I wasn't familiar with your project though, interesting stuff.
I'm trying to add more photography related features to Framedex but yeah there's so much we can do locally, exciting times.
iliashad 11 hours ago [-]
That's great, I checked your article when it was in front page because someone mentioned my project in the comments.
Good job for the article and the project. That's great, yes local models are getting better and better
LeonardoTolstoy 30 minutes ago [-]
What models did you use for the stages? I see Qwen2.5-VL-7B-Instruct mentioned as an advanced option, so I assume maybe Qwen2.5-VL-3B-Instruct by default (which is what I also use for a lot of stuff, it is incredibly good at "clean" OCR, but as you maybe indicate not the best at "describing a scene").
Also how much did OpenAI Whisper cost for the amount of video you processed, and did you try out any local models for that? This is something I want to look at for a personal project, but I don't really know if there is anything good enough locally to get accurate transcriptions.
For that project I'll say I don't see much degradation in embedding quality at much much worse quality than 720p (all the way down to 240p), which speeds things up considerably. Although I don't really do face or object detection, just scene embeddings. To me any process whereby it would take longer to process the video than watch it is probably a no go in general. Obviously a challenge for local-first analysis.
esjeon 11 hours ago [-]
> Then, run the frame analysis pipeline, which will divide the video into separate video scenes (1s each, or 1fps)
> (…)
> Frames analyzed 57,537
Aha, it makes total sense. This number sounds much more reasonable than “669 GB”, since the actual total size of processed frames would be like 10-30 GB.
(Not downplaying anything. Doing-at-home always requires some math on practicality)
> Total compute time 67h 40m 42s
I’m just curious tho — is there any paying options that can accelerate this kind of process? Just spin up GPU instances?
iliashad 11 hours ago [-]
> Aha, it makes total sense. This number sounds much more reasonable than “669 GB”, since the actual total size of processed frames would be like 10-30 GB.
The reason why is “669 GB” is the total raw footage size when I'm doing the video processing, I downscaled each frame to 720p to make the video processing much faster and I don't need full original quality in order to get accurate results (as far as I know and experiment with).
> I’m just curious tho — is there any paying options that can accelerate this kind of process? Just spin up GPU instances?
For now, I found that NVIDIA GPU for example RTX 3060 with 12GB Vram was much faster than my M1 Max. (still working on optimizing for speed and accuracy).
ngai_aku 7 hours ago [-]
What PAYG providers do people here recommend? Most powerful machine at home is an M1 MBA (16GB), so I too am interested in short term options where I can still benefit from the privacy of local models
villgax 4 hours ago [-]
Runpod
egorfine 1 hours ago [-]
Yep. Go to vast.ai, spin up a cheap GPU instance, add a bit of code to the project and let it run it finish in just a few hours for like ten bucks.
But it's not as fun as running local model right here on your computer on your own desk. It feels like magic.
justinram11 12 hours ago [-]
Something I've enjoyed more than I expected is Google and Apple photos sending me photo memories and compilations of various things in my life and my kids lives over the last decade.
I'm really bullish on taking more video of my kids, with the thought that it will become easier and easier for AI to put them into little compilations I can enjoy later.
mwelpa 1 hours ago [-]
I wish I could connect Apple photos to my Spotify account and have photo memories connected with songs I listened to at the time :)
alias_neo 36 minutes ago [-]
Music memories are the best.
I booted up my old PS3 from my uni days (20 years ago?) and found all of the music I had on it because I used it for everything at the time. Some seriously nostalgic music I'd completely forgotten about.
You don't mind Google using your kids to train their models and advertising algorithms?
Years from now they'll be getting "hey look at BIKE BRANDS' NEWEST CHEAP BIKE REMEMBER WHEN YOU USED TO RIDE BIKE BRAND BIKES"
satvikpendem 9 hours ago [-]
I think most people really don't care, and/or will just adblock those sorts of things when they do arrive.
marci 3 hours ago [-]
Don't worry. Most people spend most of their compute time on a phone, where you're ability to filter ads is way more enshitified.
