/g/ — /ldg/ - Local Diffusion General

gm saars

Blessing the thread before schizo neet malware spam

>mfw Resource news

06/14/2026

>SCAIL-2 GGUFs quantizations
https://huggingface.co/realrebelai/SCAIL-2_GGUF

06/13/2026

>PRXPixel (text-to-image, pixel space)
https://huggingface.co/Photoroom/prxpixel-t2i

>SCAIL Auto Extend
https://github.com/Brobert-in-aus/scail-auto-extend

>MotionBricks: Scalable Real-Time Motions with Modular Latent Generative Model and Smart Primitives
https://nvlabs.github.io/motionbricks

>dyfuzor-web: turns an Excalidraw scene into an Ideogram-4 structured JSON
https://github.com/karolrybak/dyfuzor-web

>sageattention-autotune: Autotuned block sizes and other QoL improvements
https://github.com/woct0rdho/sageattention-autotune

06/12/2026

>ComfyUI-Flux2Klein-Enhancer: Conditioning enhancement and reference latent control
https://github.com/capitan01R/ComfyUI-Flux2Klein-Enhancer

>InterleaveThinker: Reinforcing Agentic Interleaved Generation
https://zhengdian1.github.io/InterleaveThinker-proj

>Experimental Anima LLLite Regional Controlnet
https://huggingface.co/Sen-sou/Anima-LLLite-Regional-Controlnet

>World Tracing: Generative Pixel-Aligned Geometry Beyond the Visible
https://haoz19.github.io/world-tracing-page

>VietFashion: Benchmarking Sketch-Text Composed Image Retrieval for Cultural Outfits
https://hng0303.github.io/VietFashion

>Modality Forcing for Scalable Spatial Generation
https://modality-forcing.github.io

>VideoMDM: Towards 3D Human Motion Generation From 2D Supervision
https://videomdm.github.io

>EvTexture++: Event-Driven Texture Enhancement for Video Super-Resolution
https://github.com/DachunKai/EvTexture

>Budget-Constrained Step-Level Diffusion Caching
https://github.com/Westlake-AGI-Lab/BudCache

>ECA: Efficient Continual Alignment for Open-Ended Image-to-Text Generation
https://github.com/Snowball0823/ECA

>InterleaveThinker: Reinforcing Agentic Interleaved Generation
https://zhengdian1.github.io/InterleaveThinker-proj

>>109058751
Can you please vet these links before posting, why do you seethe about certain anons when you constantly make irresponsible post like this?

File: 01311-3546849431.png (2.46 MB, 1508x1043)

I hate being local...

File: highlights_g_109052846_1781487872.webm (2.97 MB, 822x1456)

the collage script could probably be updated a little
goon vids from anon

>>109058786
become worldwide

>>109058786
Either start sucking dick or trading stocks this sounds like a hardware issue

>>109058798
could i make a lot of money if i posted goon content on tiktok?

>>109058708
this one's not merely heun

>>109058786
:( sorry my iq isn't 90.

File: 01310-3546849431.png (1.94 MB, 1221x915)

>>109058786
I hate having a GPU...

>>109058798
>No mention of my kino Gosling Blade Runner tears in the rain gen.

Cmon bro, I worked hard on that thing.

Nobody ever thanked me for my Chroma chained sampler abstract artworks.

How well does SCAIL 2.0 work with wan2.1 loras?

Gonna try it just wondering if it's a waste of time.

File: 1781141435683844.png (75 KB, 739x814)

If Israeli jews give me local Seedance 2.0 distilled into LTX 3 I will personally rebuild the third temple.

>>109058854
>>109058786
SaaS Lives Matters

File: 354654.webm (3.78 MB, 420x291)

>>109058868
>he thinks his kinos hold any value here
keeeeeeeeeeek
https://files.catbox.moe/eh56jk.mp4

>>109058885
very nice, can't wait

File: FK9B__00003_.png (713 KB, 832x1216)

anima, but with flux 2 klein base 9b fp8 as detailer.

