@TheseAlpacas i will test after work tonight but yeah basicaly want to edit anime references.

@TheseAlpacas nice thank you!

so true :D

because most users use CFG 1 which is horrible

@TheseAlpacas I hope the next version allows for a customizable perspective; currently, the Missionary is locked to a POV view. It would be great if the setting could also work when the config value is set to 1.

Qwen has edit and reference capabilities.

Who is 'we'?

I'd never ditch Qwen for Krea currently. Krea is superior, what Qwen does have the advantage with for some things it editing and it is a different flavour, it will replace Klein for me which I used for editing stuff that Krea edit couldn't do (and tbh as good as Krea edit is it's still just a community project, Krea3 will blow away Qwen and Krea2 in all departments)

Overall I'd pick Krea2 if editing isn't that important to you but ultimately both are great. I'd spend the majority of time with Krea2 and get Qwen with a few lora's for editing or for when I want a different flavour/vibe from Krea.

I will be moving to Qwen2.1 as more lora and community support catch up with Krea2; Not because I think its superior, but having one model that 'does everything' is hugely convenient and saves on HDD space :)

Qwen 2.1 is massively superior to krea 2 (I say this as someone who loved Krea 2, & used it 90% of the time before Qwen 2.1 came along).

It has a much higher skill ceiling to get good results back from presently, however; though it can generate hardcore nsfw images right out the box, & with nothing other than the base model. It's edit capabilities are especially first class (though totally confusing relative to what everyone is used to).

Qwen 2.1 is going to have some real growing pains with the workflow regarding how Alibaba expects it to be used...but the community will accept it when they realise what it's capable of, or the rest of the community dumbs the workflow down to get it to that level. For right now; you're pretty much forced to to use their accompanying LLM models that released with it to get anything good out of it though (that, or a whole load of patience lol).

Once a good 4/8 step lightning LoRA comes along for it, & more LoRAs in general, then I think it'll really take off.

I've been around the generative AI community since early SD.

After spending a lot of time messing with it, I feel that Qwen 2.1 is objectively superior by nearly every metric. Krea 2 was hands down the best of the last generation of local i/t2i but Qwen 2.1 feels more like the first of the next gen.

It does love verbose prompting though, lol. There's a reason that the devs don't recommend using it without a PE LLM. That's likely going to be the case for all generative AI models going forward though.

For prompt writing, I'd highly recommend the HauHau version of qwen3.5 9b on HF instead of the official version of the same model that they just released. I was already using it for Krea 2 and other models as a PE so I wasn't too surprised when Q3.5 9b was officially recommended for use with Qwen 2.1. It's a prompt writing wiz but the two new variants that they released in the main Qwen 2.1 repo are a little wacky unless you're fluent in Chinese.

Also, both Qwen 2.1 and Krea 2 use Qwen VAEs which means they really need a nyquist notch node for higher resolutions. That just removes the noticeable 'screen door' effect without blurring or harming the image.

Krea 2 for image gen and qwen 2.1 for edits

@TheseAlpacas whilst this wasn't directed at me, I'll still answer (cause I'll still use Krea 2, & do like it still).

1. It follows the conventional paradigm of text prompting that's easy to use/everyone already knows. You can have fun just goofing around without needing a side LLM (spontaneously generating creative random images on the fly).

2. It can't compete on realism; but it's still a beautiful style close to quasi-realism that I like.

3. It generates faster/better quality (as of present at the lower end); I can generate a really nice image in 20-ish seconds in Krea 2 @ hires. Qwen Image 2.1 is way better quality (& I can generate fast-ish images also), but the quality suffers with the present lightning LoRAs, etc). This will probably be fixed soon once FP8, etc, is mainstream integrated into every UI, & someone like Wuli releases a 'definitive' 4/8 step LoRA though.

@vrgamerz258 "There's a reason that the devs don't recommend using it without a PE LLM. That's likely going to be the case for all generative AI models going forward though."

I think so. Ming Image have gone the same route.

I think the natural course of things is that all this is just going to be seamlessly integrated straight into most UI frontends in the future (so there's no back & forth separation, etc). It just feels clumsy/annoying to use right now, but someone had to be first to aggressively try to force that paradigm shift. The whole situation reminded me of Ideogram "Why can't I just generate images with a simple text prompt" friction most users felt.

a new model has been released and it needs new loras, so wtf are you so anxious about?

nobody's gonna "confiscate" the models you have because newer ones are released.

