First successfully trained custom variant of PlayGround V2
Use the training version above in Kohya and it should work for you. But I suggest merging the trained data with the original PlaygroundV2 model, as this model is not pure PG.
https://huggingface.co/playgroundai/playground-v2-1024px-aesthetic/blob/main/LICENSE.md
Recommended to use DPM SDE Karras/Exponential, 50 steps and a CFG of 3. But you can try lower step counts and experiment with CFG and other things. I have tested 4-8 steps and had some success even.
I like to use with a denoise of 0.88 on the Sampler.
A few caveats. It functions in place of an SDXL model, however it does not share compatibility directly with LoRA etc. As far as I have tested. They seem to effect the model differently than intended due to the nature of this model having been trained completely from scratch.
I personally think that this model is going to become the new base model for community work, going forwards, as it is a million times more flexible to train and already higher detail than many custom SDXL models.
Also this is technically an alpha release on my part and doesn't necessarily reflect the full degree of outputs the model is capable of.
Description
FAQ
Comments (21)
It cant do people too good (especially hair, tits are good xd) but all the rest is better than SDXL so I hope people will be pushing this topic. GJ!
On the HF page it says "The model was trained from scratch by the research team at Playground". So, technically it's not SDXL, correct? By the way, civitai now has new tag "Playground v2", it could probably be better to edit "Base Model" tag for clarity.
Awesome, thanks for letting me know. Yes this is not the same as SDXL. same technology/code different weight values. SO it is not cross compatible, but the models otherwise function interchangeably
@Triple_Headed_Monkey No worries. But does "not cross compatible" mean we cannot use SDXL LoRAs and ControlNets with PGv2? Or is it opposite because they "function interchangeably"? Sorry, this stuff is still confusing to me.
@green_anger So it is kind of confusing, but because it is the same technology things will work on a technical level. There should not be any errors. However, due to the nature of all the weight values being different, the expected results from certain tools which have been trained to work with SDXL weights may not occur and instead something else entirely may happen.
If anything, I'm not saying Control Net and LoRA will not work, only that the expected functions will not likely present themselves when used.
@Triple_Headed_Monkey Ah, that may explain why I couldn't get decent results with the model with my standard workflow, it had some LoRAs with low weight that usually improved most of SDXL model's output. I'll have to try this without any extras. Thanks for clarifying this!
I am enjoying this model, well done.
LoRA do affect it differently... but some of the better ones do seem to change things. Everything works differently and produces different results. It's great to have something different to play with! Particularly when so many other models are merging into one inbred amorphous blob of similarity.
I've been told that Wowifier works :D
AdvancedEnhancer at .75 and SDXLHighDetail at .5 are my current favorite loras with this one. Wowifier is also fun, but can be messy above .5. CFG 2-3.8 is the best range for me so far as well. schedular: sgm_uniform. Denoise:1.0
also freeU can mess things up so I removed it from my workflow with this model.
@zaxious Wow I have only add-detail, wowifier and Add Mega Detail - I need to get those 2. Thanks for the tip!
@zaxious I can't find SDXLHighDetail - just one mention of it inside an app on huggingface.
Can you please enable the model to allow it to be used for civitai's online image generation? Thank you for your consideration.
Yeah sure man. Sorry about that. I thought it was enabled.
@Triple_Headed_Monkey Thank you for your reply.
I just realized that your model is based on Playground V2, which is not SDXL, so civitai's online generator probably does not support models fined-tuned on Playground V2 yet. My apologies 🙏
@NoMansLand If not, perhaps in the future it will, as this may become the new defacto model format.
@Vendaciousness Yes, I hope the generator will support PGV2 based models in the future
There is no physical reason why they can't do so. They choose not to as part of the license agreement makes it so that they may have to seek business arrangements with PlayGround in the event that their models are generating revenue on the site.
this is the second model of yours i've gotten and nothing is in focus. Some of it is badly out of focus. Like the images were really small, enlarged then blurred to mask noise and blocks, and that was sampled. Right now everything I sent through is 100% unusable. And its too bad too because it looks different than the other ones I use. Weirdly some of its is a bit clear, but then other parts are blurry again.
This is very strange. Clearly I don't have the same issues running the model. I use BF16 precision with FP16 VAE in combination. Try this and see if it helps.
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.