>To make a new thread, wait for page 8 at bump (and/or image) limit in catalog.
>ONLINE GENERATORS https://tensor.art (NO LONGER SUPPORTS NSFW) https://tensorhub.art (NSFW version of tensor, It's no longer free sadly) https://civitai.com (NO LONGER SUPPORTS NSFW, use the link below for that) https://civitai.red/ (Civitai new NSFW page) https://refineforever.com/ (Free but limited options and probably slow) https://tungsten.run/ (10 generations free every day, but the premium it's cheap) https://novelai.net/ (PAID ONLY)
So they only prompting difference between these two is adding "large pectorals" to it, but it gets much more... smooth and detailed? I wanted to keep the flatness of the first one. I've tried multiples variations of big/large and pecs/pectorals/chest, all of them resulted in this change.
Here's my prompt:
hiromi_higuruma, 1boy, solo, male focus, adult male, male, short hair, black hair, white collared shirt, open shirt, white briefs, black necktie, loose necktie, bulge, cigarette, standing, skinny, muscular, tall, looking away, BREAK white background, simple background, BREAK simple, cartoon, (beitemian:0.7), (idoraad:0.9),
neg:
lowres, worst quality, bad quality, bad anatomy, signature, username, logo, watermark, artist name, old, bad hands, weibo watermark, patreon username, bad face, blurry, messy background, disfigured, extra limbs, watermark, 3d render, realistic, colored skin, ui
So I'm kind of dumb, but still confused lol. The flatness of the first one is because for some reason the (beitemian:0.7) isn't triggering even though it's included in the prompt. I should've realized that based off the art difference.
Still not sure why it's not prompting but whatever I guess?
Full disclosure: I don't pretend to know everything...But:
You are pulling the model into a lot of different directions with the Lora, 2x BREAK commands, and the fact that Beitemian is at 0.7 strength. I am pretty sure that is why the beitemian is not triggering.
Just like if you prompt in opposite directions (describe a characters face, then prompt for the head to be out of frame or facing away etc...) you are making the model choose between prompts that are in opposition to each other. The model doesn't inherently know to "mix" two styles. It just knows you want one style more then the other, so it is skipping one. Also, keep in mind that BREAK is very strong, and I don't think you are using it in the intended way. Try getting rid of the BREAKS altogether and see if that helps, would be my first step.
Interesting, I'll keep that in mind and try playing around with it. I probably don't need all the breaks, that's for sure. Never really understood their purpose lol.
For what it's worth, I've been using some variation of this prompt for almost a year and this is the first time it's happened. HomoSimile XL v4 can definitely mix two artists styles.
Both SD 1.5 and SDXL models are famous for being able to mix. But that mixing is an unintended consequence of the model being imprecise. In fact, if you read what people are saying about "newer" models like flux, they complain it can't mix the way older models can.
I do know that after a certain point SDXL will just straight up ignore prompts once it hits a certain limit, which, again, is what is happening to you (I bet). The "padding" that BREAK commands put in will make you hit that limit faster. Try something simpler like:
hiromi_higuruma, 1boy, solo, male focus, adult, short hair, black hair, white collared shirt, open shirt, white briefs, black necktie, loose necktie, bulge, cigarette, standing, skinny, muscular, tall, looking away, plain white background, (beitemian:0.7), (idoraad:0.9)
Yeah, I'll definitely play around with skipping the breaks. It was just weird it wasn't mixing before adding "large pectorals". Just kind of weird. The mix works fine though, both of these were mixed too:
>>3090593(OP) I've updated and polished the LoRA for my draenei character Care to give it a try? >https://files.catbox.moe/mefsg6.safetensors the trigger word should be "blue_draenei"
>>3090668 >any tips? pay for a gpu cluster and fine-tune it, it isnt fully trained, and even when it is I have no idea how capable the model can be with text
00
Anonymous04/29/26(Wed)02:19:20
>>3090682 but, for whats it worth, in other models (Flux klein) I had better luck when the text in the prompt was in full caps
You use Stability Matrix right? Considering switching away from Comfy with their new investment/worried about their future. Any recommendations? I used A1111 until a few months back when I nuked it somehow messing with extensions. Something simple, private and easy to install would be dope.
>>3090726 Yes, I have always used Stability Matrix, and I currently use Forge-Neo. Forge-Neo is based on the old A1111 UI and functions pretty similarly. If you liked A1111 you will be right at home.
There is a regular "Forge" branch that you can ignore, it is no longer updated. Forge-Neo is active and updating all the time. It can handle Flux/Anima/Wan/Zimage and video generation, though I personally don't bother to make videos.
A couple of tips for ForgeNeo:
There is a "UI Preset" box in the very top left corner. Self explanatory for the most part when you want to change it, but remember that Pony/Illustrious are all build from the old SDXL, so choose "XL".
Clip skip setting is hidden unless you dig through the full settings list. It defaults to 1, which is fine for Noob/Illustrious, but if you are using old SD or Pony, you may want to dig it out and set it so the option will appear on the main screen.
The Only extensions I use are:
"A1111-sd-WebUI-Tagcomplete" (a autocomplete for the prompt box that automatically pulls the tags from Danbooru (or elsewhere, you can pick) can really help you prompt if you know exactly the word the checkpoint was trained on.
"Adetailer": I don't remember how it works in comfy, but you will need to download this extension and different Adetailer detector models for it to work. I only use "Anzhc Face seg 640 v2 y8n.pt" (for face) & "CockAndBallYolo8x.pt" for penis. Although the penis one works best if you build a separate prompt, both positive and negative, which can be a pain.
"sd-forge-extra-sampler" & "sd-webui-agent-scheduler": They add newer Sampling Methods and Scheduling Types. I prefer the 2 combos: Euler SMEA Dy Negative with SGM Uniform ; and "Extended Reverse-Time SDE" with Simple. Keep in mind with Hires.fix you will need to change the sampler method to something basic (Like Euler), as the newer ones don't play well. That goes for "Adetailer" as well.