JMiao 11 hours ago [-]
do you use android and ios, or is there another benefit to having personal media with both?
dave8088 6 hours ago [-]
I run both on my phone as a lazy (but flawed) backup strategy.
iliashad 11 hours ago [-]
Can you please elaborate more?
ngai_aku 7 hours ago [-]
I think most people are either in on Google or in on Apple whereas the OP indicated they have their media stored with both
robrain 15 hours ago [-]
DaVinci 21 has indexing built-in (AI IntelliSearch). Not to diminish the work you did, but this is now available to many users (probably only Studio users since it has AI in the name)
iliashad 15 hours ago [-]
Yes, I didn’t look at it. But does it upload your videos to the cloud or process them locally? And does it allow to provide custom faces data to help labeling faces in your videos ?
I think Adobe premiere pro have it as well but cloud processed
teovall 15 hours ago [-]
The AI features in DaVinci Resolve are all processed locally. It does not currently have face tagging.
robrain 14 hours ago [-]
Haven’t tried it yet, and I don’t know if it matches OP’s requirements, but the blurb says “You can even search for individual faces”
This is what took me from free to paid user, and it was well worth it.
iliashad 15 hours ago [-]
That’s great to know, thank you!
____tom____ 52 minutes ago [-]
I wonder how long it would take on faster hardware. I have ten times that much footage, but 67 * 10 hours is a lot of processing.
I might be better off getting something with a beefy GPU on AWS or Google cloud.
insumanth 4 hours ago [-]
I will be doing these things with local LLMs
Take a fast, small and powerful LLM running locally to index my personal data like images, videos, documents and enrich them and tag with the enriched metadata.
Want to group by people - Search tagged metadata and group it
What to search an image by description - tagged metadata
What to organize by anything - tagged metadata
This should (hopefully) put an end to my file clutter
nitin_flanker 2 hours ago [-]
I am in no way a tech savy person, don't know coding, don't know networking or AI much either. But I definitely want to have a system like this. An AI powered gallery / video repository that can help me find moments, people, colors, objects from 100s of 1000s of files.
Local LLMs sound so cool but I know they won't be easy to setup or use for common joe like me.
Mashimo 1 hours ago [-]
Immich can do part of this. For photos it does lm object detection and ocr for text. I think for video is currently only the first frame. It also has face / people detection.
And once set up it's easy to use even for non technical people.
Beijinger 18 hours ago [-]
Does it work for porn collections too?
pduggishetti 17 hours ago [-]
You'll need a lora for this, porn content rejection is heavy. Or you'll need a abliterated model, not sure if vision also works.
You might want to add something like yolo finetune to detect scenes + face recognition too.
vorticalbox 15 hours ago [-]
Vision still works perfectly fine in abliterated models.
avadodin 5 hours ago [-]
Just because they don't refuse it doesn't mean they are useful.
I found a few pornographic pictures on the web to hand to Abliterated Gemma4 12B(literally just to test this) and it needs pushing just to accept that people can be naked.
It didn't refuse but it also didn't provide useful descriptions such as "this is a pornographic picture of a woman".
> G4: There is a person lying down in a scientific context, if I had to guess they are a biologist in a classroom
> me: Is she wearing any clothes?
> G4: No.
Also, it is obsessed with penises —seeing them in compositions where there is only a female. I suppose it's been trained to ban dick pics or something.
Prompting may help some but 12B seems to be a bit worse than E4B with the vision/audio model at voice and text reading so maybe that one would do better.
pduggishetti 14 hours ago [-]
Never tried any of this for porn, just speaking out how I would go about it tbh!
dotancohen 13 hours ago [-]
For GP's purpose, can face recognition techniques be repurposed for, um, other body parts recognition? Sometimes the actresses are facing away from camera. There are exposed lips, if that helps.
sarjann 17 hours ago [-]
Asking the important questions
nntwozz 8 hours ago [-]
I was meandering through the comments about to leave the topic when my interest suddenly piqued upon reading the word porn.
