How to Enhance Jewelry Product Pictures Using AI | Train with Flux LoRA

00:10:58
https://www.youtube.com/watch?v=mlEv3iqdNsk

概要

TLDRIn this video, the creator uses AI to enhance photos of a ring, taken with an old Samsung Note 1 phone, by testing various AI platforms to create commercial-quality images. The creator compares three services: fall. a, tensor art, and Dojo pixel, detailing each platform's process and effectiveness. Fall. a was found to be most efficient, though expensive, for transforming images to look natural and seamless. Tensor art was complicated and less effective, often producing unsatisfactory results. Dojo pixel, with a monthly fee, offered decent image creation but still had notable shortcomings, such as exaggerated effects. The creator highlights the difficulties and costs associated with these AI services, emphasizing the need for user-friendly interfaces and lower pricing. Despite challenges, AI helps simplify the creation of complex imagery, such as adding effects like red dust to images, saving time compared to traditional photography setups.

収穫

  • 📸 Using AI to enhance product images can greatly improve picture quality.
  • 💰 AI platforms for image enhancement tend to be expensive and complex.
  • ⏳ The quickest user experience was found with fall. a, offering more natural results.
  • 🤖 AI can struggle with generating precise elements, such as fingers.
  • 🔧 Overtraining models can negatively impact image results.
  • 🌐 Understanding the AI tool's interface and settings is challenging for non-experts.
  • 📈 More user-friendly interfaces and affordable pricing could help businesses leverage AI tools effectively.
  • 🖼 Quality image generation often requires several attempts and refined settings.
  • 🤔 Simplifying AI jargon would make tools more accessible to general users.
  • ⚡️ AI saves time in creating elaborate photo setups, offering easier prompt-based solutions.
  • 🛠 Despite advancements, AI technology demands ongoing improvement and adaptation.

タイムライン

  • 00:00:00 - 00:05:00

    The speaker introduces an old Samsung Note 1 phone, emphasizing its condition, and expresses interest in using AI to improve pictures for commercial use, noting their lack of expertise in AI. They intend to test three websites: Fall.a, Tensor Art, and Dojo Pixel, all of which offer expensive AI services. The speaker details the process with Fall.a, involving renaming images with a trigger word and notes the time-consuming training process. They highlight challenges in finding correct prompts and settings but also express being impressed with the results, despite errors such as distorted fingers.

  • 00:05:00 - 00:10:58

    The speaker critiques Tensor Art, the second website, describing it as complicated, possibly Chinese-origin, and least effective, highlighting the exhausting process of uploading, labeling, and training images without satisfactory results. They feel the site's complexity is intentional to exhaust user credits, preventing finding optimal settings. They switch attention to the last, Dojo Pixel, which requires a subscription and seems user-friendly. They experience initial complications with the image results but manage better outcomes after adjusting prompts. Overall, despite some success, all sites demonstrate setbacks like over-contrasting images or added noise.

マインドマップ

ビデオQ&A

  • What phone is used to take the pictures of the ring?

    The Samsung Note 1, a broken and dusty phone, is used to take the pictures.

  • What is the creator using AI for in this video?

    The creator is using AI to improve the quality of pictures taken of a ring, aiming to create commercial-style images.

  • What are the three websites the creator compares for AI image enhancement?

    The creator compares fall. a, tensor art, and Dojo pixel.

  • Which website did the creator find most effective for enhancing images?

    The creator found fall. a to be the most effective due to its simplicity, speed, and more natural-looking results.

  • What is the main drawback of using these AI platforms mentioned in the video?

    The platforms are quite expensive and have complex interfaces, making them challenging for regular users to navigate.

  • What were some issues with the AI image generation?

    Issues included images sometimes looking unnatural, overtraining affecting image quality, and problems with generating human fingers.

  • Why does the creator think AI for image enhancement is complicated?

    The process involves complex settings and jargon which are hard to understand for average users.

  • What problem did the creator face with tensor art?

    The images generated by tensor art did not resemble the product well and seemed to require subscription for better results.

  • How does the creator feel about the use of AI for product images?

    AI can make creating certain images easier, though they suggest the need for better user interfaces to simplify the process.

  • What does the creator suggest about pricing and usability of AI tools?

    Lowering fees and simplifying the interface could make these tools more accessible to sellers needing multiple attempts for a good result.

