---
title: "Improve Media #2"
video_id: iBhpIpokajE
url: https://www.youtube.com/watch?v=iBhpIpokajE
workflow_page: https://workflows-of-ai.com/workflow/improve-media-2
language: en
captions: generated
---

0:01 hey there um so I just did a video about
0:04 how you can improve the media entity
0:08 when it comes to the
0:09 document media type I just want to show
0:13 you how to do the same for audio and to
0:17 do the same for a image media so you get
0:20 that idea as well um this is building up
0:23 on the other video so if you haven't
0:25 seen the other video please look at that
0:27 first um to get some idea what I was
0:30 doing there um so where I will start is
0:35 that for the Audio I need this module AI
0:38 interor deep CR for the image we're
0:40 going to just going to use
0:43 um uh open AI St 3 in this case in
0:47 theory we also have the dream Studio
0:49 plugin so if you prefer stable diffusion
0:51 you can use that but I will just use it
0:53 and this is already installed in the old
0:56 video but for the audio we need the Deep
0:58 CR so we will just go in here and run
1:03 ddev to uh get this
1:07 module
1:10 um so we're downloading the Deep module
1:14 and then we will just go here to extend
1:17 I I can just preview so this is just
1:19 some audio about difference between TCP
1:22 and
1:22 UDP whenever a computer wants to
1:25 communicate with another computer the
1:27 communication so that's just a previous
1:30 someone talking basically about what's
1:31 the difference between TCP and UDP for
1:33 the image I will just Showcase with this
1:36 Lamborghini
1:37 car
1:39 um right so uh let's see we have ai
1:44 interpolator deep cram install oh sorry
1:47 download it we install
1:49 it so we go to configuration web
1:51 services decam config to set up this you
1:54 have to log into to deep CR I'm logged
1:56 in here so I'm just going to create a uh
2:00 API key test API
2:02 key uh well can set the
2:07 owner set so it expires later tomorrow
2:11 so I don't forget to remove it and
2:13 someone of you steals something
2:15 here and I just create a key here just
2:19 going to copy this key and save it here
2:22 so that's to basically authenticate to
2:25 deep gr so we have deep G set up now oh
2:29 that's not have so many things here so
2:31 what we're going to do is take the media
2:33 type audio in this case now then and
2:35 improve it I can just start by uploading
2:38 the audio so you see there's no meta
2:40 Fields uh connected on this at all uh we
2:47 just going to take this TCP versus UDP
2:49 upload it and there's yeah as you can
2:52 see there's no metadata basically so we
2:54 need to add Fields here so we're going
2:56 to
2:57 do uh I'm just going to open open the
3:00 document field so I can copy some of the
3:02 prompts actually we're going to reuse
3:05 existing Fields right so uh we can
3:10 actually use the media document uh sorry
3:13 no no no media document I mean the
3:16 original text I mean it's it's a
3:18 transcription in this case so maybe we
3:20 should call transcription but I will
3:22 just reuse
3:24 this uh and in this case I'm going to
3:26 use deep gr
3:28 transcribe right and we're going to use
3:32 uh I guess in this case it's well let
3:37 just take General there's no monologue
3:40 there and we use the audio file because
3:42 that's the file field we have right and
3:46 uh we just set this to batch mode this
3:48 is the first thing we need to do so this
3:50 is just going to replicate basically
3:52 what we did with the other things so
3:53 we're going to add search words here
3:56 enable in AI interpolator we can just
3:59 copy this
4:01 prompt that we have
4:03 here oh sorry that's the help text here
4:07 from uh oh no right right what we want
4:10 to do is we want to start with the long
4:12 description that's how we did it in the
4:14 old one so we create a long description
4:16 and then we generate everything from
4:18 that right so uh I'm just going to take
4:22 the long description
4:24 here so we do open AI text long from the
4:28 original text and we're just creating a
4:31 summary I'm doing this very fast now you
4:34 can look at the other video to
4:35 understand exactly what I'm doing
4:38 because this is basically uh doing the
4:41 same thing over again so we use gp4
4:43 turbo preview here then we're going to
4:47 uh oh the search word actually got
4:49 copied there so we have to set it
4:51 [Music]
4:52 up so we're just going
4:56 to basically uh create that short string
5:00 here so we take that from the long
5:02 description we have a prom 10 to 15 15
5:05 search words um this has to be 10 1002
5:09 because it has to happen after we have
5:12 transcribed it and made a
5:15 summary uh turbo preview we save
5:20 that and we're going to also do the
5:24 short
5:26 description so we're so we're basically
5:28 going to have the say metadata Fields
5:30 here
5:31 right uh enable that we take open AI
5:36 text long from the long description we
5:39 use as context and we're going to create
5:40 a two sentence
5:42 summary same thing here this has to have
5:45 1002 batch
5:48 gp4 turbo preview and we save that and
5:52 the last thing is we're going to do
5:54 the uh header image
5:58 prompt that creates the header image
6:03 right or preview image that's what I I
6:07 wrote that a little bit wrong so open AI
6:10 string long take from the long
6:13 description generate a prompt for a
6:17 image 2002 batch Modo GPT for
6:22 Turbo preview save
6:25 it and lastly but not least oh I hope so
6:29 supposed to
6:30 reuse we're going to do um the preview
6:34 image enable a inator d e from the
6:39 header image prompt we just do the
6:45 context and this one we need to do 103
6:48 because it has to happen after the
6:50 header image is done to a little bit
6:53 better
6:55 quality save it and then oh sorry sorry
6:59 we go back to the
7:01 media just going to
7:03 re um resave this and you will see that
7:07 all these fields will basically be
7:09 automatically filled out now then by the
7:11 A
7:12 interpolator and then we're going to do
7:14 the same for image well we're going to
7:16 do it a little bit different because for
7:18 the image of course we don't need to