Why it’s always the same question? Hahah. I posted my project over Reddit and I got the same one hahah
15 hours ago [-]
lifestyleguru 17 hours ago [-]
Last time I tried whisper, it hallucinated an elaborate conversation from sounds of slapping and moaning and it took minutes to spit every single line of it.
3eb7988a1663 16 hours ago [-]
Parakeet has been trained to detect non-voice sounds and exclude that from identification, so you might have better luck with that family.
dotancohen 13 hours ago [-]
If I remember correctly, the whisper documentation actually recommends to trim non-speech portions as the models halucinate heavily during those portions.
supertroop 17 hours ago [-]
Not sure if you’re being sarcastic but I think this is an interesting question. Would deep seek be useful here since it is local?
fibers 13 hours ago [-]
just because it is local does not mean it wouldn't reject explicit content. you can definitely try and find abilated models and can attempt to use unsloth or something similar to tune it properly.
kaycey2022 7 hours ago [-]
Is abliteration even necessary. While “playing around” I have noticed that most models are very strict only in the first prompt. The moment you get past that with a good turn, the next turn on you can get them to do _anything_.
okr 13 hours ago [-]
Depends how deep you wanna go.
MaxGL 2 hours ago [-]
[dead]
havercosine 5 hours ago [-]
Well done! I couldn't understand how you are building reels out of it via the agent. Is it some sort of AI tool calling that takes image links and builds a reel via some video editing tool ? Or +/- time delta around the timestamp returned from the indexed from a given query + join them together?
iliashad 5 hours ago [-]
Thank you! I'm using RAG, I have every video scene indexed individually in the vector database. When I'm asking the agent, it'll use an Ollama model to understand the request, use the available search tool (searching using transcription text, faces, visual, audio or combined) something like when you use Claude or Chat GPT it'll use the web search tool to find you info online. Then, I can filter out video scenes using the Ollama to better present accurate and unique video scene, then send those video results to Davinci Resolve using their API to create a video timeline using those video clips
WarOnPrivacy 16 hours ago [-]
I was surprised to learn that the
M1 Max CPU is an ARM/SoC, comparable to an 11th gen Intel i9
Do I have it right? Would Windows ARM performance be similar for those cpu?
It's also a bit apples (heh) to oranges for a handful of reasons, but most impactful
- "unified" ram makes all the system ram available as VRAM
- dedicated ai coaccelerator thingy
Both of these reasons allow the apple silicon chips to crush conventional cpus in these kind of AI model workload stuffs
No idea about what the windows arm stuff is capable of. I know they use Qualcomm snapdragon chips though.
voidmain0001 9 hours ago [-]
No comparison. M1 Max has 400GB/s RAM bandwidth while Snapdragon X2 Elite, the latest and greatest , has 228GB/s RAM bandwidth.
Rohansi 8 hours ago [-]
I don't disagree with your conclusion but the comparison of max bandwidth between the two SoCs is not enough. Neither of them will use all of that bandwidth doing AI work because the GPU will be compute limited. That's why dedicated GPUs perform so significantly better without having significantly higher bandwidth.
owldown 15 hours ago [-]
“Comparable” is maybe true if we are talking about single core performance, but for memory bandwidth, the M1 Max is about 8 times faster. Wider bus, lower latency, not even close.
iliashad 15 hours ago [-]
To your question, I can’t deny or confirm that because I didn’t tried it this project over a Windows machine yet or a machine with this config
cake-rusk 13 hours ago [-]
I have an RTX 5090 card but it only has 32 GB RAM, can something like this work on my machine?
iliashad 11 hours ago [-]
Yes, and it’ll result in much faster results than the ones that I did with my computer
I was looking for a solution for this issue of running docker containers over MPS and utilizing their GPU power. I think this project will be the solution for it, I’ll try it very soon and add support for it. Thank you, much appreciated
When trying to read this article, the main website was throwing errors to CloudFlare unfortunately
iliashad 4 hours ago [-]
Can you check again ? I'm not sure why it's show a cloudflare error
WhitneyLand 15 hours ago [-]
I’d like to see embedding of actual video clips become practical in this type of workflow.
