Blog
π
6
Search and browse Bench Labs blog articles
ποΈ Bench Labs is cool
Bench-Labs
BenchLabs
AI & ML interests
bench labs
Recent Activity
updated a Space about 1 hour ago
bench-labs/blog liked a Space about 1 hour ago
BenchLabs/Demo updated a Space about 1 hour ago
BenchLabs/DemoOrganizations
updated a
Space about 1 hour ago
liked a
Space about 1 hour ago
updated a
Space about 1 hour ago
published a
Space about 21 hours ago
liked 3
datasets 3 days ago
liked 2
models 3 days ago
upvoted an article 9 days ago
Article
Fine-tune video and image models at scale with NVIDIA NeMo Automodel and π€ Diffusers
nvidia
β’ β’ 79 liked a
Space 9 days ago
reacted to wop's post with π€π₯ 9 days ago
Post
3644
# PixelModel v1
Last month we released PixelModel β a neural network whose weights are literally the pixels of a PNG. It was a toy: 202,752 parameters, welded to 32Γ32 output, trained on six solid-color swatches. It scored FID 566.84 on the Tiny-T2I-Leaderboard, mostly by producing the same yellow noise for every prompt.
Today we're releasing PixelModel v1. It is 8.5Γ smaller β 23,747 parameters β and it beats v0 on both benchmark metrics while being trained on 20,000 real MS-COCO caption/image pairs instead of six color swatches. The entire model now fits in a 160Γ149 PNG.
That image is not a visualization of the model. It is the model. All 23,747 weights, one per pixel.
## links
bench-labs
Blog post [read it here β]( bench-labs/blog)
See us on the [Leaderboard β]( FlameF0X/Tiny-T2I-Leaderboard)
Model card [here]( bench-labs/pixelmodel-v1)
## The catch
A 23K-parameter model does not draw sandwiches. With ~1 parameter per training image, the loss-minimizing behavior is to output the average of all plausible images for a caption β caption-conditioned color, light, and layout statistics. Food prompts come out warm and brown; sky prompts come out cool and bright. That is the ceiling for this size class, and we'd rather show it than crop around it.
# cherry on top π
The model generates 600 images (cpu) in 5 (five) seconds.
Thats 5000 images in 24 seconds on cpu.
The model trained on cpu for just 30 minutes.
Last month we released PixelModel β a neural network whose weights are literally the pixels of a PNG. It was a toy: 202,752 parameters, welded to 32Γ32 output, trained on six solid-color swatches. It scored FID 566.84 on the Tiny-T2I-Leaderboard, mostly by producing the same yellow noise for every prompt.
Today we're releasing PixelModel v1. It is 8.5Γ smaller β 23,747 parameters β and it beats v0 on both benchmark metrics while being trained on 20,000 real MS-COCO caption/image pairs instead of six color swatches. The entire model now fits in a 160Γ149 PNG.
That image is not a visualization of the model. It is the model. All 23,747 weights, one per pixel.
## links
Blog post [read it here β]( bench-labs/blog)
See us on the [Leaderboard β]( FlameF0X/Tiny-T2I-Leaderboard)
Model card [here]( bench-labs/pixelmodel-v1)
## The catch
A 23K-parameter model does not draw sandwiches. With ~1 parameter per training image, the loss-minimizing behavior is to output the average of all plausible images for a caption β caption-conditioned color, light, and layout statistics. Food prompts come out warm and brown; sky prompts come out cool and bright. That is the ceiling for this size class, and we'd rather show it than crop around it.
# cherry on top π
The model generates 600 images (cpu) in 5 (five) seconds.
Thats 5000 images in 24 seconds on cpu.
The model trained on cpu for just 30 minutes.
liked 4
datasets 9 days ago
liked 2
models 9 days ago
liked a
Space 9 days ago