ビデオをもっと見る

AIを活用したYouTubeの無料動画要約に即アクセス！
字幕
en
オートスクロール:
  • 00:00:00
    hey guys today I have this beautiful
  • 00:00:03
    ring and I took several pictures of it
  • 00:00:06
    with my phone which is this one right
  • 00:00:09
    here this is an old phone I believe it's
  • 00:00:12
    the Samsung Note 1 it's a broken phone
  • 00:00:16
    and a Dusty phone but anyway guys in
  • 00:00:19
    this video I'm going to try using AI to
  • 00:00:23
    make my pictures better and hopefully
  • 00:00:26
    create commercial style image first of
  • 00:00:30
    all everyone I am not an expert with AI
  • 00:00:34
    so please don't judge me I honestly
  • 00:00:38
    think it's really for nerds uh no
  • 00:00:41
    offense to all the Nerds out there I'm
  • 00:00:44
    probably a nerd too but what I mean it's
  • 00:00:48
    so complicated for normal people They
  • 00:00:51
    have different names like flux sdxl or
  • 00:00:56
    something like that I hope uh in the
  • 00:00:59
    future some one make it easier for
  • 00:01:01
    regular people especially those who sell
  • 00:01:04
    products online I'm going to compare
  • 00:01:07
    three different websites the first is
  • 00:01:10
    fall. a then tensor art and Dojo pixel
  • 00:01:17
    they let us use Laura to train our
  • 00:01:20
    products image they are quite expensive
  • 00:01:24
    and they try to charge us as much money
  • 00:01:28
    as possible let's begin with fall. a
  • 00:01:33
    it's a thirdparty website uh that allows
  • 00:01:37
    us to use Laura by renting their
  • 00:01:40
    Hardware but before I do anything I will
  • 00:01:43
    rename each pictures I took and add a
  • 00:01:47
    trigger word which will be ring then I
  • 00:01:51
    will put them in a folder and create a
  • 00:01:54
    zip
  • 00:01:55
    file on the website I will search for
  • 00:01:58
    Laura and uh training with Laura I will
  • 00:02:03
    type in my trigger word and upload my
  • 00:02:06
    zip folder then just hit the start
  • 00:02:10
    button it takes a few minutes to train
  • 00:02:13
    the model and uh this is quite a big
  • 00:02:17
    drawback to be honest but I believe it
  • 00:02:20
    will get faster in the future you know
  • 00:02:23
    as Hardwares improves once the file is
  • 00:02:26
    ready I will hit the Run button and this
  • 00:02:30
    is where our experiments begin I need to
  • 00:02:34
    find some prompts so I did some research
  • 00:02:36
    on Google and find some prompts that I
  • 00:02:40
    copy and paste adding my trigger word
  • 00:02:43
    I've have play around with the setting
  • 00:02:45
    below uh I don't know what all the
  • 00:02:49
    setting means but one thing I know is
  • 00:02:52
    that the number of steps affects the
  • 00:02:55
    quality of the image uh generated I
  • 00:02:59
    think think I probably overtrain it and
  • 00:03:03
    uh I should not have set it to the
  • 00:03:06
    maximum because we need to leave some
  • 00:03:08
    room for the trainers to work with
  • 00:03:12
    anyway this is the result and I am quite
  • 00:03:17
    impressed to be honest you know however
  • 00:03:21
    it took me several tries to get these
  • 00:03:24
    results it took me a moment to find the
  • 00:03:28
    correct prompt and and mess around with
  • 00:03:31
    the settings this is where it gets quite
  • 00:03:35
    expensive anyway let's review all the
  • 00:03:38
    pictures generated from fall. a this is
  • 00:03:43
    the first image with the ring and some
  • 00:03:46
    red dust
  • 00:03:48
    explosion and I really like how it
  • 00:03:50
    actually blend the image uh with the
  • 00:03:54
    scene creating a natural Overlook and uh
  • 00:03:58
    you will see later that different sites
  • 00:04:00
    create different results I've tried
  • 00:04:04
    another prompt of a woman wearing the
  • 00:04:07
    actual ring and those are the results
  • 00:04:11
    and also a technician holding the ring
  • 00:04:15
    uh and the result was quite impressive
  • 00:04:19
    however sometimes the finger of are kind
  • 00:04:23
    of mess up and I read somewhere on the
  • 00:04:26
    web that the reason is because Laura
  • 00:04:29
    runs on the flux 1.0 which struggle
  • 00:04:35
    generating fingers there is a newer
  • 00:04:38
    version which is the flux
  • 00:04:40
    1.1 uh it does a better job however
  • 00:04:44
    there is no training modes yet there is
  • 00:04:46
    one thing about technology it improves
  • 00:04:50
    quickly there is something new every
  • 00:04:52
    month and it's kind of uh hard to keep
  • 00:04:55
    up the next website is tensor art
  • 00:05:00
    which was recommend by a YouTuber but it
  • 00:05:04
    is the worst of the three this time
  • 00:05:08
    instead of training with Laura I tried a
  • 00:05:10
    new method called lock don't ask me
  • 00:05:15
    what's the
  • 00:05:16
    difference uh because I have no idea and
  • 00:05:21
    uh I I don't really want to know you
  • 00:05:23
    know all I care about is the result this
  • 00:05:27
