Show HN: Rubbrband – A hosted ComfyUI alternative for image generation

57 points by jrmylee 1 year, 5 months ago on hackernews | 36 comments
The prompting feature looks great — excited to try this out.

[OP] jrmylee | 1 year, 5 months ago

appreciate it! would love to hear your thoughts after using it

everforward | 1 year, 5 months ago

That is available in regular ComfyUI, though theirs does look very easy to use.

The neatest prompting tool idea I saw was one that used an LLM to do prompt refinement for the image, though the implementation I found was more antagonistic than helpful. It garbled my prompts into near-synonyms, giving me a basically equally crappy prompt but for a very different image.

The varied prompts are very neat too, where you give it a template string and a list of options to interpolate and it generates an image for each combo or random combos or etc. Very nice when you want to see the same image with different moods, lighting, backgrounds, etc.

This is backed by both github, and YC? That's pretty impressive. Though sadly the demo on the front-page gives no results for me, just endless animations.

[OP] jrmylee | 1 year, 5 months ago

thanks! We're getting a ton of traffic right now and having some issues scaling... looking into this
What's the difference between your offering and the plethora of others? [1] Do you offer all checkpoints and LoRas even those who are not free to use in generation services or the IP-adapters with insightface as requirements? Rather fishy tbh.

Also offering workflows from your team that infringe on copyrights from the mouse is a bold move.

[1] https://openart.ai/ https://comfyworkflows.com/ https://comfy.icu https://cogniwerk.ai/ https://www.runcomfy.com/

[OP] jrmylee | 1 year, 5 months ago

We don't offer any non-free-to-use checkpoints/LORAs on the platform out-of-the-box. If you wanted to import a model/lora to your own account from a CivitAI link you can do so, but we're at this point letting users do this under good faith that they have permission to use these models.

We also have face IP-adapter nodes in our app!

I would say there are a lot of more minor UX differences about our app than other offerings. One particular favorite of mine is the ability to switch between the Node Editor screen and the Playground screen using cmd+p.

We built this feature mainly bc the node editor isn't great for generating images, but it's awesome for dialing in the exact aesthetic you want with different nodes/settings. We built the Playground screen for generating images once you have a workflow you like.

> The most controllable way

Probably worth reworking this tagline, especially given your goal is to reduce the complexity of ComfyUI. Instead this makes it sound like it’ll have lots of settings or something.

[OP] jrmylee | 1 year, 5 months ago

gotcha thanks. any suggestions there? our goal is to have more levers than something like Midjourney, but without the complexity of comfy. Hard to convey this in a tagline lol

block_dagger | 1 year, 5 months ago

Maybe something like "Simple UI to get started. Advanced options for power users who want more control."
Take a look at OpenArt and how they're positioning themselves. They ditched their free and hobbyist users and now they're doing workflows for business and marketing use cases. Word on the street is that they're making an absolute killing. More revenue than giants like Leonardo, which is a full on drawing and animation suite.

You'd be surprised what making workflow easy can unlock in revenue potential. Look at how they're selling it and lean into that. I'd be doing it if our product was shaped like that.

[OP] jrmylee | 1 year, 5 months ago

Thanks this is very helpful
Looking at your YC history, Rubbrband was initially meant to be "a way to train open-source Machine Learning models in just one line of code." I'm curious what this initial offering was and why you pivoted.

[OP] jrmylee | 1 year, 5 months ago

yeah we went through a few pivots actually - Rubbrband was initially a mobile app for musicians. It would listen to your practice sessions and give feedback. We pivoted out of this primarily because of issues with the technology and market.

pizzafeelsright | 1 year, 5 months ago

Maybe they're using the usage to train?

Maybe the humans have been trained by AI to give AI a job?

Maybe the pivot was "gotta build something"

Imagine the controls on one line of code and how long is that line?

Word wrap: OFF
>some major problems we’ve had with it. First is that it runs primarily on your own hardware

That's not a "major problem", that is the one most important core feature that anyone who is looking for full creative freedom and who doesn't want to enter "safety" censorship and surveillance hell should be looking for.

You can run ComfyUI with as little as 4 GB VRAM, which is not that much anymore.

