Show HN: Image background removal without annoying subscriptions (pixian.ai)

383 points by jacobn ↗ HN
Hi HN,

Removing the background from images is a surprisingly common image processing task, and AI has made it really easy. The technology has come a long way since segment leader remove.bg launched here on hn in Dec 2018 [1]. Chasing remove.bg's success, a legion of providers have come on the market offering varying levels of quality & service.

Despite there being a large number of competing services, most still price for very high (~95%?) gross margins. Furthermore, subscriptions make the effective unit price a lot higher than the list price for infrequent users, and requires effort & attention to ensure you're getting value for money. This has prevented a host of use cases (e.g. infrequent professional / hobbyist) and business models (e.g. ad-supported websites & mobile apps).

We see this as an opportunity where we can jump to the market's logical conclusion to gain market share and build goodwill: cost-plus PAYGO pricing, i.e. the "S3 pricing model".

So we've built yet-another image background removal service ( https://pixian.ai - introductory post 6 months ago [2], a ton has been improved since then) but with a couple of twists:

1. Quantified quality comparison (90-120% of remove.bg, depending on image category), free for you to check your own images so you can make an informed choice.

2. Customer-friendly pricing (PAYGO @ 1-10% of competitors' subscriptions) with a generous free tier (and free while in beta).

3. A novel API result format: Delta PNG [3], which offers excellent latency & bandwidth savings. Especially useful for mobile apps.

4. Operational transparency: actual volume & latency metrics public, with more coming soon (all API providers should be showing this).

There's of course more to it than just price and we see several sources of differentiation in this market: quality, price, capability, reliability, latency, and goodwill.

As a new entrant we're looking to meet-or-beat the quality bar; beat on price, capability, reliability and latency; and to build up goodwill over time.

Our goal is to make it a no-brainer for new accounts to choose us, and to provide the tools and guidance necessary for existing accounts to make the switch with confidence.

We'd love for you to try it out and to hear your thoughts!

https://pixian.ai

[1] https://news.ycombinator.com/item?id=18697601 [2] https://news.ycombinator.com/item?id=33439405 [3] https://pixian.ai/api/deltaPng

119 comments

[ 1.0 ms ] story [ 173 ms ] thread
Bookmarked, will use.

If I can upvote twice, I would.

>Philosophy >While AI is rapidly transforming the way we as a society do business, AI itself is changing even faster. What was >cutting edge only a few years ago is now rapidly becoming commoditized.

>We choose to accept and accelerate this reality.

>We therefore see ourselves less as a tech startup, and more as an outsourced MLOps extension to your engineering team. >Our goal is to be more 'S3' and less 'Adobe' for state-of-the-art AI image processing.

Well done and well said.

Wish many hearths won

The comparisons show a legend but the keys are all the same colour?

Is that intentional?

Ah, it's not, will fix, thanks for pointing that out!
seems to be fast and accurate. thx will use.
Looks great. the "see it in action" might look better with a checkerboard transparency background instead of white.
Agreed — it's hard to see the edge detection between the sleeve and the sand. From what I can tell, it looks pretty good!
I tried with a random photo from my mobile phone, it complains that the image is too large. This is bad UX to me. The webapp could resize it for me
Most images have excessive margins around the foreground. By allowing you to crop it instead of just shrinking it you get an effectively higher resolution result.

If you don't want to manually crop it, just press "ok".

That said we could probably improve the messaging in that dialog, thanks!

(comment deleted)
You can do this yourself for free, using the U2NET model... Either client-side or server side.

https://github.com/xuebinqin/U-2-Net

Pretty trivial thing to implement.

For a Linux user, you can already build such a system yourself quite trivially by [...]
sending your photos to your iphone, tapping and holding on the object, and taping the share option that pops up over the foreground item, then choosing save image. then you can send that new image with the background removed back to your linux system.

How to send and receive image between linux and iphone is left as an exercise for the reader.

I don't think trivial means what you think it means
Several hundreds or possibly several thousands of lines of python code (I just checked main.py and one from a subfolder) doesn't sound trivial to me.

That said, I am not negative, only pointing out that at least for me (as a admittedly non-native speaker,) trivial is not a word I would use to describe it.

