🚀HarrlogosXL - Bringing Custom Text Generation to SDXL!🚀

Same for me

change the cfg start at 4 then incrementally go up to 10. worked in one of my use cases

Try using the regional prompter extension for A1111 or the attention couple custom nodes for comfyui, and have different words for the prompts in each of the mask areas you want, I managed to get 4 legible words this way. Also it looks like some phrases work better than their components sometimes for some reason, I couldn't get the words "A Nap" to work, but "Take a nap" worked for me almost every time.

Regional prompter works really well: https://civitai.com/images/3433673 , https://civitai.com/images/3419503 , https://civitai.com/images/3420272

@Dever Thanks

@HarroweD many thanks for such a broad reply

Yes, it's very difficult. Try a comma or even two separate prompts.

@snowcult Thanks, will try that.

@HarroweD I understand. Thank you for making the effort to make it happen.

Must be something wrong with your settings or prompt, as the rest of us have figured it out. What are you prompting it?

@DazMakeArt xx text logo,black. but batch4 comes 3 white word or even all white word

@manboro is that the whole prompt? if you look at the description of this model, you can see that you may provide quite a bit more context for the model to understand what you want. the less it has to guess the less chances it fills it with white or other unwanted aspects. give it a try :)

@eurotaku you right, but simple only word and color do create good image, just the color not so right. i mean balck color

@HarroweD other color should be ok, but black... i just try HarroweD text logo, black. and word color mainly white. should accent, bg color must be put in?

@HarroweD im so hyped for it 😇💸

@HarroweD not sure if you know this or you have a problem with the license but you can allow onsite generation without changing other license conditions :)

@HarroweD just let us use it here 😭

@HarroweD Ahh got it, thank you so much

I'm guessing you use the updated a1111 webui. Have you checked under [settings/extra networks/When adding to prompt, refer to Lora by]. For me this fixed it. Some creators add metadata that changes the lora's name when called in promptbar. Which also causes lora's not to work when copying the prompt of the uploader. Hope this solves your issues!

Cause the creator doesn't allow for onsite generation right now, although he's working on it. If you don't have the computer to run it, there's multiple collabs or services that allow you to.

@DazMakeArt I only make prompts and images trough the website. I do not understand ecerything about AI offline. ( Not using Civitai )

I use to generate things when I'm in bed with my iPad Pro after I'm tired, or when I'm traveling by train (I move around a lot for my job), so the onsite generation offered by CivitAI is a godsend. It would be awesome to get the model be finally usable online for people like me! 🙏

@DazMakeArt which ones

Because this is a Lora XL. It's only going to show with an XL model loaded. If you can't run XL models then don't even bother. Wishing there was a non-XL version myself.

@lolatbannedfornothing138 Thank you for explaining! Had no idea about XL Models.

"big text" "big logo" and alike added to the prompt might help, but, it's more or less luck of the draw.

@HarroweD ok

Have you seen the gallery below that all of the generations are successful ? How possible that you cant get the desired word generated (even 1) in a 200 batch generation ? You should check you settings.. There must be a mistaken technique about your setup.

Trying some parameters, generated 1 in 5 batch. https://i.ibb.co/qxgxjCh/test1.png

I share your frustration. Overall, this Lora is very amazing, but sometimes it's incredibly infuriating. I have noticed that some words it hits better than others. Like even my name "Foxbite" hits maybe 6/10 times, but a friend of mine "Shroom", after like 200 generations, i never hit it 100% even once, even though "Shroom" is probably a more common word. It gave "SHAROM" / "SHOROM" like 150/200 times. I've done all the tricks like putting "SHAROM" in negative, increasing weights, locking seed and adjusting cfg/steps and all that jazz, but it just refused to make "SHROOM".

@HarroweD Ive tried all that sadly, I know its easy to assume users are being a bit silly or dont know what they doing when they have issues, and im FAR from an expert, but I would try all means before id make a post moaning, the lora is great and its fun to play around with, so kudos on your work of course!

Its just extremely hit or miss and miss more often then hit to be honest.