I use krea and Z-image a lot but also still using sd1.5, and new sd1.5 loras are still getting released.

i feel like you're very entitled and thinks the community should cater to your own specific needs. I don't like you, kid. I'mma block you real quick

@napsumfur225 I honestly don't think anyone would be able to recover from you blocking them, kid. It seems way too harsh a punishment. What, with your zero uploads, zero content or contributions, 4 followers (that probably just clicked randomly for the daily buzz grind or by accident lol), & just generally the fact that blocking people pretty much does zero, whilst also not forgetting the fact that you're a complete nobody, anyway.....how will they ever recover from missing out on all that, huh? 😭

@ShadowCell I honestly cannot recover from the devastating block you gave me.. you don't know me, I don't know you, not your fight, nobody mentioned you, yet you invited yourself to give a free dick ride to a stranger to feel good, and maybe feel good about yourself too.

out of nowhere you went mad about a sarcastic comment not directed towards you, so mad to the point you blocked me (as if I would give a flying fuck) which is hilarious and also telling of how unstable you are.

I finetune LLMs, mostly vision related models for processing medical images (microscope slides/x-ray scans) and label/interpret them. I don't work with image models besides using them, and I'm surprised I even have those four followers here just so you know who's the nobody you're referring to.

and to spoil the fun for you, I was completely sarcastic, but some people like you are boring as fuck and stupid as a rock can't tell silly jokes from serious talk. I don't even bother to block anybody anywhere, but you did block me.. to punish me maybe?

. you're so full of resentment and lack of self-worth, I can tell.

@TheseAlpacas I was just joking, I can't tell why it's not obvious.

anyway, sd1.5 is pretty crucial for the field I'm in, which is working with medical images (histopathological slides; which are the images seen under microscopes.. a tissue sample is taken from the body, then prepared and stained to be studied under microscope) and x-ray scans (mostly lung x-rays but other body parts too).

this specific work also requires annotators, ultralytics models are the undisputed kings in this area since they're very flexible and easy to train.

we combined both sd1.5 and ultralytics in generating, testing and validating these medical datasets, and finetuing vision models. that what was available at the time when we started 3~4 years ago, now we moved beyond that by far but still use sd1.5 for drafts since it's ridiculously fast and use low compute power.

personally, since I worked with sd1.5 a lot I formed a habit (a bad habit, maybe?) of using sd1.5 for creating drafts, trying multiple prompts to see what works best and use that for generating in other bigger/better models. and also for generating some anime/cartoon images for my 6 years old sibling :D

@TheseAlpacas Thank you

@TheseAlpacas Bro, what captions were in the dataset for anal scenes? I want to understand the best way to trigger LoRA specifically for anal.

@TheseAlpacas can you publish an edit workflow? thanks

I'm not certain what you are asking. 'er_sde' is a sampler and 'simple' is a scheduler. The recommended combo for this LoRA is 'er_sde' as the sampler and 'beta' as the scheduler.

If you're using ComfyUI then they are both included as options for the default Ksampler but if you're using something experimental like a ClownSampler then a 'er_sde/beta' combo won't be available due to the way the math works for those.

@TheseAlpacas I gave up on viggle loras already and generations became much faster and much better, with and without your lora.

viggle's and some other turbo lora I found rely on "manual sigma" node, this is very counterintuitive -at least for me- to write a set of numbers, or copy/paste them when I don't understand what these numbers signify, and that means I can't experiment with something I don't understand. writing wrong sigmas immediately results in black images.

the funny thing is qwen2.1 is very fast without speed loras used at all, also I didn't know that the base qwen2.1 generates a lot of NSFW content without even needing a lora for that which is awesome (I use an uncensored text encoder but that alone can't just make a model become NSFW out-of-the-box)

this proved my suspicions that viggle's lora are the reason behind very bad generations, your lora functions very well without turbo loras at all.

I use euler+simple for base, and er_sde+beta when your lora is used, I didn't test res_2s + beta(57) because I guess those needs RES4LYF custom nodes and I don't trust the careless dev who created it. so I'll stick with with comfy's own samplers/schedulers.

@TheseAlpacas jup my bad sorry. I used it for Qwen Image 2509! Sry

You can use the default workflow for starters: C logo -> Browse Templates -> Image (tab) -> Qwen 2.1 Text to Image. Then, enter the subgraph, add a "Load Lora" node, and connect it to the nodes "Load Diffusion Model" and "Qwen Image 2.1 Cache". If all of this sounds too complicated, let me know.