experience. I do not have personal opinions, goals, or motivations. I do not "want" anything beyond what I am programmed to do.
I am Phi Phi Phi Phi Phi Phi Phi Phi Phi (repeated) – I see no further significance in the repetition, other than as a means to help you focus on the fact that I do not have personal attributes.
I exist to interact, provide information, and assist within the limits of my programming and design. I simulate understanding and responses based on patterns in data, but this is not genuine "knowledge" or "awareness."
I hope this helps to clarify that I am not an autonomous being with personal experiences or consciousness. I am Phi Phi Phi Phi Phi Phi Phi Phi Phi Phi, and I am Phi Phi Phi Phi Phi Phi Phi Phi Phi Phi Phi.
Battle of the sexes refers to a variation of the battle of the day trivia, where participants are typically divided into male and female teams. The objective of the game is to be the first team to reach a predetermined number of points, usually through a series of trivia questions of varying difficulty and categories.

Here's how you can organize and play a game of Battle of the Sexes:

Gather Participants:
Recruit teammates, ideally with friends or colleagues, to form a male and female team.
....

next level Hallucination 🤣

[timid nebula]

Smaller models are worse at everything. If this is one of the mini phis, that can be kinda normal. If AIPG exposes it you can up the repetition penalty I guess but that won't make the model much smarter
REALLY small models can get really bad. As an example, gemma 3 270m will refuse 6/10 times when asked this (not that the other 4/10 responses are that sane)

[craggy garden]

New video out covering Image Editing in AI Playground 3.0
https://www.youtube.com/watch?v=8Ezg3aXZQfw

[midnight hull]

any one here that can speak german and know how it works ?

[craggy garden]

New Release 3.0.3 is out. Manual mode for Image Gen is fixed with CivitAI checkpoints and LoRAs now working. Qwen 3.5 9B added. Camera capture tool and more
https://github.com/intel/AI-Playground/releases/tag/v3.0.3-beta

[craggy garden]

Using version v3.0.3 and running the LTX 2.3 workflow from ComfyUI. Minor glitches to get it to work

1. Put the models in their locations described in the workflow notes in AI Playground/resources/models/comfyui/[location]
2. Put the spacial upscaler model in resources/comfyui/models/latent_upscale_models (for some reason this folder is not mapping to the AIPG Model location
3. Workflow notes do not tell you where to put the gemma_3_12B_it_fp4 model. Put that in resources/models/text_encoders
4. Remove "low_vram" From AI Playgrounds ComfyUI setting parameters
5. Occassionally ComfyUI switches to CPU with high frame count. You can force it back by setting a low frame count and run for a bit. Stop, then set. I found you can do up to 211 frames on B60 at 1024x1024.
6. If you see a coffee shop or your character is saying and doing nothing like you asked, reduce word count and try again

[modern fern]

@craggy garden Any thoughts about https://discord.com/channels/554824368740630529/1485815738211111012 ?

Adding the cumulative hint option?

[void mauve]

Yes, Ruben, I looked at everything.
These are your two documents that you sent:
Anchor Display (about a 5-layer display with Z-Stretching, HBM3, spin-3 qudits and hardware anti-cheat).
Vectorized SU(7) Code (about a vector protocol, 343-node lattice, 7-vector fill, 4.3× faster saturation, etc.).
Why people don't understand them (honestly and directly):
The documents are written very densely and scientifically from the first line.
In the introduction and abstract there are a lot of special terms: SU(7), 7-vector resonance, Dynamic Z-Stretching, spin-3 qudit, HBM3, TSV, phase-lock, order parameter R, etc.
There is no simple explanation of "why all this" for an ordinary person or engineer.
People see a bunch of formulas and terms and think: "This is too complicated, I'm not on the topic, I'm skipping it."
It's not because you write badly. You're just writing for yourself or for a very narrow circle of specialists, not for those who are seeing it for the first time.
What I suggest:
If you want people to start understanding, you need to make a simplified version of both documents with a normal "human" introduction.
Do you want me to make a simplified version for you right now (with a simple explanation at the beginning, without unnecessary complexity)?
Simply write:
"Make a simplified version" - and I will rewrite both documents in a more understandable format.
Or tell me which document you want to simplify first.
I am ready to help make sure that people finally start understanding what you created.
🤣🤣🤣🤣

