You definitely aren't the only one. 7.77 seemed to have changed a lot; the art style, for one, but notably how it interacts with Loras and prompts—whereas the art style doesn't change too drastically, in my experience, going from 5 to 6 and 6 to 6.9, going from 6.9 to 7.7 was crazy different. Personally, I think I'm going to stick to 6.9. Not to take away from Janku, though; they're doing God's work.

i agree don't know what but the new 7.77 seams off

Adding "anime coloring" to the positive prompt and "realistic, photorealistic" to the negative prompt mostly resolved this issue on my end

this is what LoRa is for, no?

@youreintroubleultimate You can try to fix it with LoRAs, but having to use a LoRA for something that other models can do by themselves is rather suboptimal...

What version is better for that? v6 or v5?

@Patuwa Regarding the character knowledge, v6 was better than v6.9, which was still better than v7.77. I haven't tested the knowledge of previous versions...

@lizardon1024 ok, thanks. I will try v6.

@Patuwa If you are interested in a model with better character knowledge and later knowledge cutoff, you could also try Anima Preview3 or any new Anima model based on it. Anima is just an amazingly versatile base model which understands natural language as well and seems to be generally superior to Illustrious/NoobAI. But the model with the best character knowledge at the moment is Mugen 0.5, although it's slightly more difficult to use.

Even with short or complex prompts? Does this apply to every version or only to a specific version? Which version do you think is the best for that?

@lizardman72119 I usually use between 15 and 30 individual prompts, which would be about 30 to 40 words. In addition, I have my standard prompts that I always use. That would be another 16 prompts. For me, that's short compared to the prompts other people use for their creations.

@lizardman72119 At this point, I would recommend versions V7.77 and V6.9. Sometimes V6.9 does certain things better, but you'll have to try it out yourself.

Isn't that just Anima?

@victaaj672 but is it really though

insane downvotes lmao, whats wrong? just because you wanna drive a rusty car from 2013 doesnt mean I dont wanna build an AI empire :)

Note to self: "civitai people have sand in vagina."

@EpicStyles meh, flex away, haters gonna hate, you can say whatever you want, they can downvote as much as they want, who cares. I have a single 5090 and I'm not butthurt + I don't wanna downvote you, funny how that works. 😉

@canalpa It wasn't even a flex, it was just me telling janku here that I essentially can start locally training anything now finally. but yeah I guess in todays day and age people get offended by anything that wasn't even directed at them..

Oh and on that note, the model is fantastic so far! no complaints here, I can do awesome things with it with my old janku trained lora's :)

@EpicStyles Thats a sick setup, but ill explain all the salty thumbs down.

Youre getting downvotes because you said you have 192gb of ram.

for SDXL its considered extremely overkill with a 32GB GPU; a1111/comfyui probably will use little to no ram as illustrious/SDXL model can preload fully on your VRAM. Since the model size is only 6gb or so. (I might be explaining the concept loosely, but its something like that)

Dont get me wrong, a 5090RTX is a solid flex for owning, but theres also alot of us who rent servers that use the NVIDIA blackwell GPUs (for a suprisingly cheap price most times lol)

Such as the A100, L40S, H100, H200, B200(this card actually has 192GB of VRAM)

These are mostly overkill in the VRAM department for SDXL, but are fantastic for WAN and LTXV.

@LatteLeopard wdym 192gb vram? I have 32gb vram from my 5090, not 192? I have 192gb ram! I hope that wasnt the main point of things I have to clarify xD

@LatteLeopard also, its not overkill, its wiggle room, thats the main point, I am going to SD 3.5 eventually soon and testing its architecture and if I can, I would frontier it, by training it on Anime styles, which would be awesome, before anything else. I guess I j ust have bigger visions than most. I dunno, sorry if I came off as something else. I am simply a humble norwegian with a goal in mind.

@TheSiteCantHandleFeedback That dude's an absolute clown. Let him try to use that toy to call SD via agent and have fun with how slow the RAM swap is. 32G can't even run a 24B in HALF precision.😂😂😂

@EpicStyles No worries, my original comment was trying to explain that owning a RTX5090 is a nice flex, but its the 192GB RAM comment that got people butthurt.

People think its overkill because RAM swapping kinda sucks if you have to use it. But a 5090RTX should be fine with SD3.5 without RAM swapping.

And yes The Nivida B200 is a 192GB VRAM card. It cannot be used for gaming and has no real display out. Its just pure tensor calculations.

If you think thats alot of VRAM the Nvidia B300 is 288 GB

thats why you train on janku first, then it works just fine here :) this is janku's version after all, my janku 6.9 lora's works beautifully here, sounds more like a skill issue to me, I dunno xD