The animation when you drag a photo is extremely nice
The problem I have with these is that they do 98% of the removal well, but flub the last 2%. This has been the case with every one of the online and app flavors I have used (note I refuse to subscribe to Adobe.)

My use case is mineral photos, as it turns out. And I would be very surprised if the AI had been trained on these. The sad thing is -- mineral photo backgrounds tend to be very simple and smoothly-varying. Should be a slam dunk. Ah well.

A one-shot background remover doesn't give you the opportunity to interact with it and suggest that it got things wrong here and there.

Yes, I tried one of my mineral photos, and the app made several errors that it shouldn't have, as the foreground was clearly distinct from the background.

I don't know if I'm allowed to post a photo link, but here it is: https://imgur.com/a/V9H1pRH .

Smooth background is likely worse because it carries not much background semantic. A wood table might do better there. Anyway the white box approach does so much more to item display photos, including all around soft lighting, background removal ai still are quite a way from those.
Have you tried Segment Anything from Meta? There are a bunch of click guided segmentation models out there.

https://segment-anything.com/demo#

It worked fine on your example with a couple of clicks.

As someone who is actively working on using SAM, I would say that it leaves a weird border (2-5 pixels wide) of the cut out object so depending on the tasks, it may or may not be suitable. And in the demo, the server returned a way better image embedding than the open source one from Vit_h. There are a lot of issues on github talking about this. So take what you see on the demo page with a grain of salt.
iOS photos has this built-in AI thing where you double tap a photo and it separates the subject into a png with alpha. Imgur made available only a very low res version of your shot but still this is what I got, seems pretty good actually: https://wormhole.app/mbrY3#ab0qbAz7p26SYR0CEONDpQ

(also, cool shot!)

(I'm a dev on the project.) We have another background-removal service, ClippingMagic.com, that is built around an editor to let you fix the errors in the automatic result. You may want to give it a try for your mineral photos.
Love it. It also provide highres output out of the box. 100% I will be a customer, albeit my yearly needs are in single digit range.
Adding it to HomeScreen as a WebApp on Android, creates a link without a logo. Could you make it appear like other WebApps, like Squoosh?
There's some mention that this can be self-hosted, but it's not that easy.

Even a pre-trained model and self-hosted API like RemBG[0] performs a lot worse than this "pixian" service does.

https://huggingface.co/spaces/KenjieDec/RemBG [0]

All in all, I'd say they put a LOT of effort into scaling (it's FAST) and the models (they're extremely accurate).

you have conflict of interest with them? very generous review
Not at all affiliated. I have used these kind of APIs for ecommerce for about +15 years and are always looking for the best quality/cost ratio and this one really impressed.
Tried couple of pics, it's way better than other services and pricing of it seems really good! Thanks for the app!
How does it compare to Command+Shift+K in Preview on Mac OS?
You can also right-click -> actions -> remove background on recent (not entirely sure when it was added) Mac OS versions
Exactly & doing it in browser is also not that far fetched
pretty good! will use it to generate WhatsApp stickers. perhaps you could use Enterprise sales as a way to subsidize consumers with 1 use/year needs? :)
Problem is there are so many cheap companies out there it seems you can easily run yourself into a "sybil resource exhaustion attack" by employees working their way around the "free for consumers paid for businesses" rule.
I appreciate the free tier with resolution-limited downloads. And you've priced the middle tier right, because I wouldn't hesitate to drop five bucks even if I didn't know how long it would take me to churn through all 250 redemptions.
>Retention

>Records not associated with an account are deleted or anonymized within a year of creation. Image processing records associated with an account are retained indefinitely.

>We retain them on your behalf so that you can view and download them as you wish, and to provide customer support.

Thank you, but no thank you.

(comment deleted)
Where did you read that?

https://pixian.ai/remove-image-backgrounds says:

> Right now, we retain images and results for five days after they are uploaded, after which they are permanently deleted. Please note that our data retention policies may change over time, and this current policy does not bind us in the future, or require your affirmative consent to change.

https://pixian.ai/policies/terms says:

> User Submissions and any associated Results will expire 2 weeks from the point of upload.

> The Service may provide users with the option to delete User Submissions to have them expire before their normal expiration time.

> Expired User Submissions and any associated Results are subject to deletion or retention at the Company's sole discretion.

