The model I think is fairly self explanatory.
Got bored again, decided to give my shot at a turbo anima model for realism, doesn't seem like too many have had much success in it when I checked months ago, not sure about now, part 2 of this set will be a LoRA I'l release after titled [Malebolgia] Custom Goober 80 (Realisticish) Character Style Anima that should help as well, as it was part of the development, as well as was the base of this, either can be potentially of use to anyone who wants to refine upon it, I don't mind if anyone takes this and does merges of their own and put it up for generation, I probably wont, I don't make enough buzz to put checkpoints up for use as well as get paid. (That's not a cry for donations, I'm just stating that I'd rather not put out checkpoints into an already flooded market and take away from any other creators income.)
For use with these model previews my positive prompt was at the base minimum:
masterpiece, best quality, absurdres, realistic, photorealistic, realistic,
My negative prompt was:
blurry, low quality, text, watermark, speech bubble, logo, loli, shota, child, toddler, baby, censored, censor, logo, signature, website, monochrome, greyscale, comic, manga, 3d, unrealistic, anime, cartoon, 2.5d, fake girl,
Previews made with ER SDE sampler, Beta Schedule, 12 steps, but you can probably go down lower to like 10 or 8, images were rendered at 1280 x 1600 without hires fix, in Forge Neo, at a CFG of 4,
And I think that's everything relevant to this model that I can think of, let me know if I left anything out that you need to know, have fun.
https://vocaroo.com/1dCKgpymuk1M
Back to my devilish tricks.
Works well at a strength of 1
Previews made with [Malebolgia] Fraud - Ditch 2 'Turbojuicy Fraud Realistic' Custom Checkpoint Merge Anima
Description
Not sure if I figured out how to put the negative prompt in the 'ernie image' generation workflow, not really sure of the quirks of int 8 or how to gen with it, so take the preview images with a grain of salt
FAQ
Comments (12)
There is something wrong with your INT8 version file. Loading it results in the following error:
ValueError: Unknown quantization format for layer blocks.0.self_attn.q_proj
Other INT8 models run without problems.
Look at the other comment thread on the model, you need to use it in comfyui with a special sort of loader, that's what the 'about this version' mentions of the 'ernie image' thing, you don't use a normal model loader.
Or wait, other ones work? Let me check again after this model I'm baking is done, maybe I uploaded a wrong version
No clue what could be causing your issue, I've appended two images using my originally titled "[Malebolgia] Fraud - Ditch 2 'Turbojuicy Fraud Realistic' Custom Checkpoint Merge INT8 Anima_00001_.safetensors" as well as the civitai automatic renamed "MalebolgiaFraudDitch2_int8VERSION", the images will pop up down below in the gallery after some time, but it seems to work, you can copy the workflows and see if that helps. I've never dealt with the 'ernie image' thing before, but it works with that.
Or rather it'll append them to the already present gallery, goddamn silly site design, because I didn't use the full 20 previews. Check the last two that are appended to the previews.
@Malebolgia Ok yes, it does work when using the custom node with model type ernie. But that doesn't seem to be the best way to provide an INT8 version of your model. I have tested dozens of other INT8 model versions before and it was never a problem to just load them via the normal model loader. I don't have any experience quantifying models myself, but it seems like the most common way with the best possible results is via INT8 ConvRot.
@Lizardon1025 I did try to figure out CovRot with chatgpt, but it's explanation didn't really make much sense to me, I'll try looking into it tomorrow or so, and come with a new one, probably name it 'INT 8 VER non-ernie', but it's about 2 am, so I'll probably fall asleep soon.
@Malebolgia Yeah don't worry about it, it's not urgent at all ^^
Fixed!
@Malebolgia That was way quicker than expected xD
@Malebolgia I'm sorry but there seems to be something wrong with this non-ernie version:
https://civitai.red/images/142771262
The quality is way worse compared to the original bf16 and ernie INT8 versions using the identical settings.
@Lizardon1025 No clue how to fix that yet, I'll try to come back to it later, no clue. It was odd to me that the file size was so unusually small, but I'll see if I can figure it out later, not going to spam the site with model revisions since that would be goofy, but if I can figure it out, I'll see what I can do later.