No.5392
>>5390Didn't know there was a Monoe LoRa for Pony.
Just found out there's a Monoko one too. I wonder how horryfying it would look with Pony Realism.
No.5393
>>5392> I wonder how horryfying it would look with Pony Realism.Give it a try and show us the results.
Also, two more images I made earlier and another two with an addition of some
questionable artist's style LORA.
No.5394
>>5393Hum, doesn't seem to be compatible, guess it was developped on the SD 1.5 Pony version.
No.5396
>>5394Try to include something about realism in the prompt maybe? This might work, I've started to mess with Pony Diffusion just a few days ago.
Currently upscaling this one.
No.5397
>>5396The issue isn't the prompt, A1111 doesn't even see the LoRa when i load Pony Realism (which is based on SDXL).
I guess it would work with the original Pony models based on SD 1.5 but i don't use them anymore (and it won't achieve the realistic result i'm seeking anyway).
No.5401
"depressed"?
No.5402
>>5400It only takes a dozen of seconds on my RTX 3090 even using hi res fix with the same sampling method. I was thinking about a lack of memory but CUDA should throw an error if it runs out if VRAM… Or maybe it falls back on the CPU ?
No.5404
Ok, so Waifu Diffusion weigth around 3.5GB in full precision while Pony XL is around 6.5GB, this hits quite hard on a 8GB card. Try using the –medvram flag if you're running A1111, it won't do any miracle but it should help.
No.5425
>>5423
Mixing LoRas feels like playing mad scientist.
No.5430
>>5428
Why does it look like…
Looks suspiciously like…
No.5433
>>5430That's exactly what you think it is.
No.5434
What has science done
No.5448
>but i feel depressed just by looking at this one.
seriously what did OP mean by this?
No.5449
>>5448He wishes Mado were real
No.5450
>>5448Its realism and eeriness conveys the feeling of hopelessness and depression feeled by madotsuki.
No.5455
>>5453Yes, but it's tricky, uncertain when you want multiple characters and it's really hard to mix different characters LoRas.
No.5462
>>5461Well, SDXL based models have tons of LoRa, are versatiles and still quite capables (indeed Pony is pretty good), but diffuser-transformers are a big step forward.
No.5472
im gonna kam to these pictures
No.5473
I'm going to train a monoe LoRa for Flux (hoping my card has enough VRAM), I'll upload it on Civit.AI when done.
No.5474
>>5473So far, so good.
First picture is prompt at start of training, second after ~800 steps.
No.5552
>>5551Jesus christ you should all be shot. Go back to twitter if you want to soy out over your faggy algorithms
(USER WAS FED TO AI MODEL FOR THIS POST) No.5561
>>5551ComfyUI is harder to master, but its shit tons of available nodes and the control it gave you over the models is unparalelled.
SD webui is perfectly fine for generating "simple" pictures, but once you step into complex productions, this is another story.
There's a lot of workflows available on the net you can use as bases.
Anyway, happy new year, hope this year we'll get even more tensor power and VRAM.
No.5577
>>5573It's another model based on SDXL and yes, Illustrious is quickly taking over Pony, a shit ton of LoRa is already available for it.
Personally, I've switched to Flux, Diffusion Transformers are way more powerful than U-net based models.
No.5578
>>5577Not like you need a LoRa anyway, at least for art styles. HassakuXL (the one I use) can imitate art styles of pretty much any somewhat popular artist.
> FluxIt still can't do anime style, isn't it?
No.5579
>>5578I was more referring to characters LoRa.
And Flux can do anime style image/characters without any problem, I've started training a few LoRas for specific characters (like the Monoe and Monoko ones I've posted earlier).
There's also some specific checkpoints of F1 dev that have been trained to produce only anime/manga pictures, pretty much like Pony and Illustrious.
No.5580
>>5579I've had no need of characters LoRa either, as the model I use recognizes all the (somewhat) popular characters too.
Good to know Flux can do anime style, but is it still requires more than 8 gigabytes of VRAM? If that's the case - guess I'll stick with Illustrious for now.
No.5581
>>5580Q4 quantized version of Flux Dev only needs around ~6.8GB of VRAM to run and produces pretty good results.
I've tested it on a mobile RTX 4060.