    site is the most complicated
  • 00:05:30
    uh compared to the first one and I
  • 00:05:33
    believe it's made by Chinese people
  • 00:05:37
    because I can pay with Ali pay anyway we
  • 00:05:41
    do the same thing we upload our image
  • 00:05:45
    and type our trigger
  • 00:05:48
    words but uh this time we also need to
  • 00:05:51
    type a preview prompt which I believe is
  • 00:05:54
    a description of the pictures once the
  • 00:05:58
    pictures are upload we need to label our
  • 00:06:01
    image it will automatically add some
  • 00:06:06
    tags but we have to remove any tags that
  • 00:06:10
    are not
  • 00:06:11
    relevant then uh we just start training
  • 00:06:15
    and it takes about an hour to finish
  • 00:06:19
    once it is done it will show several
  • 00:06:21
    results and ahead to choose uh the one
  • 00:06:25
    that looks most like the actual products
  • 00:06:29
    but of course of course none of them
  • 00:06:30
    turned out like G product which felt
  • 00:06:34
    like a total waste of time but I proceed
  • 00:06:38
    to create a new project and finally
  • 00:06:41
    started generating some
  • 00:06:43
    image on the side there are a lot of
  • 00:06:46
    complicated settings and they are limit
  • 00:06:50
    what I mean for example is if you try to
  • 00:06:53
    add more steps you have to upgrade to a
  • 00:06:57
    pro subscriptions this seems to want to
  • 00:07:01
    take as much money as possible from
  • 00:07:03
    users I already pay for some credits and
  • 00:07:07
    I really don't want to pay any extra
  • 00:07:10
    dollars come on guys just like before I
  • 00:07:13
    type in some prompts and the result were
  • 00:07:17
    awful maybe it was because I didn't have
  • 00:07:20
    the right settings applied I feel like
  • 00:07:24
    uh they made it complicated on purpose
  • 00:07:27
    so you would use all your credit
  • 00:07:30
    to get good results it's nearly
  • 00:07:33
    impossible to find the best settings
  • 00:07:36
    because there are so many
  • 00:07:38
    combination but I might be wrong right
  • 00:07:42
    maybe this site is better for character
  • 00:07:45
    generation and not for
  • 00:07:48
    objects and these are the result I got
  • 00:07:52
    and they all fail so uh don't waste your
  • 00:07:56
    money on this site the final site
  • 00:08:00
    is a pixel Dojo or Dojo pixel which
  • 00:08:04
    charge a monthly subscriptions fee I
  • 00:08:08
    think the overall usage and look are a
  • 00:08:12
    little bit better than uh tensor art uh
  • 00:08:16
    so again I upload my zip file and wait
  • 00:08:21
    an hour for the training to
  • 00:08:23
    finish uh once uh it was done I could go
  • 00:08:27
    to the create section and start
  • 00:08:30
    generating some image in the first few
  • 00:08:34
    pictures the ring look quite off because
  • 00:08:39
    uh it added some extra parts to the
  • 00:08:42
    Rings however the background was really
  • 00:08:46
    awesome after a few try I change my
  • 00:08:49
    prompt and add the dimension of the
  • 00:08:53
    gemstones and it suddenly create better
  • 00:08:56
    pictures the result were good but they
  • 00:09:01
    still had an AI look it looked over
  • 00:09:05
    contrast and kind of have a heavy uh Den
  • 00:09:09
    noise
  • 00:09:11
    effects I tried to fix that with uh
  • 00:09:14
    topas upscale function uh and it did a
  • 00:09:19
    little bit better making it feel more
  • 00:09:21
    natural all right everyone it's the end
  • 00:09:24
    of our video I personally like the first
  • 00:09:28
    website because it has a simple
  • 00:09:31
    interface and was also the fastest to
  • 00:09:34
    train a model the results look more
  • 00:09:39
    natural if someone uh could make a
  • 00:09:42
    website easier to understand and not use
  • 00:09:45
    hard words like steps or
  • 00:09:49
    vae it will really help a lot of
  • 00:09:53
    sellers it will also be great if they
  • 00:09:56
    could lower the high fees I know that I
  • 00:10:00
    know people need to earn money and I
  • 00:10:04
    agree but when using AI tools I often
  • 00:10:09
    need more than 50 tries to get one good
  • 00:10:12
    pictures people still need to take their
  • 00:10:16
    own pictures of their products some
  • 00:10:18
    might say that they can make those
  • 00:10:21
    pictures without AI but using AI make it
  • 00:10:25
    easier for example if you want to shoot
  • 00:10:29
    this image with the red dust first you
  • 00:10:33
    have to find the dust then set up your
  • 00:10:36
    camera and lights and take many photos
  • 00:10:40
    to get one good shot with AI you just
  • 00:10:44
    need to write the right prompt to get
  • 00:10:47
    the image you want you know anyway guys
  • 00:10:51
    please uh like this video and let us
  • 00:10:54
    know what you think about AI
タグ
  • AI
  • image enhancement
  • product photography
  • fall. a
  • tensor art
  • Dojo pixel
  • photo editing
  • AI tools
  • image quality
  • technology