@lxe
Latent Space
Joined July 2007
ocr is not solved
50
Aleksey Smolenchuk retweeted
AI Toolkit now supports training LoRAs for YuE2-3B, an amazing music model. Support would not be possible without the tokenizer and help from @sin_ceriously . Thank you! Also special thanks to @Machinedelusion. More in 🧵
13
36
5
298
20,274
Aleksey Smolenchuk retweeted
Training for YuE2 now in comfy with my custom node pack! Shout out to @sin_ceriously for the RVQ Adapter that made this possible github.com/filliptm/ComfyUI-…
12
23
6
198
18,152
my arch nvidia driver always works against the latest unstable Linux ABI :) no power management issues, no performance degradation i stopped making PRs to it too… they usually already have this work pending for the next release anyways
set up omarchy mac on my m1 and it didn’t support external monitors. astra wrote a patch that fixed it in 20min. holy shit.
1
159
Aleksey Smolenchuk retweeted
Astra built me an entire stage simulation and control system in Sentinel. I paid nearly $10k last year for software (Depence2) that could barely achieve stage simulation of this quality in real-time. Can't wait to bring this into reality.
33
96
13
1,422
70,588
Aleksey Smolenchuk retweeted
Another Update: in openai, we ARE continuously working on Prism, the scientific/technical writing surface. It is owned by a small team; improvements are coming, although we wish we can ship more. The best way to get update is our discord channel. @vicapow and I are there for you.
36
37
9
913
132,660
Aleksey Smolenchuk retweeted
Rosalind Workbench connects scientific questions to specialized models, tools, and reviewable outputs in one workflow, from protein structure and sequence analysis to sequencing pipelines. developers.openai.com/blog/r…
46
50
11
525
111,819
Aleksey Smolenchuk retweeted
Looks like someone beat Fal to the punch! H3 generating 15 seconds in 47 seconds on a B200 GPU They do this by doing the following: 1. 49 transformer calls → 4 calls 2. About 90% of video-to-video attention blocks skipped 3. Blackwell-specific sparse CUDA kernels 4. Compilation, operator fusion, sequence parallelism, and parallel VAE decoding The resulting inference path is approximately: Prompt ↓ Original H3 text encoder ↓ Noise latents ↓ 4 distilled H3 transformer calls ├─ Video attention: only ~10% of selected 64-token tiles ├─ Text/audio attention: still dense └─ H3 fused + compiled operations ↓ Original video/audio VAEs ├─ compiled video decoder └─ multi-GPU parallel decoding ↓ Video + synchronized audio huggingface.co/FastVideo/Fas…
19
19
3
363
26,716
ok @Meta , maybe “you might like” isn’t the best cta here
1
160
pretty ballsy of meta to inundate us with propranolol and modafinil Instagram ads while they are in a middle of a massive lawsuit stemming from their “haha yeah this is probably legal so let’s ship it” attitude
1
165
Aleksey Smolenchuk retweeted
Thanks for Huggingface team @multimodalart , day 0 support open weight music model!
MiniMax Music 3 is just out on @huggingface 🎶🎵 SoTA song generation, now open weights 🔊, a 8B LLM + 2.7B DiT that turn prompt + lyrics into full songs fits smol consumer GPUs with 🧨 diffusers or comfy model weights + app on Spaces, come play ▶️ huggingface.co/spaces/MiniMa…
8
17
2
210
18,338
Aleksey Smolenchuk retweeted
Fast H3 implementation for Metal. Enjoy, modify, and so forth: github.com/antirez/h3.c Contains code from @liuliu which is welcomed in taking back whatever parts he likes for @drawthingsapp in case there are H3 plans there.
51
177
44
2,205
563,370
Aleksey Smolenchuk retweeted
Replying to @MiniMax_AI
@MiniMax_AI More from my H3 system, im addicted..
38
46
5
542
16,349
Aleksey Smolenchuk retweeted
Replying to @MiniMax_AI
@MiniMax_AI H3 model is some of the most fun iv had in a long time in diffusion. Open source Has been stale for video for so long. This allows people to make some seriously impressive stuff. This is a single render, 1088x1920. I built a custom prompt scheduling system that allows me to schedule conditioning (prompts) to cut in and out as i please. So naturally i made it beat synced to t he song. There is zero editing done to this video
50
65
5
862
32,941
This is the second time I’ve proactively reported smoking on Bart and second time nothing was done. ⁦@SFBARTalert⁩
2
3
314
oh they delivered!!! Thanks @SFBART
1
2
386
oh shit let’s see
1
20
oh damn it worked! Thanks @SFBART
30