set your filter in civitai for comfyui. Save a picture and drag it into comfyui. Many people leave the metadata in the pic so it will populate the workflow in comfyui. If you open the picture in windows you can look at the comments field. The comments normally store the workflow.

Use forge UI is my recommendation.

@bobby888 That is what I use as well. Waaaay easier and faster and you can get most of the information out of the comfyui comments/metadata and piece it together. Get the png info plugin and you can drop the picture in it but the metadata format for comfyui and forge ui is different so you will have to piece it together. comfyui metadata usually just gives the modelid instead of the name. you can use the civbrowser plugin to figure out what those are. good luck!!

@b0bsm1th Thank you! I'll give this a try.

@rfsecondary883 my pleasure. Let me know if you get stuck or need some other info. There have been quite a few people who have helped me and I want to pay it forward. Cheers!

SwarmUI is best / easiest

@Cyberdelia I use a lot of things, but I'm moving more & more over to SwarmUI. It's biggest advantage is that it has a fully working ComyfiUI backend which you can switch over to in 5 seconds (or without even firing Swarm up). & SwarmUI gets the latest/most updates to support new things/optimisations.

Personally, I'm not a fan of all this workflow/node bullshit lol. It's just too time consuming, & clunky most of the time (especially if you need to edit someone else's spaghetti workflow). With SwarmUI, you have the added control of that, or if you don't need it, & just want to change settings quickly, then you have the Forge like UI frontend of SwarmUI you can use instead.

https://github.com/mcmonkeyprojects/SwarmUI If anyone is curious; scroll down & "Download The Install-Windows.bat file". You basically just double click that, & it auto-downloads everything you need in 10mins or less. It's the easiest setup ever. For n00bs; It'll also automatically auto-download any clips or VAE you need for checkpoints if you have the wrong one, or none at all.

@Cyberdelia Yes! Easier and better UI! As simple as you want to it to be with flexibility for advanced customization (and even comfyui view if you really want, thoI just use comfyui for that hah). The project is also actively updated with new features/improvements. Highly recommend!

invoke is your pro tool

@Cyberdelia  thank you for all your hard work,

@cleonthethird I tested it and I disagree, for now. It is fast and I like the setup and combination with comfyui. Why I will stay with forgeui for now: It is not flexible and comfyui is not integrated well enough with the main GUI. For example when I use Reactor and want to run the face fix after the face swap, but you can not change the order. You have to load it in comfyui and change the order of the nodes there, but if you load that comfyui into the main GUI it will make the config menue messed up and you can not save it.
Also, advance settings are to complicated and you have to search for them deep in different tabs. It is not intuitive.

Dev has to look for a different solution and also has to change the GUI for that to make it more intuitive.

The thing is, you have a good and intuitive GUI already with A1111 and Forgui, but you want to make something different and NOT better. You see that many times.

Focus should be in optimizing something and using things that are already good.

@bobby888 Tbh, I wish elements of this weren't true....lol. Because I actually prefer ForgeUI design, but there's just some technical areas where it just doesn't cut it for me (even fails badly). Hence why I switched over a lot of stuff to different things, & SwarmUI became the best of a bad lot of options for me (most of the time).

That's not to disrepect any of the developers involved for any of them, because I think a lot of the UI is just preference based.....but it still leaves me wanting more/something else (basically A1111/Forge, but updated/supported like ComfyiUI main; which seemingly just doesn't exist. Closest thing seems to be Neo).

"Also, advance settings are to complicated and you have to search for them deep in different tabs. It is not intuitive." - I use the search function for this btw (at the very top of the menu). You can just just search for the setting you want (because like you said, going through all the tabs is a bit clumsy at times).

I thought that was your stage name when you do your full monty routine ;) lol

I don't see how it's much of a problem. "happy" or "blank" are about the whole range of nuanced facial expression the average anime weirdo can distinguish between, anyway lol.

@ShadowCell  Just for fun, use any animated model and ask for an expression. like, anything from pony, any of the others. if you're unable to distinguish between "happy" and "literally nothing going on" and "anything the f else" you might have a neurological order, but don't assume it's common.

@gingkgobalboa *Slow clapping ......that would be the joke lol.

Which is funny, because if you aren't autistic.....then I'd say that would have been very obvious; especially the "nuanced facial expression the average anime weirdo can distinguish between" bit.

That said, it was just a bit of fun; I wasn't trying to actually demean anyone with ASD. The way you overreacted just there though, kind of makes it seem like you actually do like to take cheap shots at soft targets for real.

Haha, you're absolutely right!