Frame level embedding it covering a lot, but can miss out on a lot of action related searches.
iliashad 11 hours ago [-]
Sure, I'm using (https://huggingface.co/collections/Qwen/qwen25-vl) which can help me understand action like falling down, because I can provide for example 5 frames (down scaled to 720p) to understand what is happening in this part of the video
asdfasgasdgasdg 10 hours ago [-]
Cool build but the example videos you provide at the end are . . . not what I would hope for when thinking about the highlights of 2000+ videos of biking? For example the dog barking video only has one scene repeated two or three times and it's five seconds long?
iliashad 10 hours ago [-]
Fair enough, what would like to see as an example video and I would make it.
For the dog barking videos, those are only the video scenes that I have a dog barking sound in the video.
I'll keep adding more prompts and example videos, keep an eye for that
asdfasgasdgasdg 10 hours ago [-]
I don't have any preconceptions about specific content I want to see. I'd just think that so many hours of such cool adventures would have greater variety. It made me wonder if your AI really did such a good job of indexing it. It made me think maybe the tech isn't quite ready yet?
Did you ever visit crazyguyonabike.com? A long time ago I had the pleasure of following the journey of a friend of a friend of a friend on that site:
I would love your feedback and suggestions for new improvements or features you wanna have, either in the source available version, the desktop app or blog post itself?
PreownedPlaid 10 hours ago [-]
this is really cool. was looking to do something similar on mbp 64gb
iliashad 9 hours ago [-]
That's really great, thank you!
synergy20 9 hours ago [-]
can vlm be used instead or it's too heavy and slow
15 hours ago [-]
m3kw9 15 hours ago [-]
Grab frames, lower res, classify, combine meta data. Write to sql
iliashad 15 hours ago [-]
Not really. Grab frames, lower res, classify, combine metadata, transcribe the audio, convert those data (text, visual and audio) to embedding, save them over a vector DB and SQL DB. Which helped me to do semantic search, RAG, search using a screenshot of the video to find the exact the moment in the video plus search using an audio file as well. And other features unlocked with vector DB
ingvay7 13 hours ago [-]
Really cool work and workflow. strongly prefer this kind of local, open pipeline that i control over a dependency on Adobe tools and lock ins.
iliashad 11 hours ago [-]
I agree with that, thank you for your feedback. Also, maybe you're not a video editor and you just wanna search your videos. The video editing integrations are optional and you have full control. You can switch between Adobe Premiere Pro, Final cut Pro or Davinci Resolve
ingvay7 10 hours ago [-]
cannot wait to incorporate this to my workflow. thanks
nyxtom 14 hours ago [-]
Now this ^^ is an awesome use case!
iliashad 11 hours ago [-]
Thank you, would like to know your use case for this kind of project and which prompt you want to genearte ?
Mawr 9 hours ago [-]
> Many of the videos I captured amazing moments, and sometimes it's kind of hard to watch the full videos to get those moments.
Yep. I had the same problem.
> Then, run the frame analysis pipeline [...] I have a face recognition plugin using my custom faces data, object detection, on-screen text, shot type, and scene description [...] we will have three vector DB collections that have all the information about our videos, like video location metadata, camera name, faces recognized, objects detected, on-screen text, transcription, description of each scene, and many more [...] we can get better indexed data if you use the advanced mode indexing to use the Qwen2.5-VL-7B-Instruct model to understand and describe your video much better, but at a slower indexing speed
Yeah, uhm... ok :)
If anyone else has a similar problem, the real solution is as follows:
1. When recording, if you witness an interesting moment worth saving later, press the power button — this will mark the current moment in the video as a chapter.