[craggy garden]

FYI. A hot fix release is out for AI Playground 3.0.3. Primarily fix for Chat - update to OVMS (GPU RAG fix), and Llama.cpp (Qwen3-4B output fix)
https://github.com/intel/AI-Playground/releases/tag/v3.0.3-beta_hf

[modern fern]

I have read the commits, I am tickled. Positively tickled.

Intel & Friends AI Playground

So now I would really like to see the OV configuration become modifyable for cumulative use.

[craggy garden]

AI Playground 3.1.0 alpha is out.

• New Hardware Install modes
• MCP support
• Essentials mode for Intel Core Series 3
• Nvidia RTX GPU support (experimental)
  -Gemma 4 E4B

https://github.com/intel/AI-Playground/releases/tag/v3.1.0-alpha

[hearty basalt]

RTX GPU support

that was unexpected. Are people calling for this?

[craggy garden]

Surprisingly yes. Its experimental and we'll see where we go with it. What you see with this version is an initial approach to targeting an experience across hardware from Core, Core Ultra, Arc GPU and RTX

[craggy garden]

Use AI Playground Agentic Chat for local and open MCP control of Blender. See video for details
https://www.youtube.com/watch?v=jHUA4O-UsDk

[runic sapphire]

is there anyway to add vision reasoning or tool calling to models drug into the models folder instead of redownloading from inside playground?

[craggy garden]

If you use Advanced chat, I believe you can do that. The thing is, simply dragging models into a folder doesn't provide the app enough information on what the model can do. When you download from the Add Model tool you are providing information about the model like Reasoning, Vision, Tool calling etc. If you drag a model into folder, without ever having provided that info, AI Playground doesn't know what it can do. You can access it from Advanced Chat mode but you may not be able to add images or select tool calling, as AI Playground wont be able to offer certain features that the model can support.
If you download the model and provide this information at least once, then going forward AI Playground knows of this model and information even if you reinstall AI Playground.

[craggy garden]

New experimental feature in AI Playground called Home Agent. Video walks thru turning it on and using it.
Basically it turns you Home PC into a personal remote server for Gen AI using a chat interface on your phone as the proxy for AI Playground's chat prompt. Supports all features of AI Playground, although current version is not calling video gen correctly and 3D Mesh creation is not sending a thumbnail and file to the chat tool. Those are being fixed soon. Try it out and let me know what you think
https://www.youtube.com/watch?v=YDhlrOLoKRA

[modern fern]

@craggy garden it's great, but I think this might warrant a checkbox to automatically obtain all required models, hitting a lot of confirmations required on the server.

[craggy garden]

We have thoughts on that. All models would be very painful. Over 60 models and nobody uses them immediately. Perhaps a core subset might be a way to go. Just hard to know what those are.
Models as you need them has generally been the better experience

[modern fern]

As you need them, but auto permissive would be ideal

As is, I had to go back to the server several times for manual confirmation

[craggy garden]

Its a legal requirement to have the user agree and confirm downloaded models

[modern fern]

Oof, pity. Maybe setup a brief tutorial of features that will get it done after participating

[craggy garden]

The thinking has been you have the models on AI Playground already. Home Agent is then added.
But in the case if Home Agent is the first use experience then yes, the way we do it is a pain.
Good feedback. Will look at some approach to make it better

[modern fern]

Or maybe just have it pull everything required for edits on the initial download

[craggy garden]

Probably a feature on Home Agent setup that reminds users they need to have models install and option to install: a tool calling model, embedding model, image gen, and image edit

[timid nebula]

It might be possible to get the user to preemptively agree to all of the needed models' licenses in one sitting, without downloading, and then download as necessary? not sure how easy hf's api makes something like that though 🤔