218 xl

[feral elbow]

but its weird because I use other models that say vpred and they come out fine

[lime sandal]

try with other samplers

[feral elbow]

oo good idea

ive been using euler cg ++

hmm didnt work with euler a

but good suggestion!

[floral mulch]

you have to force vrped and turn ztsnr on

manually

in forge u can only do that with this extension https://github.com/DenOfEquity/forgeFlux_dualPrompt

I'll eventually release a 2.0 but idk yet about the details, still want to try some things for the vae

[wide beacon]

could just use SwarmUI

[floral mulch]

or that

[lime sandal]

or use reforge

[feral elbow]

why does my reforge not have this tab?

what fork is that

[wide beacon]

are you on forge or reforge make up your mind

[lime sandal]

real

im using reforge not forge

[feral elbow]

reforge

[wide beacon]

[lime sandal]

this? https://github.com/Panchovix/stable-diffusion-webui-reForge/

[feral elbow]

oke ill try this

[wide beacon]

lmao

[gritty lynx]

how is the training thing going imi

[lime sandal]

my try with pop5 it works

[floral mulch]

Not good, can get rid of noise and texture but things stay blurry. I'm doing one last thing and then just merging the existing eq models, that was easier

[wide beacon]

https://civitai.com/models/2111688/koeri

vpred illustrious

[floral mulch]

@edgy talon

[wide beacon]

yes ik

i didn't ping because it would be funny

[edgy talon]

latent space dreams did the main conversion and i just tuned it more

[wide beacon]

why illu tho

[edgy talon]

cause noob vpred has issues with it's text encoder

also it's base is 0.1 so smaller dataset

[floral mulch]

noob has bigger dataset than every illust model combined

[wide beacon]

you can fix that

[wide beacon]

not true

[edgy talon]

initially i tried training on a noob merge

[wide beacon]

noob has e621 and danbooru

[edgy talon]

i don't know the dataset of il 2.0 but it probably has more than both given it's 20 million

i did add some furry stuff in this version and i will add more in the 2nd version

[floral mulch]

noob is way more than 20 mil

it's the ENTIRE danbooru and e621 of 2024 august~ and screencaps from anime, and whatever else curated datasets euge got from chinese sites or whatever

[edgy talon]

by retraining from scratch on illustrious

cause you can't undo clip at a high lr

anzhc trained another clip version for eps 1.1 and it improved things but it's only compatible with that version

[edgy talon]

yeah i assume its pretty similar then

i just didn't get good results when training on noob overall

maybe there's a checkpoint amidst their test repos that converges better

i tried 24r2, v29 etc. wouldn't get the noise right

[floral mulch]

testing more timestep samplers

[floral mulch]

neither of them are perfect

v29 seems to be thebest in one angle but hard to say

pick your poison type thing

[edgy talon]

how did you stabilize the noise on your merge 🧐

i think it's the best noob merge overall

[floral mulch]

dont think I did, it's just rng

run the same prompt 10 times get 10 different result

but this xyz looked good so I kept it as a gotcha

good to ragebait noob server nerds

[edgy talon]

lol

[edgy talon]

i trained with log snr uniform and it worked well

[floral mulch]

can play around with this but it's not gonna fix fundamental issues

[edgy talon]

snr windowed is like uniform but a little stronger

i would try uniform first then windowed if it's not working somehow

it clips the snr

[floral mulch]

only tried the adaptive ones

its not good though, adaptive focuses on hardest and noisiest is hardest so all the bins are at 1k

I'll try the other 3 next but lots of settings to try for each so

[edgy talon]

code i used

[floral mulch]

doesnt seem to learn noise as much in from here so far

[edgy talon]

finetuned seele pop3 (imi's merge)
https://civitai.com/models/2125852

[jaunty vortex]

why this name

[wide beacon]

18k finetune on noob eps https://civitai.com/models/2105269/mai-menriallinstruct

[edgy talon]

idk

i can say it really fast so its good

[floral mulch]

beszari

[nocturne cosmos]

new noob model

• Hugging Face：https://huggingface.co/ChenkinNoob/ChenkinNoob-XL-V0.1
• Civitai : https://civitai.com/models/2167995

[wide beacon]

Yeah unfortunately is eps

[nocturne cosmos]

zimg noob when

[wide beacon]

z image team seems to want to finetune an anime version themselves

[feral elbow]

its not that great unfortunatley

[feral elbow]

THIS will probably be the new SDXL very hyped

[floral mulch]

gonna merge it later ish

new knowledge isnt related to the mode lbeing eps or vpred

[plucky dove]

Full 1.0 expected to release by feb 2026 or so

[wide beacon]

imi is merging not training off of it

[plucky dove]

oh ya i know, but from my testing its not trained enough on the new knowledge yet