On their Privacy Policy under Image Processing Records

https://pixian.ai/policies/privacy

(I'm a dev on the project.) The privacy policy is an old and generic one that we use across a bunch of sites. It should be updated. Our retention policy on this site is as stated on the front page FAQ. After five days, the records are deleted.
I just assume nothing is deleted when sent over an electronic medium. Makes life easier with that assumption.
The main challenge and deficiency with this and other services like this is that they assume some content type to be foreground - here it is likely "portrait" style people - it's hard to say as this information isn't included in the UX. Does it handle other foregrounds, such as cars? Who knows.

A better solution is to have controls, at least in optional manner, which the user can use to scribble examples of the foreground(s) and backgrounds.

Absolutely, that was our original take on https://clippingmagic.com where the editor lets you easily make changes to the result.

But almost all API users and most regular users just want a result and to not have to futz with it, hence this offering.

Doesn't seem to accept .heic photos, which is really odd given that the majority of new photos in the US are probably created in that format.
If anyone is already running auto1111, or simply uninterested in paying, there's an addon that does this very well available here https://github.com/KutsuyaYuki/ABG_extension, additionally I've had very good results using the masks generated by Facebook's SAM, which is also available as an addon here https://github.com/continue-revolution/sd-webui-segment-anyt...

auto1111 has a one click installer and these extensions can be installed by going to the "extensions" tab and pasting the github URL into the "install from URL" box. auto1111 available here https://github.com/AUTOMATIC1111/stable-diffusion-webui

I made a UI for Meta's SAM that works very well for background removal: https://github.com/dibrale/samist

It runs from a python script.

Great tool!

Some explanations on the differences of these models would be nice for a noob.

If I recall correctly, the difference between the SAM models is just a parameter number versus accuracy tradeoff. I have the parameter numbers listed under 'Installation', but the relative quality of the models would be task-dependent and subjective.

I would think that part of the motivation for releasing the smaller models in addition to the larger ones would be use in video image segmentation and mobile filters. The smaller models might actually be more fit for purpose with regard with regard to those applications than the biggest one. However, I'd reccommend the biggest model (vit_h) for desktop or laptop image processing.

I have set up auto1111 just as a hobby to make art (if it's okay to call it that) for myself.

Surprisingly (?) my convoluted setup is slowly becoming this actually useful toolkit for various tasks I sometimes need to do.

macOS does this natively. There's models that can run in the browser literally so any premise on discounting doesn't exactly pan out lol
(I'm a dev.) Not everyone has macOS, and running a DL model in your browser is not exactly mainstream. Which is why segment leader remove.bg gets an estimated 35M in monthly organic traffic. Also, we believe our results are significantly better than those offered by the open-source models we've seen.
I'm talking about lightweight ONNX models which for this task is more than enough.
Ok; two things:

1. For "objects" and "artwork" it says that quality is above 100% (108% and 119% respectively), which is weird and doesn't inspire confidence? It's also unclear generally what those percentages mean.

2. When trying this from work it says "Unable to connect to the worker. Is your firewall or proxy blocking WebSockets?" -- it's possible that the firewall is the culprit but there should be a workaround? (All methods give the same result, drag'n drop, ctrl+v, or picking from the explorer).

You can click through to see the report. Above 100% means they performed better than the competition, not they handled every picture well.

> Pixian.AI had 241 images that were rated good, or not rated and identical. That's 87.0% of your 277 images.

> The competitor had 201 images that were rated good, or not rated and identical. That's 72.6% of your 277 images.

> Pixian.AI achieved 241 / 201 = 119.9% of the competitor's performance

> The report is based on a set of user-provided images. We then reviewed the services' respective results separately and rated them as either good or bad. The comparison was done in a blind manner, without labels indicating which result was Pixian.AI's and which was the competitor's. We then tallied up the results and produced this report.

https://pixian.ai/comparisons/cwsslt8d78zl7vq/share/5e51cd3d...

On iOS you can long tap on a subject and then save it as an image. Mac OS has a similar feature on right click and Preview :)
I will give this a try, but I’ve been happy to shell out for a few credits on remove.bg when intricate hair strands were involved [in isolating portraits]. Whatever they do just works.
remove.bg is great, and they handle hair really well. But so do we! Give us a shot next time ;)