Oh hey, a newbie just caught the AI image bug

[cerberus]

The SD 1.5 models are the only ones ready to generate content relevant to our interests out of the box, yet I'm entranced by the potential of XL and Flux assuming I can find the right LoRAs to turn off the built-in censor. Already signed up for CivitAI cause I figure I'll be spending a lot of time there.

While the bulk of my work this year was done on 1.5, I currently use XL.

Yep, everything on my two main threads and my DA account is XL. I started out with 1.5 but switched over about April. I was happy with 1.5 when I was using it, but I wouldn't want to go back now. XL is a good step up in quality and, with a good upscaler (I use 4x-UltraSharp), can produce excellent hires images. On the negative side, there are less resources, it's a bit less responsive to prompt terms and weighting, it eats graphics cards for breakfast and it's got weird sense of humour.

[MadMax359]

my Fave is in the hallway, she needs all the space she can find

Always nice to meet another enjoyer of the well-endowed

very much!!

[Sekani]

Small update (no pics), been goofing off with Flux.1. Unfortunately I don't exactly have a turbocharged PC (or actually a PC at all, I'm using an M2 iPad), so I can't do much with the higher-quality models. The schnell version actually showed a ton of promise; it kicked out half-decent images fairly quickly, and the prompt engine is magnificent in a way that anyone that's ever tried to specify a sinking depth can truly appreciate. Only issue is that for the life of me I can't get the fucking thing to create mud that doesn't look like river water. I'm probably gonna spend some time with SD 3.5 now that it's out and see if I can get better quality at the cost of a more archaic prompt system. I'm slightly concerned about bumping down to SD XL since I know that this tablet may not have the muscle to give better results than what Perchance was doing.

[Sekani]

Finally made something respectable after a week of futzing around. I may actually be getting the hang of this, dare I say

[cerberus]

Small update (no pics), been goofing off with Flux.1. Unfortunately I don't exactly have a turbocharged PC (or actually a PC at all, I'm using an M2 iPad), so I can't do much with the higher-quality models. The schnell version actually showed a ton of promise; it kicked out half-decent images fairly quickly, and the prompt engine is magnificent in a way that anyone that's ever tried to specify a sinking depth can truly appreciate. Only issue is that for the life of me I can't get the fucking thing to create mud that doesn't look like river water. I'm probably gonna spend some time with SD 3.5 now that it's out and see if I can get better quality at the cost of a more archaic prompt system. I'm slightly concerned about bumping down to SD XL since I know that this tablet may not have the muscle to give better results than what Perchance was doing.

I have to admit I winced when you said you were using Stable Diffusion on an iPad and were considering SDXL etc. My own experience, on a PC with 12GB of video RAM, was that that was OK for SD 1.5 as long as you didn't overdo the image size or run large batches. It used all 12GB and more if I used Reactor or ControlNet etc but generally coped. Currently, I'm using SDXL but am avoiding Reactor and ControlNet, I have a max image resolution of UHD (4K) before image generation fails with insufficient memory (12GB on the video card plus it'll additionally take up to 15GB normal RAM). I don't know what you'd need to run Flux or SD3.5, but comparing checkpoint model sizes - SD1.5 2.25GB, SDXL 6.5GB , SD3.5 13.9GB and Flux.1 22.3GB, I'm pretty sure my PC doesn't stand a chance of running anything above SDXL.

Getting rid of the water effect varies from checkpoint model to model. Putting, "water" in the negative prompt helps in some cases (I'm assuming you've tried this), increasing it's weighting can also help (e.g. "(water:1.5)" may work), it's less effective with SDXL though.

Finally made something respectable after a week of futzing around. I may actually be getting the hang of this, dare I say

What did you use to get this? It's a credible start and you've obviously got past the censorship issue in Perchance.

[Sekani]

I have to admit I winced when you said you were using Stable Diffusion on an iPad and were considering SDXL etc. My own experience, on a PC with 12GB of video RAM, was that that was OK for SD 1.5 as long as you didn't overdo the image size or run large batches. It used all 12GB and more if I used Reactor or ControlNet etc but generally coped. Currently, I'm using SDXL but am avoiding Reactor and ControlNet, I have a max image resolution of UHD (4K) before image generation fails with insufficient memory (12GB on the video card plus it'll additionally take up to 15GB normal RAM). I don't know what you'd need to run Flux or SD3.5, but comparing checkpoint model sizes - SD1.5 2.25GB, SDXL 6.5GB , SD3.5 13.9GB and Flux.1 22.3GB, I'm pretty sure my PC doesn't stand a chance of running anything above SDXL.