Yeah, I will go back to older models I think. the newer releases just seem to be hard not being NSFW

@Veesyr Once when I was generating an image, I don't know why it directly produced a pornographic picture of an old man and a beautiful woman having sex. The key point is that the child was right beside me, which scared me to death.

@Cyberdelia I used the prompt words of cyberpunk, planning to generate a male punk with cybernetic limbs. The prompt words can't be found now as they were randomly generated. There was no adult content at all in them.

can you give example for mixed race

@warphorizon for me, it seemed to recognize prompts with combinations like "Peruvian, Asian" as dintinctly different than "Asian" or "Peruvian" by themselves as it was apparent in the face that it was a new person, or even certain ethnicities like "Korean" as unique from the overall "Asian" race. "Asian" is still a very powerful prompt by itself, so the mixed "Peruvian" had to have a strong weight to come thru in the resulting generations.

I've generated literally thousands of images already with it, but I'll try to show you some of the insane images I've created at ultra widescreen resolutions......in like 30 seconds or less! It's absolutely crazy.

Qwen has a real problem with plastic, CG/AI curse skin. And less than convincing subject lighting. Even looking at scene composition it's often very...awkwardly contrived. Pony/SDXL can also look very fake. Does, generally, without help. So does Flux (non-Krea). But when something like Cyberrealistic Pony hits, you don't even need a skin helper lora and it can have much better, more naturalistic and photographic looking lighting.

That sounds a million miles off the experience I'm getting.

https://civitai.com/posts/27351245

https://civitai.com/posts/27351177

It generates all these in just 15 seconds each. These aren't rare cherry picks either.....just slop prompts, & it'll bang these out 95% of the time. At one point I actually left shortly & came back, & looked through the image gens after. It generated over 100 images.... only 3 of them (yes 3) had minor anatomy defects.

PS - I should warn anyone reading any of this btw lol; I've seen other people's image generation times, & I immediately thought they must have configured it wrong.... from what I can see, it doesn't play nice on old/weak hardware, so don't get your hopes up. On the other hand, if you have a high end GPU then Qwen 2512 is an absolute banger (the older models perhaps not so much, I haven't really used them tbh to comment on how big an improvement has occured since).

These are just boring spam pics, but when I have the time I'll upload the really interesting ones; & they are insane! Prompts like "Catwoman appears at a window in the background with a bat signal appearing overhead as a light above her shining, etc, etc" Literally paragraphs of this nonsense......& it F' does all of it......EXACTLY AS YOU ASK! I kid you not. Just to be clear, this is just random background nonsense on the end of the 4-5+ paragraph long prompts; & it still nails it! It's insane.

@ShadowCell that's better than most but there's still a very artificial quality to the skin. Like it was trained on heavily photoshopped, uncanny valley magazine images. The light is also very unnatural and biased towards contrived, modern studio lighting.

@BRai4now something wrong with your generation settings, prompt, or what type of attention your using, or your on the old qwen model which did have worse quality. qwen 2512 looks better than flux 2 max, chatgpt image 1.5, and nano banana 2 a lot of the time

@goeypants400 you people keep saying that and I don't see it in any of your generations. Synthetic, artificial skin and contrived, unnatural lighting. Loads of tells it's generative.

I'm commenting on YOUR gens and other examples out there. Not issues with my settings.

@BRai4now so you tried it yourself or not? if your just looking at what other people made, a majority of them dont know what they are doing like you and the image comes out like shit. horrible way to compare models bro. generate images using same prompt on your pc and compare with other BASE models like SDXL base instead of pony, dumbass, qwen 2512 will come up looking the best without a doubt. your comparing this model to another model thats been finetuned and with loras on thats retarded (but it beats them usually still). if your asking for 1-1 real life light physics then no model does that. your comparison to pony finetunes in lighting or skin is flawed as fuck, they are firstly finetuned with millions of images and unet models do textures and lighting different than DIT models. there isnt a single sdxl based model that can make a image not looking "generative". put a few loras on qwen 2512 and prompt for amateur and it looks like real life, retard over here comparing a massive retrained finetuned model to a base, imagine a qwen 2512 trained on 10 million more images then only you can compare it to pony

@goeypants400 how about you go fuck yourself little boy. Cry more about shit that doesn't concern you and miss every point along the way, while just saying more shit that's both objectively nonsense or stupid while making claims you don't back up and can't even be bothered to point to anything remotely backing what you say.

