Brandon G. Neri
@nerijs
Followers
1,191
Following
102
Media
2,399
Statuses
6,178
Making cute corgis @ RetroDiffusion
CDMX
Joined November 2012
Don't wanna be here?
Send us removal request.
Explore trending content on Musk Viewer
Ronaldo
• 729090 Tweets
Portugal
• 594682 Tweets
Venezuela
• 418192 Tweets
Canadá
• 319813 Tweets
#loveislandusa
• 83141 Tweets
#SmackDown
• 69499 Tweets
Vinotinto
• 62833 Tweets
Rondón
• 46777 Tweets
Serena
• 42293 Tweets
#CopaAmerica2024
• 38007 Tweets
選挙妨害
• 37302 Tweets
Yankees
• 36395 Tweets
リンツのチョコ
• 35425 Tweets
Jana
• 34083 Tweets
Kenny
• 33128 Tweets
ボボステ
• 32469 Tweets
Self Esteem
• 25565 Tweets
Angel Reese
• 24733 Tweets
#ManoTengoFe
• 19029 Tweets
ボーボボ
• 16076 Tweets
Laura Loomer
• 15640 Tweets
Romo
• 15231 Tweets
サラダ記念日
• 13000 Tweets
ソフトン
• 12659 Tweets
#CANMNT
• 12398 Tweets
蓮舫赤軍
• 11579 Tweets
Red Sox
• 11255 Tweets
Love when non technical people discovers that virtually all human knowledge are compressed into a 2gb file
84
415
5K
Btw, I'm not mocking at all this kind of thoughts. All the contrary, I love when people really learns how these system works.
4
5
891
NVIDIA released new drivers and a AUTO1111 extension for using TensorRT with SD 1.5 and SDXL
Works like a charm! Conversion on a 4090 is ~2mins and support multiple resolutions, batch sizes and loras!!
14
100
485
⚠ Model release: isopixel-diffusion-v1
SD v2.0 768-v fine-tuned model to produce isometric "pixel art"
12
47
217
One of my favorite spanish creators just used the pixel art XL LoRa, his channel was a huge inspiration for me to join the AI world 🥹
Jugando con el LoRA de pixel art para Stable Diffusion XL, se obtienen resultados muy chulos! 💖
13
29
422
5
8
163
Oh boy, this is amazing.
Basically, SDXL can extrapolate well to 768x768 and 512x512 even with LoRa.
But the thing is, it's basically transposing the concept of pixel art to different sizes. This LoRa was trained ONLY with 128x128 samples and can extrapolate to 96 and 64 🤯🤯
6
13
149
🔥 Some cool outputs from my upcoming 128x128 pixel-perfect Isometric Pixel Art LoRa for SDXL
👇
12
10
142
I don't want to spoil anything, but this is HUGE.
That's a 256x256 pixel perfect output from a SD v1.5 model (the theoretical limit was 64px at 512x512)
🤯
10
9
103
@TheGraphicsFrog
Yeah! hahaha
I love to tell my non-technical friends that this AIs basically learns like humans, just observing the images
4
0
90
ControlNet Tiles + img2img + Pixel Portraits LoRa is 🔥
1.0 Denoising strength, "shiba inu, doge, much wow"
2
10
74
⚠ Model release:
coralchar-diffusion
SD v1.5 fine-tuned model for cute cartoony character portraits
4
20
73
@keybrackson
Yup, I think that analogy fit better a LLM, we’re literally training all internet + books on them 🤯
1
0
72
I think I definitely can push SDXL to 256x256 pixel perfect outputs. The concept is there, just need a lot more work :P
3
5
61
Can confirm that SDXL 1.0 is extremely easy to train and wonderful to work with 🤯
5
5
59
LEGO BrickHeadz XL LoRa also available on
@huggingface
You can try it there with the widget 🚀
1
9
58
Some samples from my upcoming Pixel Art LoRa for
@deepfloydai
IF-I-XL
Coming soon to
@huggingface
🚀
Pixel perfect 64x64, no downscaling, no post-processing 👀
2
7
55
🔥 Some samples or my upcoming 128x128 generalist pixel art model for SDXL
(It's different from the Isometric one)
Coming Soon ™️
6
3
38
Exploring more prompts.
