dstack to manage clusters of on-prem servers for AI workloads with ease
ā¢
7
chansung park PRO
chansung
AI & ML interests
None yet
Recent Activity
reacted
to
their
post
with š
1 day ago
Simple summary on DeepSeek AI's Janus-Pro: A fresh take on multimodal AI!
It builds on its predecessor, Janus, by tweaking the training methodology rather than the model architecture. The result? Improved performance in understanding and generating multimodal data.
Janus-Pro uses a three-stage training strategy, similar to Janus, but with key modifications:
ā¦ Stage 1 & 2: Focus on separate training for specific objectives, rather than mixing data.
ā¦ Stage 3: Fine-tuning with a careful balance of multimodal data.
Benchmarks show Janus-Pro holds its own against specialized models like TokenFlow XL and MetaMorph, and other multimodal models like SD3 Medium and DALL-E 3.
The main limitation? Low image resolution (384x384). However, this seems like a strategic choice to focus on establishing a solid "recipe" for multimodal models. Future work will likely leverage this recipe and increased computing power to achieve higher resolutions.
posted
an
update
1 day ago
Simple summary on DeepSeek AI's Janus-Pro: A fresh take on multimodal AI!
It builds on its predecessor, Janus, by tweaking the training methodology rather than the model architecture. The result? Improved performance in understanding and generating multimodal data.
Janus-Pro uses a three-stage training strategy, similar to Janus, but with key modifications:
ā¦ Stage 1 & 2: Focus on separate training for specific objectives, rather than mixing data.
ā¦ Stage 3: Fine-tuning with a careful balance of multimodal data.
Benchmarks show Janus-Pro holds its own against specialized models like TokenFlow XL and MetaMorph, and other multimodal models like SD3 Medium and DALL-E 3.
The main limitation? Low image resolution (384x384). However, this seems like a strategic choice to focus on establishing a solid "recipe" for multimodal models. Future work will likely leverage this recipe and increased computing power to achieve higher resolutions.
upvoted
an
article
1 day ago
Open-R1: a fully open reproduction of DeepSeek-R1
Articles
Organizations
chansung's activity
commented
a
paper
about 2 months ago
fix outdated gr.inputs and gr.outputs
1
#2 opened 11 months ago
by
chansung
Is this a merge?
4
#3 opened 11 months ago
by
mrfakename
Crazy good!
10
#1 opened 11 months ago
by
Rrrila
ā” WebGPU Benchmark Results (33.83x speedup)
#39 opened 11 months ago
by
chansung
Apply for community grant: Personal project (gpu)
9
#1 opened over 1 year ago
by
chansung
Upload klF8Anime2Fp16.safetensors
#2 opened over 1 year ago
by
jphan32
Upload packages.txt
#3 opened over 1 year ago
by
jphan32
š© Report: Not working
9
#3 opened over 1 year ago
by
prasunad
Can I run this locally?
4
#2 opened over 1 year ago
by
BBLL3456
This app can explain llama.cpp to me very well but most of others won't!
6
#3 opened over 1 year ago
by
inthf
Apply for community grant: Personal project (gpu)
2
#2 opened over 1 year ago
by
chansung
Apply for community grant: Personal project (gpu)
1
#1 opened over 1 year ago
by
chansung
Please tell us the prompt format!
4
#2 opened over 1 year ago
by
TheBloke
why is it called "mini"?
1
#4 opened over 1 year ago
by
chansung
Adding `safetensors` variant of this model
#1 opened over 1 year ago
by
SFconvertbot
how many tokens generator per second about inference llama-13b-hf on A10G
4
#2 opened almost 2 years ago
by
baby1
Combine fine tuned model
2
#4 opened over 1 year ago
by
Bloodofthedragon
Refresh model
5
#2 opened almost 2 years ago
by
Whitepaper