@simartem07 I literally said that some words are a very easy hit, I have plenty of pictures I could post below too that would make the LORA look like its a 100% hit rate to any spectators. But that would be cherry picking which I dont want to do when I have such trouble with alot of letter or word or phrase combos.

And no Im not doing anything wrong that I can see, the fact I can get some things to work so easily (i tested out the words in most of the pictures posted on this page to make sure i wasnt doing something wrong, and nearly all gave a good hit in a few seeds, so I then knew it was just a hugely random lora depending on the words)

And no offense but the picture you posted of LOKIPOKI to try prove a point is an awful picture, you can get the name to appear alright if you put the weight at like 2.0 and dont use a creative CFG at all, but then you might aswell go write the letters in MSpaint and it would be less ugly then the picture you posted. To try get it to look good, and still get a nice hit is where the work comes in.

Some words and phrases work fine and they look really cool and creative, others just Will Not Work, fullstop. Bar losing all creatively and style to force it very hard, which ruins the look, like the one you posted.

Which is very frustrating when trying various words and phrases (Ive noticed anything with repeating letters is the worst offender, even with careful use of negative prompts)

@LokiPoki1 I can understand your frustration about your case, but the image I posted was not about proving it is a masterpiece, but still acceptable when compared to have zero result in a 200 batch. 1/5 in a batch is still acceptable if you had a zero chance before. Most of the things in AI has now trying to develop and we used to be doing these things in 2d, 3d platforms Spending hours and days, before. if you believe for your case that unknown words can not be generated in this lora, why waste lots of time and energy for a 200 batch just use mspaint (that was your opinion that could generate better)if you believe you can still get a better result than a 2.0 weight prompt, it just would be more accurate, 1/5 is still a ℅20 success even if it can be classified ugly. I am not insisting on that, but even with a 2.0 weight prompt, with other dynamic parameters involved, we still will have endless combinations over each seed. For best results with the unknown words thing, xyz script would help a lot, with dynamic cfg, weight prompt and sampler related changes, even the model could be a dynamic parameter in this scenario. sorry for the cherry-pick result :-)

@HarroweD FFFFFFFIIING AWESOME!!!!!!!

@HarroweD if I can figure out how to make it more consistent I might be able to afford a license for it, but for now, its too random to actually use it for work. Right now I am only using it to make team art at work, not company based but more for "our team has a cooler logo than you" sorta stuff.

@r0adkill76 "I have been been using it for work"
"it's too random to use for work"
Have you seen the gallery? Looks like you just don't know how to use it lol

@AthenaShadow you may be right there, feel free to show me some good prompts for it. I am always up for learning from the masters, and I do not pretend to be anything other than a cheap knock off.

@r0adkill76 I'm not really sure how to help you there. The instructions on how to use it are clear as day in the lora description, and there's hundreds of examples in the gallery containing the metadata that you can literally copy and paste and just change the wording on lol

@AthenaShadow Maybe I am using a bad platform/checkpoint, what platform do you use?

@r0adkill76 I use it locally on Stable Diffusion A1111, HarroweD uses ComfyUI. Works great with both and the best results come from the Starlight checkpoint from my experience, though there are many others that it works with as well. You could try posting some of your prompts here and I can let you know what's wrong, as it's probably something to do with your prompt.

PonyDiffusion works very differently, to the point where I'd say it's incompatible.

The recommended checkpoints are StarlightXL, PixelartXL, I've tested it and gotten amazing results with MixtapeXL, SonitusXL and Protovision

@DazMakeArt thanks for the reply. I had figured as much. I'll give those a go and see what I can make.

@HarroweD Thxs !

It struggles with the simple Roman alphabet sometimes, I think there's pretty much no hope for intricate Chinese characters until SD is more precise.

that seems to be the onl;y thing i can generate, jibberish...

There's more than enough information provided to get results. Learn to read.

@ZUSIMO more talking about the fact that its impossible to print text using table diffusion, no methods work, unless ur skitzo

Why? It would defeat the entire purpose for this lora LMAO

@ZUSIMO Indeed, but I like the logos it makes, and sometimes I'd prefer to get just the logo and add the text myself