Fun fact: I'm not even running a good quality SDXL workflow, I improvised a really quick one with only with text encoder 🫠
2
4
36
LEGO Minifig LoRa is up on HuggingFace! 🚀
You can try it with the amazing widget!
2
5
36
I'm loving SDXL 1.0
1920x1080, pixel perfect, no artifacts, just pure corgi goodness
3
0
35
LCM Lora works great with Loras!
Just updated the Pixel Art XL model card with some best settings for using it with LCM:
3
4
32
Can confirm Würstchen can do cute pixel art corgis out of the box
1
1
30
Yesterday I adopted this little doggy. He was abandoned on a heavy rain.
Welcome home 🥹❤️
5
0
26
Some outputs of my newly trained 32x32 pixel art model
cc:
@hardmaru
@PublicPrompts
@EMostaque
@NerdyRodent
@emmanuel_2m
@RiversHaveWings
@KaliYuga_ai
@prompthero
⬇
6
0
25
@RealAstropulse
@TheGraphicsFrog
Yeah! hahah
Even harder for anti-AI people, that thinks these models go to Google and make a collage from images from there 🫠
1
1
23
🔥 LoRa release: Pixel Survivors Character
Some fun experiment aimed for modding of a certain game 👀
2
4
22
First attempt at training a
@deepfloydai
IF LoRa 👀
Test on Pixel Portraits v1 dataset, 1200 steps on Stage 1 only
Thanks to
@huggingface
@multimodalart
for the compute 🔥
4
2
22
This is probably as close as we can get for now on high-res pixel art on SDXL :(
I'm probably releasing this model and the workflow. Although not pixel perfect, outputs looks great for wallpapers or inspiration. Even as a base for manual fixing.
2
3
22
SSD-1b is awesome, when used in conjunction with all the optimization techniques, you can pretty much do a x4 performance compared to SDXL 🤯
Also the VRAM footprint is way smaller
1
1
21
Your own pixel detailer Lora coming soon to RetroDiffusion 👀
Flat shading can also be added with negative weight to make images with more shading!
-20%, 0%, 20%, 50%, 80%
2
0
19
2
1
21
Here's a fixed GIF of the animations, for now limited to 4 frames each direction
4
2
21
Ah yes, the famous Pisa Tower in Italy 🫠
(Look at that perfect outline btw)
1
2
19
@Carlos_Bes
@DayoScript
Aquí es cuando digo: ¿Realmente importa? El 80% del público nunca ha jugado Rol y creo que el objetivo principal de la campaña es entretener y pasarla bien. Igual no es justificación pero a veces creo que debemos de tomarnos la cosas un poco más a la ligera...
2
0
16
I'm feel no longer limited by my artistic skills, and I think that's the beauty of AI.
Way more people can express themselves through a game/writing/artwork/code with the help of AI 🥹
1
1
18
🔥 Just release my first public pixel-art LoRa
Pixel Portraits v1
2
4
18
Really happy to join the RetroDiffusion team 😄
Excited to see what comes next 🚀
I'm super excited to announce that
@nerijs
is working with me on developing more base models and loras for
#RetroDiffusion
!
Brandon is one of the most talented model trainers I've seen, and I'm absolutely thrilled to be working with him. Stuff is going to be CRAZY!
3
2
32
5
1
17
It was a low quality dataset I had for SD 2.0 back in the day.
I'm extremely impressed by how little SDXL needs to learn for new concepts or styles. The x3 increase of param size definitely shines here.
Maybe with smaller LoRa can achieve better results and more training
2
2
16
"vaporwave diorama"
SDXL 0.9 is kinda hard to train. I've been doing lots of experiments with LR and rank-alphas, pushing for small file size and training speed.
So far, ~160mb, 15 min runs on a A6000 for pixel perfect 128x128.
4
0
16
I trained a model on licensed art I have for a personal project and turned out amazing!
I'm still not sure if I can confirm or not that 99% of the art of the game is AI generated since Steam is not allowing AI art :(
Hopefully you can see it in the future
3
0
16
Thanks for all the support, It's been a wild month ❤️
Pixel Art XL is currently the most liked style model on Civit 🚀
3
2
16
@lapetitemachine
Yo estuve experimentando hace un tiempo y pude sacar ésta, aún no tenía tan buen calidad como los modelos de ahora.
