@AlexanderRedde3

Software engineer working on ML image generation techniques & inference. Obsessed with programming and life. Research Engineer @ Luma AI

 Here and there
Joined March 2022
  • Tweets102
  • Following123
  • Followers562
36Media
Zippy retweeted
Luma
@LumaLabsAIOct 19
Exact sequence parallelism for packed multimodal masks. It picks ring attention, a sparse all-to-all or Ulysses per layer, matches single-GPU numerics, and beats a FlashMask + Ulysses hybrid in 47 of 48 cases on up to 128 H100s. lumalabs.ai/news/packed-mult…
7
17
4
99
12,316
Zippy retweeted
Luma
@LumaLabsAIMar 23
Uni-1 is here! A new kind of model that thinks and generates pixels simultaneously. Less artificial. More intelligent.
450
760
689
4,744
6,514,097
Made an fp8 implementation of Flux which gets ~3.5 it/s 1024x1024 on 4090, >2x faster than other methods. github.com/aredden/flux-fp8-… #flux1

GitHub - aredden/flux-fp8-api: Flux diffusion model implementation using quantized fp8 matmul &...

Flux diffusion model implementation using quantized fp8 matmul & remaining layers use faster half precision accumulate, which is ~2x faster on consumer devices. - aredden/flux-fp8-api

github.com
1
3
1
18
1,493
Finally figured out how to speed up my #sdxlturbo frontend! It's so fast that the only way to show the actual speed is to delete the prompt, since I can't type fast enough 😆 .. built with next.js frontend & tensorrt backend.
14
28
7
215
43,858
For anyone who may think that this is a fast-forwarded video, or using image caching, I can guarantee it is not- the images are all base64 and it's using a websocket, the terminal below is the output from the tensorrt rest-api server in real time as I ctrl+y / ctrl+z
3
23
1,456
Now we're real time! 😎 - with SDXL turbo & tensorrt optimized further.
6
2
20
1,531
Now I'm at 38 fps, from switching to taesd- but have no way to view the images that fast 🥹
2
6
315
Just broke 20 fps!
2
291
Optimized further meaning I'm getting 17 fps for full clip encode -> unet pass -> vae decode -> serialize to jpeg bytes.
5
271
Made a neat GUI in react for fun & hooked it up to an optimized sdxl-turbo tensorrt api backend I built for image autocomplete! It generates so fast that my browser cant keep up 🥲 - so fun! #sdxlturbo #sdxl #AIart #stablediffusion
3
5
3
17
2,270
SDXL unet finetuning is going pretty well! FIgured out how to train without deepspeed on 4090s 😀 #SDXL09 #diffusers #AIArtwork
2
10
909
And some more good examples 😊
6
242
Blockadelabs built a neat thing! 🥳 (which I built the ML backend for) 😳
Blockade Labs
@BlockadeLabs5 Apr 2023
REEEMIXXXXX! 📢📢📢 Infinite 360° worlds just got easier to fine tune. Transform anything - including visual style, time of day, & weather while keeping the structure intact! LIVE NOW: skybox.blockadelabs.com #skyboxlab #texttoskybox #aiart #gamedev #aitools
4
3
30
4,232
We release a new ViT-G/14 CLIP model with OpenCLIP which achieves 80.1% zero-shot accuracy on ImageNet and 74.9% zero-shot image retrieval (Recall@5) on MS COCO. As of January 2023, this is the best open source CLIP model. laion.ai/blog/giant-openclip… huggingface.co/laion/CLIP-Vi…
15
159
17
803
298,909
Magical alien garden with shrooms! 🍄 #stablediffusionart #AIArtworks
2
9
707
Will probably be hard to appreciate via twitter, but- here is a true seamless 15616 by 1024 image 100% generated by stable diffusion 🤯
12
8
3
67
11,789
And- here's a 19712 by 1024 image
6
484
Here is a closeup of the right end, so you can get a sense of detail
5
497
Some beauties I made with protogen v3.4 #AIart #AIArtworks
2
1
10
1,037
WOW lexica.art's aperture v2 is pretty impressive! #AIart #AIArtwork
12
1,068
Load more