Could u share some of the prompts you’re using for editing

Can you share a workflow with us ?

@Lorian Because it's an edit model, you might find this disappears real quick from Civit. Might want a backup on HF.

@HonZuna "Can you share a workflow with us?" 😈😈?
The default one works

@Lorian My bad, I didn't ask clearly. I meant the training step.

@Lorian How long did that take? Did you rent an H100 or something like that?

@Lorian Or maybe someone will finetune Klein. Whichever comes first, we're sure waiting a while for z-image base lol.

@sneedingonmyligma420 There was a report that its coming sooner than we think.

@LiveBurton we were told the same a month ago lol

@sneedingonmyligma420 respect Chinese culture!

@qiaoy123 not useless

Definitely not, because that would break TOS and cause the model to be removed from CivitAI and Hugging Face. Don't even think about trying it.

Works

I sugest using a "FaceDetailer" with the appropriate segm/bbox, and a SDXL-model.

Klein is 100% better at using multiple LORA. Much less quality degradation.

All models suffer from this btw, some just do better.

@Loraholic I never thought about using that for anatomy parts other than face and eyes. Totally forgot about it. I'll have to look at what will work best. thanks

Yes Klein handles multiple loras MUCH better than z-image

Nobody trains on disitlled Flux.2 models, lol

Why would you even think that anyone trains on the distilled model if the base is available...?

@zym0x Download the lore and try it))

@sizyking "A lot of latent noise is clearly visible"- Did you change the Sampler to EulerCFG++? Don't. Use Euler with CFG 5.0 and 20-50 steps. With the .. LoRA .. you'll want the strength at +0.70, even when combining it with other LoRAs. If you still have Noise, test with a simple prompt to see if Klein is changing the genitals to underwear and fighting your prompt tokens.

@NeonScreams Thanks for the advice, I'll try applying these settings.

"32b please" type comment

I did manage it with asinine prompting. But we're talkin like 3 paragraphs of a Romance Novel describing every damn detail by name, location, position, alignment, connection, texture, coloration, depth, width, blah..
However sure enough, there's a tentacle in there. pats self on back

YEs my device wont run 9b , only 4b works

Yes I am right there with you, really want 4b. I'm going to try to use it anyway though, probably won't work.

🤫

share workflow please

please what values for settings, lora values for strength and cfg values. Thanks!

@FiberCreations2035 the regular edit workflow with the lora works fine. throw in some nsfw tags rather than just mentioning "nude".

@Lorian what means no faces? so without head? How did you prepare that?

I'm currently training a model using 9 Control and 9 Target images. Did you use this same paired method, or did you just go with a single dataset? Cuz yours way works really well for editing

@Lorian Wow I'm shocked normally for editing models it should be trained paired

@Lorian did you use control images? If so what were they like and how many?

Thanks for the tip. Any change you would be kind to share the basic values. CFG, loras strength and clip strength. is there a generic lever to keep the loras at in general ?

@FiberCreations2035 I'll list my sampler settings, they give pretty good results. But fair warning, I use the base model (not the distill) so expect much longer gen times if you do the same. The quality is much higher though.

Sampler/scheduler: res_2s + bong_tangent

Steps: 10-12

CFG: 3.00

Shift: 1.00 (unchanged, also note that the bong_tangent scheduler ignores shift anyway)

You could probably use the same sampler/scheduler with the distill by doing 2-3 steps with CFG 1.00 instead, but I haven't tried that so don't take my word for it.

If you don't want to download RES4LYF for res_2s/bong_tangent you can get pretty good results with Euler/normal 20 steps too.

I haven't used any other loras so can't offer any general advice for those with Klein, sorry! Generally speaking though you don't need to mess with clip strength much; either leave it the same as the lora strength, or don't bother including clip with the loras at all (I don't).

@nsfwVariant Thank you! Had to remove the sample in place and replace it with the recommended one. Redo the connections and adjust setting! Got alot closer! Not exactly but close enough! You're awesome!

Can you share your edit workflow?

@Lorian i use one of your's prompt and look how bad it looking instead your's.

@210881175 try to decrease number of steps

@OddDimension99 Thanks, I put 8 steps instead 4 and it looks better. Much better. Try to play with it.

@210881175 yes, most fixes, particularly in low step models like this, is giving the model enough time to think and not panic and rush. thats exactly what going from 4 steps to 8 does. it doesnt have to rush and collapse the latent into something so by step 3 its polishing and done at step 4, it can explore and think at 8 steps.

gooners goona hate