Actually managed to get Flux, 3.5, and XL running on this iPad. It's a newer model with Apple Intelligence on-device and they're not horrible. Flux is still pretty slow even with a Hyper-mod, But I can generate a batch of 4-8 pictures on 3.5 in about three minutes and XL in about 55 seconds. Haven't tried 1.5 yet.

Getting rid of the water effect varies from checkpoint model to model. Putting, "water" in the negative prompt helps in some cases (I'm assuming you've tried this), increasing it's weighting can also help (e.g. "(water:1.5)" may work), it's less effective with SDXL though.

That was definitely the cheat code for 1.5 Pretty inconsistent with XL and does nothing yet in 3.5. Flux doesn't accept negative prompts and it doesn't seem to want to accept any of my creative combinations of words so I might end up training my own LORA just to get a decent mud texture. Le sigh.

What did you use to get this? It's a credible start and you've obviously got past the censorship issue in Perchance.

This was combination of Flux for image generation ("A cinematic image of a topless woman with large breasts partially submerged in mud, with the bottom of her breasts touching the surface and more mud splattered on her body") and XL for refining the image with the "no water" trick. Still not very consistent, I actually feel like I got really lucky that XL didn't fuck up the hands on this one.

[cerberus]

I have to admit I winced when you said you were using Stable Diffusion on an iPad and were considering SDXL etc. My own experience, on a PC with 12GB of video RAM, was that that was OK for SD 1.5 as long as you didn't overdo the image size or run large batches. It used all 12GB and more if I used Reactor or ControlNet etc but generally coped. Currently, I'm using SDXL but am avoiding Reactor and ControlNet, I have a max image resolution of UHD (4K) before image generation fails with insufficient memory (12GB on the video card plus it'll additionally take up to 15GB normal RAM). I don't know what you'd need to run Flux or SD3.5, but comparing checkpoint model sizes - SD1.5 2.25GB, SDXL 6.5GB , SD3.5 13.9GB and Flux.1 22.3GB, I'm pretty sure my PC doesn't stand a chance of running anything above SDXL.

Actually managed to get Flux, 3.5, and XL running on this iPad. It's a newer model with Apple Intelligence on-device and they're not horrible. Flux is still pretty slow even with a Hyper-mod, But I can generate a batch of 4-8 pictures on 3.5 in about three minutes and XL in about 55 seconds. Haven't tried 1.5 yet.

Getting rid of the water effect varies from checkpoint model to model. Putting, "water" in the negative prompt helps in some cases (I'm assuming you've tried this), increasing it's weighting can also help (e.g. "(water:1.5)" may work), it's less effective with SDXL though.

That was definitely the cheat code for 1.5 Pretty inconsistent with XL and does nothing yet in 3.5. Flux doesn't accept negative prompts and it doesn't seem to want to accept any of my creative combinations of words so I might end up training my own LORA just to get a decent mud texture. Le sigh.

What did you use to get this? It's a credible start and you've obviously got past the censorship issue in Perchance.

This was combination of Flux for image generation ("A cinematic image of a topless woman with large breasts partially submerged in mud, with the bottom of her breasts touching the surface and more mud splattered on her body") and XL for refining the image with the "no water" trick. Still not very consistent, I actually feel like I got really lucky that XL didn't fuck up the hands on this one.

I'm quite impressed that an iPad was able to run SD, let alone Flux and 3.5. Personally, I'd stick to SDXL, there isn't a lot available in terms of resources for Flux and 3.5 at the moment.

The weighting still has an effect, it's much reduced, I still use it and sometimes it helps enough. I use terms like "drained lake" and "dried up river bed" along with "estuary mudflats" to indicate muddy with an absence of water, these have worked OK for me. With respect to LORAs, I've found them to be a bit flaky and inconsistent; if they work, they tend to result in a drop in image quality in a lot of cases. Try a few different checkpoints models, some are less prone to watery settings than others. The "hand" problem varies with models as well, some are better than others. Apart from that, as you said earlier, it can be best to stop trying to get something specific and let the AI see what it can come up with.

[MadMax359]

you guys talk shop, i'll just admire the tan lines

[Sekani]

Starting to look better but not as consistent as I would like

[MadMax359]

well practice makes perfect!