@BRai4now bro doesnt have enough braincells to even comprehend basic shit about ai models, lmao this is why you think pony looks better than qwen 2512. what a dumbass i barely understand what your saying, i didnt make a single claim just stated facts. you must not even have your own pc to generate images. dumbass probably didnt even know pony is an SDXL model LOL. how about you shut the fuck up nobody wants to hear your dogshit opinions

consider using Grok... works very well with NSFW

venice.ai. I wish someone had told me about it sooner. They are open source models that it uses, but it is large parameters so it does justice and it even has a "disable telemetry" option along with anonymizing. I have quantized versions of the same ones locally, but it isn't as great with prompt adherence. If you want a free alternative, download LM studio and get abliterated models (models with NSFW filters are removed).

Regardless of which you choose both don't save data, venice.ai, saves to web cache locally and LM studio can be offline fully on your device. The only caveat is no web searches so find someone's prompt that you like for that specific model, ask it to find the structure pattern, and use that as a template.

If you decide to use Grok: don't ask for multiple images at once (e.g: generate 10 different images) because then it will generate them with 10 different verbiage styles which makes 10 different structures. Do it one at a time. This is why I personally prefer open source, it's less creative which is what you want with structure especially with generating the same patterned prompts.

Frugal tip: if you use venice.ai, Grok, or any web LLM, use incognito on a Brave or Opera browser (they have built in browser VPNs) without logging in. each time you close the window and open a new incognito you get a new free session. Endless free prompts. Hopefully this helps, happy generating!

I have the opposite problem. V17 is ignoring my prompting completely and 16 followed to the letter

https://civitai.com/posts/27333538 First wave of bonus gallery images for v17

https://civitai.com/posts/27335978

https://civitai.com/posts/27336418

https://civitai.com/posts/27336525

(Will upload more) The comparison images use old (low end settings), these are higher end image outputs.

I saw someone else upload some pictures with my prompt data in the feed; so I thought I'd just add some things you can change.

Increasing to 35, 40, 45, 50 steps is all good (35 is a really good trade-off for quality/speed). Stop at whatever level starts to hurt performance & generation times too much. Sometimes going higher isn't even better, even just aesthetically speaking.

Euler A Karras is good (per the prompt data), but I also actually really quite like using

DPM++2M & DPM++3M

Also DDPM (it is often either really awesome, or really junk, but worth trying)

DPM++2S Ancestral is also nice; it creates a randomization sequence in the step count that's less linear, so you can get away with fewer steps, but it can also cause some erratic weirdness when you change steps (even by just 5 increments). This is a good thing, & a bad thing. If you use something like Euler, you can often increase steps to consistently generate a better quality image of a cherry picked image you like (& often fix minor defects with anatomy or fingers, etc, instead of going into inpainting/edit mode). On the other hand, something like Euler A will often dramatically change the whole scene composition if you increase or decrease a further 10 steps the same way.

SDXL/Pony just has an issue with sharp faces. On that note; here's my favourite upscalers

4x_foolhardy_Remacri (Everyone has heard of this one; it's cheap, quick, & good)

4x_NMKD-Siax_200k.pth (probably my favourite)

8x_NMKD-Superscale_150000_G (Also good)

4x_NMKD-Superscale-SP_178000_G (Also very good)

4x_NMKD-UltraYandere_300k (Trained especially for anime outputs if you're into that)

4x-UltraSharpV1 (Pretty good, but I prefer other options)

4x-UltraSharpV2 (Very good quality, but it comes at a noticeable performance increase).

https://openmodeldb.info/?q=4x_NMKD-Siax_200k There's a ton of options, & they're all trained for different things, so experiment & see what you like. Stick them in your Upscaler folder & reload your UI, & you should be good to use them.

SDXL/Pony is trained at 1024x1024 resolution basically (that's a box aspect ratio/looks a bit lame). Try to stay within that pixel ratio, however (if you go too high anatomy will start to break constantly! This is where upscalers come in.

1344 x 768

1104 x 896

1104 x 864

1200 x 968

There's some less standard resolutions that generate consistent/good anatomy images. Switch them around to change from Landscape to Portrait/vice versa. Consider what you're trying to generate....because SDXL/Pony will try to fill the image.....even if it doesn't really make sense! So take that into consideration to avoid weird stretched abdomens, etc, that stretch too long across the image frame.

Pro tips - Find you really like an image, but it has busted fingers or a third leg/hand popping out of nowhere like the monster from The Thing....

Here's some n00b stuff you can try to avoid going into inpainting/edit.

1. Increase/decrease step count by 5 increments.

2. Add a useless nonsense junk word on the end of a prompt; if it's so irrelevant it won't actually add anything to the scene, but it will change the image to a degree.....perhaps enough that the added randomness fixes whatever anatomy issues you're having.

