ComfyUI is mainstream in studios now but to put things into persepctive, our last survey showed ~10% of users are ML researchers & ~5% devs building on top. as a community, it really does span the entire GenAI pipeline: research → models → inference → hardware → pro →

ComfyUI is powerful, but not exactly approachable. We always accepted that tradeoff to a certain extent. But as more teams adopted it, we kept hearing the same thing: onboarding difficulties are not just a problem for new users - they also make it so you can't just hand a

Next up: - Deploy workflow as API endpoint - Copy and paste code snippet to call the workflow with our SDK Blink and it will be shipped.

We just added this to ComfyUI the number of people using it right away in Comfy is a little crazy. I guess it's really insane when dropped into existing workflows/pipelines?

cool

we did our own temporal tiling on the VAE. zero-copy seam reuse + fully streaming encoder/decoder. I feel it's unlikely people are beating it without custom CUDA. but that is expected I guess

flux 2 is the first model we released on comfy cloud and open source on the same day

KEEP GOING. go multiplayer, go desktop-cloud sync, go sharing hub, parallel execution progess states, deploy workflow as API, teams, external storage integration - give us the next challenge we want more

the people spoke and we listened, it's coming to the cloud

Will be on ComfyUI on day-0 Super fast. Perfect realism portrait. 2K, 6 seconds

safetensors is sort of bad and it really shows on big models like flux.2 i guess there's just no attempt to optimize for mmap efficiency like it's sort of embarrassing but still respect for being open source

weird that such game-changing features can sometimes only take a day or two to build

Imagine if Lego was so powerful it let you build spaceships. world building tool

ya I saw comfy was in an apple ad, but here's the real flex up there with the goats of open source

"no more forcing artists to learn machine learning before they can create" but also, still let them learn it too. we can have both

all these huge flagship releases this week and somehow what everyone in our community is talking about and using is Z-image never would have guessed

ComfyUI x BR 🇧🇷 translations added

a lot of people have not realized how good this setup is: i generate locally (left) with low res, low steps, experiment and tinker when i get something good, i just drag and drop into right (comfy cloud) and turn up the resolution, steps, switch to fp16 model

we use something faster called matrix mult

sort of insane to still be paying for mj at this point