2. Find the chapters later when editing and cut them into clips.
3. You're done :)
This has two main benefits over the insanity above:
1. It's trivially simple instead of insanely complex and inefficient.
2. It will reliably catch all the stuff you find interesting, since you're the one doing the marking.
The downsides:
1. Doesn't work retroactively.
2. It may miss interesting stuff if you miss it at the time as well.
3. Only works for this use case.
4. Nerds won't salivate over your usage of cutting edge tech.
Noumenon72 8 hours ago [-]
What tool has this "press power to mark chapter" feature?
tredre3 6 hours ago [-]
The GoPro, it's called HiLight Tag.
PixComicOS 6 hours ago [-]
[flagged]
dadachi 9 hours ago [-]
[flagged]
knightops_dev 15 hours ago [-]
[flagged]
knightops_dev 15 hours ago [-]
[flagged]
tosief 8 hours ago [-]
[dead]
GreenSalem 8 hours ago [-]
A lawyer I know who specialises in rape,
and is excellent at getting the obviously guilty exonerated,
lost a case last year because of GoPro videos.
Her client was recording while committing the abhorrent crime.
The criminal would otherwise have got off.
From my perspective, the GoPro camera produced a good outcome.
Still, one has wonder why anyone to record their criminal actions.
Yiin 8 hours ago [-]
word "her" in this context gave me heavy feelings, what makes one to pick such a career move...
GreenSalem 3 hours ago [-]
Beggars cant be choosers.
She would rather have done corporate law but did not have the academic credentials or the networks needed for a job at the likes of Latham Watkins or White and Case.
Still it is good for society that criminals get the worst lawyers to defend them.
https://news.ycombinator.com/item?id=48222733 https://blog.simbastack.com/indexed-a-year-of-video-locally/
I wasn't familiar with your project though, interesting stuff.
I'm trying to add more photography related features to Framedex but yeah there's so much we can do locally, exciting times.
Good job for the article and the project. That's great, yes local models are getting better and better
Also how much did OpenAI Whisper cost for the amount of video you processed, and did you try out any local models for that? This is something I want to look at for a personal project, but I don't really know if there is anything good enough locally to get accurate transcriptions.
For that project I'll say I don't see much degradation in embedding quality at much much worse quality than 720p (all the way down to 240p), which speeds things up considerably. Although I don't really do face or object detection, just scene embeddings. To me any process whereby it would take longer to process the video than watch it is probably a no go in general. Obviously a challenge for local-first analysis.
Aha, it makes total sense. This number sounds much more reasonable than “669 GB”, since the actual total size of processed frames would be like 10-30 GB.
(Not downplaying anything. Doing-at-home always requires some math on practicality)
> Total compute time 67h 40m 42s
I’m just curious tho — is there any paying options that can accelerate this kind of process? Just spin up GPU instances?
The reason why is “669 GB” is the total raw footage size when I'm doing the video processing, I downscaled each frame to 720p to make the video processing much faster and I don't need full original quality in order to get accurate results (as far as I know and experiment with).
> I’m just curious tho — is there any paying options that can accelerate this kind of process? Just spin up GPU instances?
For now, I found that NVIDIA GPU for example RTX 3060 with 12GB Vram was much faster than my M1 Max. (still working on optimizing for speed and accuracy).
But it's not as fun as running local model right here on your computer on your own desk. It feels like magic.
I'm really bullish on taking more video of my kids, with the thought that it will become easier and easier for AI to put them into little compilations I can enjoy later.
I booted up my old PS3 from my uni days (20 years ago?) and found all of the music I had on it because I used it for everything at the time. Some seriously nostalgic music I'd completely forgotten about.
Years from now they'll be getting "hey look at BIKE BRANDS' NEWEST CHEAP BIKE REMEMBER WHEN YOU USED TO RIDE BIKE BRAND BIKES"
I think Adobe premiere pro have it as well but cloud processed
https://www.blackmagicdesign.com/products/davinciresolve/wha...