3. Get in the habit of structuring your prompts neatly together; Style prompts, subject matter, setting/background. Use prompt BREAKS if you're going to have loads of different subject matters; this will stop them mixing together. This will save you headaches in the long run. If you're going to add something; don't just add it onto the end of your prompt.......add it to the correct section. Remember; things at the start of your prompt also have more weight than things towards the tail end of it!

4. Can't get something to add to your image? Use (Big breasts:1.2) (Big breasts:1.4) etc. Don't start too high, or it might really screw things up, you just want the bare minimum needed until you get the result ideally. This is really useful if you want something like Short Hair, but it's not short enough. (Short Hair:1.2) etc.

I'm tired as hell, so I'll fix this up later.

https://chat.qwen.ai/ I've been looking for an alternative to ChatGPT since it turned to useless trash. There it is if you want something for good academic research (& no that's not a euphemism lol). Free Qwen 2.0 images (that's the newest unreleased model). (edit - Seems it's ended already, I hit a quota after 20 images the other day)

PS - it's censored. It will generate a woman in a bikini for instance without being a total cuck like ChatGPT though.

https://perchance.org/ai-text-to-image-generator (It's not great, I won't lie....but it's free unlimited, & can do xxx.....it's like SD 1.5 quality though. Thought I'd mention it for people with only a phone or total junker computer, & no money). I'm not aware of any other free 'unlimited' options like this?

If anyone else has any other good free stuff (not fake shill junk); keep em' comin' ^^

Incognito mode does not save you from being tracked, fingerprinted and telemetrised dude. It is just about cookies. You are all so naive people here...

@ArnAbbas Tbf, they never said it did....no one did. Might want to work on your reading comprehension next time, instead of tone deaf, obnoxious rants. They were referring to bypassing free limits. "new incognito you get another free session. Endless free prompts"

& that last bit you claim about it just being about cookies is also completely wrong btw (even if they were talking about that).

"Incognito mode does not save you from being tracked, fingerprinted and telemetrised dude. It is just about cookies. You are all so naive people here..."

@ArnAbbas as @ShadowCell said, I never said that. In fact, I never even mentioned cookies, so I think that you may want to stay to your native speaking to avoid any confusion or kindly translate my words with a translator. Telemetry is not the same thing as cookies. Telemetry captures CPU usage, application errors, feature usage rates, network latency. Cookies collect session identifiers, login status, shopping cart items, preferences etc. (basically, anything to identify you)

Venice minimizes data collection by prioritizing local storage and decentralized processing. Similarly, you can disable general browser telemetry (diagnostic reports and usage pings) in the "Brave Shields & Privacy" settings.

Just in case you didn't put 2 and 2 together (by the name calling I will assume you did not): using Brave's automatic session wiping and Venice’s local-only history together ensures that your AI interactions leave no traces on your device or their servers, effectively decoupling your identity from your data.

Or you can just use Chat GPT, get angry at anyone technical in the comment section trying to help the community, and have your data audited by Sam Altman who will then show your data to Congress. Whatever floats your boat, dude. B)

@ArnAbbas bro nothing that I just said had anything to do with what I said or even implied. I'm geeking out laughing over here. I appreciate that laugh. You said that like ur mom just found ur ChatGPT chat history of failed NSFW prompts and that it is all my fault.

Thanks for the tips, can you link the Amateur v2 LoRA you mentioned? There are too many to pinpoint the one you suggested. Merci!

This is a problem with most checkpoints (they are trained on beautiful people/models for the most part). That said (& this isn't a joke lol), try "old, fat, ugly"...... extreme prompts like that will still basically just get you an above average looking American that's middle aged lol. If it doesn't work try

(ugly male:1.2)

keep going higher (ugly male:1.4), etc, until it works with terms like that. Use the prompt BREAK system to segregate those terms.

Tbh the only solution for you is to use different models than PONY (ZIT, QWEN, FLUX)

Example

https://civitai.com/images/54932204 "Her boobs aren't big enough, or I can't generate anything other than anime women", etc (or whatever your issue).

If the prompt data is there like this link, I can just copy it in & manually change the prompt to fix the issue on my end.

I can see replies if you post to this thread. Remember, an image is worth a thousand words....

this checkpoint is making T2I not I2V to my knowledge

@krsny1326 in the gallery below they all look like I2V

@dillonburns45717 you make an image using this model, and use dedicated I2V model to make vids using the image generated with this one

@dillonburns45717 i hate that shit, for a beginner its so confusing to have these cool popping gifs on the page but you'd be led to believe oh this model makes this kind of art. i dont think moving images should be allowed if its not for a video model.

https://huggingface.co/cyberdelia/CyberRealisticPony/tree/main

from huggingface - cyberdelia/CyberRealisticPony