7:20 create a preview image the preview image
7:22 is just the actual image um we
7:26 also don't necessarily need to
7:31 uh have a base text field we can create
7:34 the long uh description directly from
7:37 the image right because the image itself
7:40 doesn't uh doesn't have any text in
7:46 it
7:48 so this might take some time while we do
7:51 that we can actually start right away
7:53 with the image just to set it up the
7:55 same way so we have a scene how it gets
7:57 set up uh so in this case we're going to
8:00 start with long
8:03 description that's still running enable
8:06 AI interpolator but in this case we're
8:08 uh going to do let's see
8:12 here
8:14 um text from docu oh right right so this
8:17 is a little bit special case because uh
8:19 the AI interpolator forces you to use in
8:21 context unless then you use the advanced
8:25 settings uh Advanced uh mode which
8:28 requires that you have the token module
8:30 installed so I'm going to install that
8:32 now and I'm going to show you how I can
8:34 do this so I just reload because the
8:38 long description is there on the image
8:40 we enable theator and then we can do uh
8:44 oh this doesn't do advanced mode yes it
8:47 does open AI text long advanced
8:52 mode uh because what we can do now
8:56 instead in this case we don't really
8:59 care about uh context of some other text
9:03 right so we can just write in the prompt
9:06 based on the image I just still need to
9:10 see how I wrote The Long description so
9:13 we we can still copy this but we're not
9:15 going to copy the context So based on
9:18 the uh image create a summary of to say
9:22 one paragraph describing the image
9:26 right uh and then it's important that we
9:29 use one of the uh Vision models so we're
9:32 going to do gp4 Vision preview and then
9:35 we can say that we use the image field
9:37 as the actual image we upload right set
9:40 batch mode and save this is then going
9:43 to take the uploaded image and create a
9:45 description of it and then we can just
9:47 do the same thing that we did before we
9:50 do create a short description here just
9:53 copy that
9:58 one
10:00 so P AI text long from long
10:05 description and this just has to have
10:08 one higher so oh 101 in this case GPT 4
10:14 Turbo I
10:16 want save it and then we also get the
10:20 search words in the case with the search
10:23 words I think it actually makes more
10:25 sense to uh for AI to look at the image
10:29 because the long description might just
10:31 describe it very abstractly while the
10:33 search word you might want to have very
10:36 specific Keys like what colors and stuff
10:39 right um so because of that we actually
10:43 are going to base that on the image in
10:45 this
10:46 case so I just copy this but I'm not uh
10:51 let's see here oh right right no no we
10:54 have to uh base it on the long
10:57 description so this is something that
10:58 will will come in the next version of
11:01 the open AI module that you can use
11:03 short text to to uh get from
11:08 the um get from the image you can't do
11:12 that actually at the moment so it only
11:14 works currently with um no it does work
11:19 sorry I'm I'm I'm not I'm just not
11:23 thinking right now sorry here so what we
11:25 want to do is sorry Advance
11:27 mode So based on the image I don't even
11:32 know my own model my own code sorry
11:36 based on the image suggests 10 to 15
11:38 search words that are
11:40 unique
11:42 descriptive and important when searching
11:44 for this document single out
11:48 colors and
11:53 forms no not single out colors and forms
11:58 are
11:59 important let's right like that and this
12:03 can be uh this can be 100 basically
12:06 because it's based on the image right
12:08 which does exist so we're going to do a
12:12 gp4 vision preview in this case because
12:15 we once again we need a vision model do
12:17 it on
12:20 image uh then preview image we don't
12:22 need header image we don't need original
12:24 text we don't need right in this case so
12:27 this should actually be enough
12:29 so I'm just going to close this now so
12:31 we're going to start with this one that
12:33 we did before the audio field and you
12:36 see that the whole transcription is
12:39 there uh there's a long description and
12:42 there's a short description about TCP
12:44 UDP right uh and of course TCP and UDP
12:49 it's important search words in this case
12:51 but also a lot of other important search
12:54 words uh yeah visualizing TCP UDP is
12:58 probably not the easiest let's see what
13:00 it did
13:01 here yeah okay uh yeah we I should
13:06 probably add to the prompt that it
13:07 shouldn't try to do text because it's
13:09 really bad at doing text still but you
13:11 see this like a network it's kind of
13:14 symbolizing
13:16 youp so let's just do the last thing
13:19 here now then we're going to upload our
13:21 image of our uh Lamborghini here right
13:26 uh Lambo and then all this should just
13:29 hopefully be filled out so the search
13:31 word the short description the long
13:34 description so let's see if that works
13:37 as
13:38 well and then um soon there's going to
13:42 come plugins to do this on video as well
13:46 well first of all deep gr soon in the
13:48 next version it's going to be able to
13:50 take video as input and if you have
13:52 something like f fmeg running on your
13:54 server it's going to be able to convert
13:56 this and uh get the trans transcription
13:59 of of the
14:01 video but uh after that there are models
14:07 that can read video already so hopefully
14:09 this will be integrated soon so you can
14:13 see here if we click edit that it has
14:17 done a long description a short
14:19 description and also some search words
14:22 red sport car right that's what it is
14:26 Sleek body work and so on
14:29 so yeah that's it so video is not there
14:34 yet but if you want to have a automated
14:36 process where you get
14:38 descriptions uh where you get search
14:40 words or even preview images in those
14:43 cases where it's not the image from this
14:45 uh get go then uh this is how you do it
14:49 automized so thank you for
14:52 watching