1
1
13
Like I said before, a higher resolution foundational model it's key for better pixel-perfect pixel art.
This are some raw outputs from SDXL, insaly basic prompts like:
"isometric fantasy assets"
"fantasy game icons"
0
2
15
👀👀👀👀
#RetroDiffusion
Update v7.5.0!
Key features:
Minecraft style 16x16 item & block models
LoRA support
Built in specific pixel art styles (Top-down, Front-facing, Flat Shading, & 1-Bit)
Check your email for the update, or buy it here:
2
4
38
2
0
13
I just captioned an entire dataset with GPT-4V for an extremely specific model that requires some really good captions and I'm blown away by how powerful it is. 🤯
Can't wait for API access, hopefully tomorrow will be announced :c
2
0
14
I'm seriously considering stop reading the Stable Diffusion subreddit.
According to them, everything thrash, SD 1.5 is the only good thing with extremely overfitted anime/hyperrealistic models.
NO NSFW = thrash
💀🫠
6
1
13
Dunno how I ended up in this part of the latent space but I like it
0
1
14
One year ago I was struggling to make a prototype for a game since I have no artistic skills, I managed to buy some art but since the budget for that project was literally $0 it was hard to continue buying new assets.
Now, I'm limited by my own imagination 🔽
3
0
14
👀 soon!
Specially for cute corgis
Two of the styles coming to
#RetroDiffusion
in the next couple weeks:
Top-down, and Front-facing perspective.
Super consistent and beautiful LoRA's designed by
@nerijs
, perfect for
#gamedev
, modding, or just making cute corgis.
2
3
47
0
0
11
You can try it here!
.ckpt and Diffusers version available
2
1
10
Working with DeepFloyd IF it's a completely different world.
T5-XXL is way different from the typical CLIP from SD models, can vastly improve a undertrained Lora
I still have doubts on how well Lora for both stage 1 and stage 2 will work 🔽
1
0
11
Lots of interesting things with high res SDXL pixel art.
This is a sample from a quick test I'm doing, low quality dataset and captions
As you can see, the effect is good, since there are way more pixels, even though it's not pixel perfect, looks kinda good
2
0
9
Looks like this qualifies for the
@HelloCivitai
contest
Post images and add it to your collections 👀 (I’ll give you a not-so-cursed Corgi in exchange)
2
0
8
🔥🔥🔥
0
0
11
Some technical details:
Really low training steps and low rank and alphas for the LoRa (64,64)
It still needs way more experimentation, mainly with LR and optimizer.
It's trained on SDXL 0.9, btw
2
0
10
Check out the new RetroDiffusion update! Lots of amazing stuff! 🔥
#RetroDiffusion
9.5.0!
Too many new things to cover in one post, so I'll summarize and give more details later.
8 new modifier models, game characters, isometric, and game console specific styles!
Background removal!
LLM prompt enhancer!
20% speed boost!
Links below!
7
7
69
0
0
10
I'm having so much fun with the Wojak LoRa, it's hilarious :P
1
0
9
🥹
We've fixed an issue in the
@huggingface
Hub that showed 0 downloads for some 🧨 diffusion models repos, including LoRAs!
Check out the LoRAs ordered by downloads:
2
2
20
2
0
9
Saying that SD compresses all human knowledge into a 2gb file was quite a statement 😂
Also, just learnt that people on twitter can take out of context anything so fast :P
1
0
10
Quick test of a new XL model for whole scenes.
This one I'm testing outpainting
1
0
9
Weird things happen when you push the latent space too far.
SD 1.5 at 4096x4096
0
0
9
Raw output from a new model I'm working on.
Full fine-tune of SD v1.5, trained on 1024x1024 samples, processed with neural pixelization technique.
Looking promising, I you ask me, this are pretty useful asset with minimal interaction.
Prompt is "isometric farm"
4
1
10
Testing out a new pixelization method for vastly improving diffusion models for generating pixel art.
The first one is the raw output from Stable Diffusion.
The second one is downscaling to 64x64 with NN.
The third one is the new method.
3
0
8