Thank you very much!

I don't know - I don't use ComfyUI or Clip Skip. So whatever would be normal in Forge, I suppose you would have to replicate that. Happy to hear you got things working 👍

@IceB2 Yeah, the only thing I've seen since it's been updated is 'Clip skip SDXL', which is just a toggle, and which I don't use and always had poorer results with/or doesn't do anything.

@IceB2 I assume this is the same as how I was testing it with xyz plot parameter. I still see no, to very very minor, difference between 1 and 2 for example. Tried a couple samplers.

Nothing like clip skip on SD1.5.

@IceB2 They might be using an AI model for translation and it's translating in that British(?) slang lol. I'm also on 8gb vram with everything I do. Should be fine for inpainting. With that said though, I rarely ever use inpainting and try to avoid doing so with that rng mess as much as possible. I opt to manual fix something and then upscale if I edit a generation.

@ocat85 as for ComfyUI, I don't use it so not sure I can be much help. But as you said, you took out some nodes to optimize the generation. Perhaps there is something else you can do to normalize things or add to help with your vram limitations. It'd be slow regardless I imagine as the model is 6+ gb. Forge UI has low vram mode stuff built in and has functioned fairly well for me, but I'm not sure what ComfyUI might have to help with things. If you're adding lora, that will also slow things down. You can try using smaller resolutions too. I've found Euler A/Euler samplers are the fastest compared to others. If you're not running CFG at ~3 or so, try doing so.

No list - just experiment. Whatever is typical for SDXL, example image ratios, etc. Prompt/samplers may effect this as well. For example, I've had some unique stuff on some models at 1600x512 aspect ratios

Base styles, yes. However, you can still create similar styles with prompting (style/artist tags, etc.). Some versions share similar base styling as well.

Seaart is a mess, they do whatever they want and even make accounts for people without their consent. It's ridiculous and having gone through their support avenues in the past was even more of a mess which resulted in nothing. They would remove models and then just have them reappear, it was a joke.

Pixai I try to manage sometimes, and recently got rid of some duplicates there and left one listing of the uploads for the time being due to inquires and use. That one has each version listed on the page now.

Tensor art, as far as I know, requires verification for uploads - which I fully respect and appreciate. I've uploaded some to their site in the past. I could possibly upload Plant Milk there. Hopefully Civitai will have more generation options in the future again though. If there are models uploaded by others on Tensor art now, I guess I'll have to see about that.

My main point with these types of sites are firstly, the mimicing accounts/auto-uploading, and hosting models for profit connected to their on-site currencies. Alongside the fact I do not want to go to 100 different sites and have to sign up to a ton of random services to "claim" a model. That's crazy ridiculous and this scum-like trend/system needs to change.

'View More' is more a generalization for viewing other various models/resources that aren't on Civitai. To be fair, I haven't uploaded a lot there. Especially with SDXL stuff. However, I plan to have version info and whatever else that will be posted there in the future. So perhaps earlier or experimental version of some things.

Which version are you referring to in particular that you like?

Flax was available as a generation option before the Auction changes/additions. How/if the auction stuff would effect that was unknown. You can see about inquiring for some clarity or to see if the model can be enabled this week for their system. In the future, we'll just have to try to keep remembering about the auction system or see if they change things.

A lot of people have shown to like Flax if you ever do try them out - thanks for the feedback 👍

Posted from another reply:

Flax was available as a generation option before the Auction changes/additions. How/if the auction stuff would effect that was unknown. You can see about inquiring for some clarity or to see if the model can be enabled this week for their system. In the future, we'll just have to try to keep remembering about the auction system or see if they change things.

"Hemp milk as a model positions itself somewhere between oat & coconut."

Interesting to hear this, the model stems from Flax Milk and Alfmix_v1 (+ a few other resources). This version has a much more 'drawn' appearance in comparison to coconut imo. Coconut's base styling is closer to semi-realistic/rounder. I agree colors are on the more vibrant side for both however. Oat is much smoother detail wise, but I can see some similarities with some aspects, but not overall. To me, Hemp Milk like a wilder/messier cousin to Flax Milk.

"will generally require a CFG rescaler if you dont want to use Eula or Heunpp2"

Nah, i've never used this feature. I just use normal cfg features. I found Euler to have the best overall balance throughout different prompts in terms of general composition/character styling, detail, etc. Heunpp2 was very similar. Other samplers can also work well. All depends on your preference, your prompting, and what you're going for.