[wide beacon]

Oh yeah Raehoshi 8 has quite a lot of knowledge of gacha characters

Could be worth merging too

[floral mulch]

[floral mulch]

I should remove watermarks right

pretty easy now with detectors and qwen edit or similar

its either that or

adding watermark/signature to the detected ones

idk whats better

1st is more work but

probably better overall

idk

[wide beacon]

yeah

you could weigh the watermark tag higher so clip and unet are more aware

[wide beacon]

https://civitai.com/models/2194383/karmix-xl-v0

karmix on civit

[floral mulch]

car mix

https://imgur.com/jBO4rrt

I decided to do it properly

so gonna edit off any watermark that's not a stylistic addition to the image

keeping other text otherwise since they're non uniform other than being text

but watermarks per artist are usually too obtrusive and static

especially the big stamp ones or the ones overlaid on subjects

alsao tried chenkins 0.1 model and its so shit

hopefully 0.2 is better but idk

as it is I don't even want to try it for merging

it will just fuck up my model

also moved to comfyui

just need to make a custom node for prompt syntax like scheduling and alternating

didn't find any node that did this

[lime sandal]

https://tenor.com/view/drdisrespect-poogers-forsen-forsencd-doc-gif-22215115

[wide beacon]

https://civitai.com/models/846917/raehoshi-illust-xl

this got a vpred version although it's buzzwalled for 10 more days

[lilac shard]

Oh

Nice

But also damn

[wide beacon]

@floral mulch if you wanna merge it

character list https://rentry.org/xbbfsdua

[floral mulch]

has some stuff i want

[wide beacon]

this got an update

[floral mulch]

https://civitai.com/models/1208658?modelVersionId=2636888 this one too

I think I used v1 or 2 as a ingredient, was top 3 vpred model at the time

[wide beacon]

random finetune from silvermoong a week ago
https://civitai.com/models/2430193

[floral mulch]

has been trying to do something with it but new version comes out soon

[floral mulch]

kinda stuck with merging tbh, cant make the rf merges look sharp enough

doing this most of the time

[olive onyx]

1. sell goon merges
2. gazillion dollars
3. train new RF finetune on 100 H100s

[wide beacon]

are ya gonna experiment with 0.3 ?

[floral mulch]

feel like I'm being gaslit

Ichecked the images on civit/hf

and they are blurry

but the text says "it's very sharp now"

but I'll try to gen with it and see how it goes

[floral mulch]

gonna finetune one of the merges when my sd sacripts is stable enough

[olive onyx]

rightmost looks nice

[floral mulch]

blurry and some artifacts

[lime sandal]

pop3 and the last one look better

[floral mulch]

[wide beacon]

i agree

[olive onyx]

does this image not have metadata or is it just not reading any

[wide beacon]

there's no metadata

[floral mulch]

grid made oujtside of comfy

dont have a good comfy grid yet

[wide beacon]

swarm has a really robust grid script but the image rendering is a bit goofy

[lime sandal]

someday they will achieve the a1111 grid quality

[wide beacon]

it's mostly aimed at making an html grid tho

that you can visualize in your browser in detail

[floral mulch]

thats useful too

I'll make a custom node eventually

just need to think about how it should work

[olive onyx]

consensus on trying to merge RF ?

[floral mulch]

with what

[olive onyx]

anything really

to a degree where it's actually considered a merge

[floral mulch]

if you only merge attention it's easier I guess

[olive onyx]

did you get around to this ? because I'm pretty much at the same conclusion

[floral mulch]

the merges look too noisy and non as stable as vpred models

[floral mulch]

trainer is almost in a trainable state, also got like 32k images by now not counting paintings yet

[wide beacon]

[floral mulch]

got schizo ideas for the future

• allow data-related concerns to run on their most appropriate device or execution path, overlap where safe, and avoid holding training hostage to full upfront preprocessing completion

• move from a fully prebuilt static manifest to an incrementally maintained sample registry with explicit readiness/cache state

• preserve true on-the-fly text-conditioning semantics while reducing GPU idle time through asynchronous TE prefetching.

• what does an epoch mean if the data that exists, the data that is assigned to training structure, and the data that is actually ready are all different sets?

• The current per-image caching approach is simple and works well for small to medium datasets, but it is likely to hit major scaling limits once dataset size moves into the 100k–1M+ image range.

• learn a sparse feature representation of TE hidden states, then perturb that sparse feature space instead of raw token rows.

[wide beacon]

can't be worse than blob ideas

[floral mulch]

what did she do

[wide beacon]

tried to make a vibe transfer for anima

[floral mulch]

[floral mulch]

idk if that will help images look not like this

feels like anima is closer to sd 1.5 than sdxl is