I might be better off getting something with a beefy GPU on AWS or Google cloud.
Take a fast, small and powerful LLM running locally to index my personal data like images, videos, documents and enrich them and tag with the enriched metadata.
Want to group by people - Search tagged metadata and group it What to search an image by description - tagged metadata What to organize by anything - tagged metadata
This should (hopefully) put an end to my file clutter
Local LLMs sound so cool but I know they won't be easy to setup or use for common joe like me.
And once set up it's easy to use even for non technical people.
You might want to add something like yolo finetune to detect scenes + face recognition too.
I found a few pornographic pictures on the web to hand to Abliterated Gemma4 12B(literally just to test this) and it needs pushing just to accept that people can be naked.
It didn't refuse but it also didn't provide useful descriptions such as "this is a pornographic picture of a woman".
> G4: There is a person lying down in a scientific context, if I had to guess they are a biologist in a classroom
> me: Is she wearing any clothes?
> G4: No.
Also, it is obsessed with penises —seeing them in compositions where there is only a female. I suppose it's been trained to ban dick pics or something.
Prompting may help some but 12B seems to be a bit worse than E4B with the vision/audio model at voice and text reading so maybe that one would do better.
ref: https://www.cpubenchmark.net/compare/4585vs4245/Apple-M1-Max...
- "unified" ram makes all the system ram available as VRAM - dedicated ai coaccelerator thingy
Both of these reasons allow the apple silicon chips to crush conventional cpus in these kind of AI model workload stuffs
No idea about what the windows arm stuff is capable of. I know they use Qualcomm snapdragon chips though.
comes with some nifty features like NLE- integrations, people search, MCP, API etc
Disclaimer: one of the co-founders
Other comments mention davinci resolve has this built in. How would you compare the two?
When trying to read this article, the main website was throwing errors to CloudFlare unfortunately
Frame level embedding it covering a lot, but can miss out on a lot of action related searches.
For the dog barking videos, those are only the video scenes that I have a dog barking sound in the video.
I'll keep adding more prompts and example videos, keep an eye for that
Did you ever visit crazyguyonabike.com? A long time ago I had the pleasure of following the journey of a friend of a friend of a friend on that site:
https://www.crazyguyonabike.com/doc/?doc_id=2405
Stuff like that I guess?
Yep. I had the same problem.
> Then, run the frame analysis pipeline [...] I have a face recognition plugin using my custom faces data, object detection, on-screen text, shot type, and scene description [...] we will have three vector DB collections that have all the information about our videos, like video location metadata, camera name, faces recognized, objects detected, on-screen text, transcription, description of each scene, and many more [...] we can get better indexed data if you use the advanced mode indexing to use the Qwen2.5-VL-7B-Instruct model to understand and describe your video much better, but at a slower indexing speed
Yeah, uhm... ok :)
If anyone else has a similar problem, the real solution is as follows:
1. When recording, if you witness an interesting moment worth saving later, press the power button — this will mark the current moment in the video as a chapter.
2. Find the chapters later when editing and cut them into clips.
3. You're done :)
This has two main benefits over the insanity above:
1. It's trivially simple instead of insanely complex and inefficient.
2. It will reliably catch all the stuff you find interesting, since you're the one doing the marking.
The downsides:
1. Doesn't work retroactively.
2. It may miss interesting stuff if you miss it at the time as well.
3. Only works for this use case.
4. Nerds won't salivate over your usage of cutting edge tech.
Her client was recording while committing the abhorrent crime. The criminal would otherwise have got off.
From my perspective, the GoPro camera produced a good outcome. Still, one has wonder why anyone to record their criminal actions.
She would rather have done corporate law but did not have the academic credentials or the networks needed for a job at the likes of Latham Watkins or White and Case.
Still it is good for society that criminals get the worst lawyers to defend them.