"Does not play well with high Lora use"

I wouldn't recommend high lora use for any model personally, this series or any other. People tend to overuse them quite a bit, especially on generation service websites imo. When stacking them together, I find low to medium values work best. And if you're highlighting 1-2 specific lora models, up to ~7-8 as a better value to retain some flexibility and base model functionality a bit more. But that also depends on how the lora model was trained and what weights are best utilized. For example, some lora I've seen don't do much until 1.3-2 ranges.

"Lots of broken or missing character geometry (hands, legs, toes...etc)

lots of texture merging

requires a ton of wrestling and mucking about to get what you want"

This version is messier, yep. This version is meant to have a sort of 'freestyle' approach with unique creative abilities. So in terms of doing some less dynamic stuff, you'd want to keep that in mind. However, to help shape things and give some structure that may be lacking where you want it; lora, artist tags, character tags, or reshaping your prompt in general so there are no conflicting concepts etc. may help out quite a bit. Also samplers and other parameters are a consideration. As far as inpainting, I never recommend that unless it's something that can be changed pretty easily. Otherwise, inpainting is an rng mess that isn't an efficient use of time imo. If I edit an image, I tend to manually fix something and then upscale the image personally. Much more effective most of the time and gives much more control.

"Almost feels like a small SDXL version of counterfeit"

Not sure what you mean by this tbh.

"Less fixed in a set style"

I somewhat get what you mean here, but also kind of disagree. The style I see per say is the messier line details and such, and I think that is pretty prominent. As far as some of the detailing (patterns, fine metal work, that type of stuff) - the detailing is there, however isn't the sharpest or clearest at times. Still adds some qualities overall though.

"Details subjects quite nicely..., ...Great compositions"

Happy to see this being recognized. This was one of the main factors for uploading this version. You can get some very creative and dynamic compositions/character orientation and angle. Using the same prompt and having a high batch count should output a range of pretty different results and compositions.

Thanks for the feedback man 👌

Thanks man - and yeah if you get too much of an 'overbaked' effect with some things such as details/lines, you could try to lower the step count.

Posted from another reply:

Flax was available as a generation option before the Auction changes/additions. How/if the auction stuff would effect that was unknown. You can see about inquiring for some clarity or to see if the model can be enabled this week for their system. In the future, we'll just have to try to keep remembering about the auction system or see if they change things.

Posted from another reply:

Flax was available as a generation option before the Auction changes/additions. How/if the auction stuff would effect that was unknown. You can see about inquiring for some clarity or to see if the model can be enabled this week for their system. In the future, we'll just have to try to keep remembering about the auction system or see if they change things.

@whoretaker go 'bid' on them then with the 'bid' button

These are checkpoint merges, so whatever the merge parts contain would be additions. Whether that be lora, merges themselves, or other trained models on top of popular base models.

Thanks for the feedback, have you tried the other versions as well? Each is aimed at being separate from the others as different 'flavors'.

@Hanekawa01 Flax is a solid option imo and has seen a lot of use. Balanced with compositions, accuracies, detail, creativity, etc. Each version has their own strengths though. Enjoy man

Thanks for the feedback man, really appreciate seeing and hearing different perspectives and how different versions are being utilized - you've been sharing some solid images👍

These merges are based on illustrious v0.1 and Noobai EPS models primarily. Some v-pred version may be mixed in very minor ways here and there depending on merge parts. As for merged in lora models, this ranges from base xl, pony, illustrious, noob, etc. based trainings. Illustrious 2.0+ are not available for use (2.0 is now), and whether or not 1.0/1.1 are mixed in depends on each merge part. Not every merge part lists what was used, so this is mostly unknown. I suspect however, most of the 'trained' models used in early merges (as well as my own) were trained over base noobai v1.0/1.1 eps and illustrious v0.1 models. I've since carried on the various 'series' and added in some new ones from then until now.

According to 'auction' page, 5 versions of Plant Milk should be available. ~2 hours have passed since. If there's no issue, I would expect they will be available soon. But if not, ask Civitai team if they know what's going on. I wasn't refunded the buzz, and the versions say they weren't refunded.