model
res + beta scheduler
noob eps finetunes -> vpred + noob vpred finetunes -> various incest maneuvers
(illPersonalMerge_v30, noobieater_v30, obsessionIllustrious_v3, obsessionIllustrious_vPredV10, catTowerNoobaiXL_v15Vpred, noobaiCyberfixV2_10vpredPerp, EasyFluffXLVpred, some loras and manual layer scaling)
7/9/2025 READ THE VERSION DESCRIPTIONS ON THE RIGHT
versions based on incest from february until june
1.5 - 1.0 + various finetunes, illust 1.0, illust 1.1, animagine 4.0, unreleased noob vpred
1.69 - same + routwei + nhl001 lora - > STYLISTIC DOGSHIT
2048a - 1.5 with more of illust 1.1 and 2.0 -> THIS IS AN INGREDIENT ONLY MODEL
2048b - 1.5 with less of illust 1.1 and 2.0 -> SERVICABLE 2048x GENS
alternate 2048b equivalents here, they're better or worse sometimes
https://huggingface.co/Enferlain/juice/tree/main/noob/2048
someone please fix model starege
and join sd mecha discord
Description
same as 2048a but less stable at 2048x1536/1536x2048. more vpred parts taken from v1.5. retained some of the resolution capability at the cost of burnt toast and some style loss. it's alright I guess. in hindsight fuck cross merging parameterization and resolution between pure anime and e621 trained models
FAQ
Comments (33)
hype
oh shit we back
is 1.5 the new one or thats 1.69?
@alternative_Universe all of them are new, check the version descriptions and example images and you can decide if any of them are interesting. But to answer your question, 1.5 is probably the one I'd consider an update to 1.0
see pictures
I know what kind of man you are
1.5 is amazing !! left a previous comment on my old account, but the colors in this version are much more unique especially when it comes to color saturation and contrast in scenic images, and the initial details ΣIH brought feel less random and more pronounced in meaning. excellent merge ♡
thanks! I haven't decided yet whether I will add slightly more detail or not. I've had fun using it as it is but maybe it won't hurt to try
illyaeater whatever you do, here for the ride and to support, ΣIH is my favorite <3
DIH.safetensors have reached real far
I really like 1.5, feels good to use. Thank you!
What the hell does "INGREDIENT ONLY MODEL" mean?
for merging or finetuning
@illyaeater Welp I got one of those i guess. was just mixing illustrious and nai together lol to get a illustrious and nai mode l merge with.
Btw: I was looking at models to merge with my new merge nodes, ig this will be one of the base models lol if you got another recommendation that isn't your model I'd like to know.
Edit: I guess ill merge this with my illu 2.0 stable rouwei nai mix lol.
@Arctenox I recommend seele for 2048 res now. Illustrious is too difficult to merge with vpred noob and has less use now unless there's something specific you want
@illyaeater I am sadly one of the few who can merge with v-pred even though it's annoying, got a few nodes though that makes this process easier lol Ive. noticed. Might vibe code a node to make the process less sell annoying, usually I'm able to do like 10 - 40% v-pred clip but... Last time I merged with v-pred and eps it was 50% a lot of the blocks you'd have to keep as the epsilon though which I don't mind as there is something specific im going for which is a stylized ckpt.
@Arctenox yeah just saying it's not worth it, seele is much better than anything illustrious has to offer on vpred and you don't need to lego with layers
@illyaeater Completed the model like 6 hours ago lol. Vibe coded a node that makes it save as epsilon.
1.5 is by far the best Noob vpred frankenmerge I've tried. It's a little unrefined pre-upscale but hires fix irons out any issues. Doesn't have style frying or color issues a lot of other mixes have. Nice job.
I have tried many models, and this is my favorite model. It responds well to the artist's name and has clean lines. Many of the models I have tried have messy lines. Thank you to the author for bringing such a useful model.!!!!!!!!!
I got two questions:
1. is res and res multipstep (in comfy) the same?
2. How did you figure out you need res? What test prompts did you use to test all the samplers? So unusual >.<
OH and a 3. : You sure this is vpred? In comfy I got two identical images with and without the ModelSamplingDiscrete Node put on v-pred and active znsr. Without that node in the workflow the results are just the same.
They are based on the same paper but not the same implementation. The one I use is this https://github.com/DenOfEquity/webUI_ExtraSchedulers/blob/neg/scripts/res_solver.py
for comfy it's in this repo https://github.com/Clybius/ComfyUI-Extra-Samplers
I figured it out by it looking better for my prompts than the other options, although I haven't kept up with all the new samplers, but I see no need to change it unless something faster with same quality appears.
And yes, it's vpred. I think you don't need a node for models that have the vpred and ztsnr keys, they are picked up by default? But I don't use comfy so idk
The best model I've used for artist prompt adherence, and it's not even close.
pure awesome sauce!!!
This model has pretty much ruined any other checkpoints for me, lol. Nothing else comes close to this in terms of adherence to certain characters and their details.
My only major complaint is that the outputs frequently tend to be overly saturated, even after putting tags like 'saturated' into the negative prompt.
(I'm running 1.5)
Yeah it looks a bit saturated especially coming from other stuff or when comparing outputs. Maybe it will be addressed in the RF version. You can also try some of the later models with "pop" in the name on the hf link it you feel like it. The one that says seele-pop3 has been scoring well against other models in prompt adherence and aesthetics.
Do you have any specific prompts where the saturation is really noticable?
@illyaeater Not quite. It appears to be variating between frequent and occasional, depending on the art style, with no style prompt leaning towards the former.
Generated images of characters without prompted backgrounds tend to be put into backgrounds with highly saturated orange, yellow and red while the background itself tends to be under-detailed (which I can just chalk up to being the model leaving me to fill in the blanks). The effect may or may not be more pronounced depending on how much the color red is associated with the character's design, so this is what I can see being the most likely to trigger the oversaturation (prompting a specific background may or may not help, as I've seen some building interiors being bathed in such colors). Another prompt that might trigger too much saturation might be 'magazine cover', but I'm not too sure about that.
Putting 'red theme' and 'orange theme' into the negative prompt helps alleviate, but it then sometimes implements a different color theme onto the image—usually blue—so then I have to put that color into the negative prompt, too. It leads up to backgrounds or lighting looking rather strange instead of having diversified colors.
I can give such models a try. Chances are, too, that there might be some technical aspects that I haven't taken into account. Instead of ComfyUI or the like, I'm using Fooocus, without additional styles implemented.
@MoltenJotun hmm yeah, not sure. the previews I used here are from reforge. in general I am aware though that images are more saturated, especially compared to eps models. it's just a byproduct of trying to get as close to noiseless and solid colors as possible. I'm guessing the less noise a color has the more it collapses to being uniform or "clear" which results in it appearing as brighter, but it's just a random schizo thought
Congratulations on being available for onsite Gen❤️. Truly one of the Best Models!!
NO WAY? SOMEONE BIDDED IT, THIS IS A MOMENT OF VICTORY!! TAKE A PICTURE!
*history I’m so dumb
I don't even know what it means but pog