>>109058907
lxt face

>>109058923
>fp8
*puke*

remember when "detailers" meant passing the latent from the first model to the second with left over noise instead of just a dumb i2i pass?

>>109058930
The science says that fp8 produces outputs within the error of f16. I wouldn't doubt science if I were you.

>>109058751
Fuck off malware distributor

>>109058938
I don't think you can pass the latents between models, can you? let me know in the comments below, or send me a postcard addressed to your mom.

>>109058939
indeed, as the science says, trans women are women just like fp8 is lossless fp16

>>109058923
wtf sampler / scheduler were you using here lmao, this shit looks like Pony V6 did if you used DPM++ 2M Karras

>>109058938
Any difference?

>>109058938
If Anima and Z Image share the same VAE, then ZiT could be a great refiner?

>>109058956
I'm an artist. You are not an artist. Your opinion is completely irrelevant.

>>109058981
they don't though, Anima uses the Qwen / WAN Vae. Z uses the Flux.1 VAE.

are these good settings for z image base lora training, i cant seem to figure it out?
linear_rank: 32
linear_alpha: 32
dtype: bf16
save_dtype: bf16
optimizer: adamw8bit
lr: 0.0001
weight_decay: 0.0001
batch_size: 1
timestep_type: sigmoid
ema: false
train_text_encoder: false
resolution: [768, 1024]
steps: image_count * 110

They're calling him the greatest ai artist ever.

worst thing about porn is you imagine if you ever actually had sex with such a hot woman you'd be going for hours and all that, but in reality you prob can't hold it more than 5 pumps before you bust and then you're embarassed

>>109059014
Use anima

>>109059038
IDK man, when I was in college I would just jerk off if I knew I would probably be banging a chick later. usually worked

>>109059075
whys it so Pony

>>109059014
>resolution: [768, 1024]

this makes no sense

>>109059093
It's Anima + high res fix with Z Turbo with 20% denoise . I'm testing with another style right now, coming back

>>109059014
just train on zit with adapter on, zib is trash

>>109059094
Too boring in your containment general thread schizo? Have you asked yourself why this is the case?

>>109059093
Base Anima is radioactive slop. It needs a serious finetune. Honestly I would have been much happier if Anima had been trained only on styles and not on characters, it would give a lot more freedom.

>>109059100
im using ai toolkit which has multiple resolution buckets
>>109059104
i want to use multiple loras and zit trained loras cant be stacked without breaking unfortunately

File: ComfyUI_01397.jpg (3.42 MB, 1500x2000)

>>109058288
LLMs are incredibly retarded when left to their own devices, you have to supply the rules, guardrails for when it runs into something not stated in the rules and goals for it to achieve. I like to caption in two passes because you never know when the LLM will decide to skip steps in a rush to finish.

File: FK9B__00005_.png (685 KB, 832x1216)

>>109058923
more steps, different sampler flow.

>>109059160
*s/sampler/scheduler
bleh

File: 1781465226151852.jpg (128 KB, 1300x1150)

>>109059150
jebby pls marry me

>>109059106
He's far gone, this is the only thing he lives for which is as sad as it is pathetic.
>>109059119
apache2 when?
Oh wait never because the the so called animation kin...oh wait he failed that too what about his UI.....oh that also failed

does anyone have any tips on how to use ipadapter in the latest builds?
Most of the time it doesn't do what I want and just feels like an i2i regardless of ipadapter.
I have a character, I just want to maintain most of the facial features and generate more on that. Instead, it approximates it to the point it might as well be a new gen...
Is that the limit?

File: dff.png (9 KB, 595x70)

what do i do with all this?

File: 2463245.webm (1.37 MB, 420x291)

>soilennials thought this was the craziest moment in anime history

File: Screenshot from 2026-06-14 22-03-14.png (550 KB, 691x1145)

>>109059166

Is Mr. Catjack here? I have a question

>>109059214
Why do you cycle through the same rotation of crushes schizo?

>>109059221
?