[OP] jrmylee | 1 year, 5 months ago

this is true. However I would my friends and I only have Macbooks so for us we wanted something that ran on the cloud

spmurrayzzz | 1 year, 5 months ago

Are you saying that the torch performance with MPS support enabled didn't meet your performance expectations? Or are you using an intel macbook and/or one with a tiny amount of ram/vram?

jokethrowaway | 1 year, 5 months ago

The M1 was released in 2020, chances are by now a macbook has ARM and neural cores which are ok to do local inference
They’re OK, but certainly a lot slower than with an Nvidia GPU.

isoprophlex | 1 year, 5 months ago

I run it on a 6 year old i5 box, no gpu, nothing. Granted, I'm fine with waiting 30 minutes for a generation, but indeed to be that's a core feature of the tool.
What's the power consumption on that ?seems like its more expensive to run than a gpu, but if you're not doing it often i guess it's not worth upgrading.

isoprophlex | 1 year, 5 months ago

60-70 W? Let's say 40 Wh per generation. Not so nice, but that works out to ~ 0.8 cents. So for a few generations a day, I can probably manage...
I have a similar setup: I'm running a Discord bot for family and friends that can do image generation. It's a 6-core AMD EPYC VPS with no GPU.

By using a SD 1.5 Latent Consistency Model (LCM) it generates an image in 3 steps or fewer. It takes about 30 seconds to generate a 512x512 image on that machine.

https://civitai.com/models/4384?modelVersionId=252914

>some major problems we’ve had with it. First is that it runs primarily on your own hardware

that's a feature not a bug

Man, commenters here are... grumpy. Yes it’s true you can do diffusion on a pi, if you’re okay waiting 4 years for your 50 steps to complete lol.

I want something like this to work, and am sure someone will eventually make the “photoshop of stable diffusion.” Personally my generations use custom fine-tunes and syncing 7gb checkpoints to a remote server can be a pain.

The bigger issue for Rubberband, I think, is that Civit weights are the killer app, and all Civit needs to do to win is make their generator a little more visible. Same with Civit’s trainer. They have an easy to use pipeline to go from training images -> generations, and they seem to run it close to cost, but it doesn’t even appear in the navigation until you sign in.

That doesn’t mean they’ll win, of course. Plenty of companies screw up, and maybe it’s not worth “winning” to begin with. See also: HuggingFace, which has Spaces, but is mostly used for the free bandwidth.

Anyway, best of luck!

[OP] jrmylee | 1 year, 5 months ago

Thanks Ed! really appreciate the feedback and support.

CaptainFever | 1 year, 5 months ago

A hosted version of image generation could be useful to me because my hardware isn't very good. However, one of the main reasons to use local generation in the first place is privacy and anti-censorship. I don't see a TOS on your service. How does it handle those cases?

[OP] jrmylee | 1 year, 5 months ago

This is a work-in-progress, will get back you
Similarly here - I clicked over and looked at the site and went first thing to find terms of use.. and nothing.

I have a couple of uses for such a thing, but perhaps you may offer one click deploy totally private use as you wish kind of thing,

I must assume a majority of people using comfyui or anything similar are going to be making stuff that is not allowed to be made on the more popular platforms.. your company may find itself quickly in the fire of many saying you should not be hosting X thing or Y thing or allowing people to do W or V.. and then there may be jurisdictions with whatever..

cloudlfare tried being a dumb pipe and espoused free speech until they did a 180..

So whtether it's investor pressure, or gov pressure or whatever it will be a thing at some point.

So also, what about export and backup options?

I would not create on another platform that can willy nilly pull the plug and leave you without your data, and the days of you pulling the plug are not far away imho.

Would use though, I hate setting this stuff up for myself, and I have a few other that can benefit from this sort of service - especially if it could be usage based at one price, and cold storage of training data at a lower price.

random thoughts.

[OP] jrmylee | 1 year, 5 months ago

yes definitely - thanks for the comments. We'll think about these points and get back to you later this week(via your contact in your bio)

lovegrenoble | 1 year, 5 months ago

Very interesting product. What is the license for the images generated with your service? Can they be used as designs for a commercial game?

[OP] jrmylee | 1 year, 5 months ago

We're currently working on our license/terms of service. Will have an answer for you shortly

smusamashah | 1 year, 5 months ago

For anyone wanting to try different styles of images using SD, I made a list of cheat sheets of styles a while ago https://gist.github.com/SMUsamaShah/218e602d508e891a123929ce...

I don't play with SD anymore mainly because now i find it exhausting to keep up with the tools and tech and low prompt adherence when trying to generate something specific. It's still fun if you are happy with whatever it gives you without putting in much effort though.

"Play around with Rubbrband for free. No credit card required." yes but WHAT is the pricing if I like it? I need it upfront. Is it credits? Subscription? Pay-per-image? How does it work?