<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:dc="http://purl.org/dc/elements/1.1/">
  <channel>
    <title>DEV Community: AI Videographer Notes</title>
    <description>The latest articles on DEV Community by AI Videographer Notes (@aivideographernotes).</description>
    <link>https://dev.to/aivideographernotes</link>
    <image>
      <url>https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F4242828%2F72ba9b5e-8cb8-4058-b64c-cf2d1da7fb00.jpg</url>
      <title>DEV Community: AI Videographer Notes</title>
      <link>https://dev.to/aivideographernotes</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/aivideographernotes"/>
    <language>en</language>
    <item>
      <title>The year signs came out wrong until I split the still from the motion</title>
      <dc:creator>AI Videographer Notes</dc:creator>
      <pubDate>Fri, 13 Nov 2026 11:34:42 +0000</pubDate>
      <link>https://dev.to/aivideographernotes/the-year-signs-came-out-wrong-until-i-split-the-still-from-the-motion-3l19</link>
      <guid>https://dev.to/aivideographernotes/the-year-signs-came-out-wrong-until-i-split-the-still-from-the-motion-3l19</guid>
      <description>&lt;p&gt;I wanted a short New Year clip from a single photograph: two copies of the same person on a wet night road, a hug under signs that read 2026 and 2027, a suitcase, and a car pulling away. It is a simple scene to describe and a messy one to generate.&lt;/p&gt;

&lt;p&gt;The first attempts went into a video model in one shot. I asked it to invent the road, the rain, both figures, both signs, and the timing of the hug. The motion was often fine. The details were not. One take spelled a year as 2072. Another bent a digit until it was unreadable. A third gave me two people who were clearly not the same person, and once a stranger appeared in the background.&lt;/p&gt;

&lt;p&gt;That is the usual failure mode. The model has to invent a set, keep typography stable, copy a face twice, and time a ten-second departure all at once. When one of those jobs slips, you feel it immediately, because the years are the joke and the two faces are the point.&lt;/p&gt;

&lt;p&gt;What worked was splitting the job. I made the still first, in an image model, and I did not start the motion pass until I could read both signs and see that both figures were the same person. Checking the frame before a video render saved the takes that would have failed the moment a 2 became a 7.&lt;/p&gt;

&lt;p&gt;The scene prompt I settled on was narrow on purpose. It names exactly two copies, says which year sits on which side, and forbids extra text:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;Use the single adult from the input photo to create exactly two identical copies of the same adult, preserving face, hair and clothing. They gently hug on a quiet rain-slick night road. Overhead signs read exactly 2026 on the left and 2027 on the right. A dark sedan and a small suitcase wait on the right. Cinematic natural rain, vertical 9:16, no other people, no logos, no extra text.
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;A few rules mattered more than any adjective in that paragraph. Write each year as one number and say it is the only text in the image. Extra words on signs are where models invent letters. Say which copy stands on which side, so the person who stays is under 2026 and the person who leaves is under 2027. Ask for exactly two people. Models like to add a passer-by.&lt;/p&gt;

&lt;p&gt;If you would rather not run an image model and a video model yourself, a &lt;a href="https://aivideographer.net/goodbye-2026-welcome-2027-ai-video" rel="noopener noreferrer"&gt;Goodbye 2026 Welcome 2027 template&lt;/a&gt; builds that same rainy road scene around one photo and then animates it. The finished file still needs a look at the signs and the faces before you keep it. A prepared scene does not remove the need to look.&lt;/p&gt;

&lt;p&gt;The motion pass is easier once the still is honest. I gave the picture to an image-to-video model with a second-by-second action list, not a vague instruction to make it cinematic:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;From 0 to 2 seconds finish the hug. From 2 to 3 seconds the person on the right lets go and takes the suitcase. From 3 to 7 seconds that person walks to the car and gets inside. From 7 to 10 seconds the car drives away with red tail lights reflecting on the wet road while the person on the left stays and watches. Camera gently pulls back. Natural rain and engine sounds, no music, no dialogue.
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Keep the camera slow. Fast moves made the digits wobble and the faces drift, even when the still had been clean. Ten seconds is enough for a hug and a look back. It is not enough to hide a bad still.&lt;/p&gt;

&lt;p&gt;The source photo did as much work as the prompts. A front-facing, well-lit adult, with enough of the body for a medium full shot, gave the model one clear face to copy. Sunglasses, a cap over the eyes, a tight crop, or a strong side angle made the second copy drift. A busy background invited a third person. I also stayed inside ordinary upload limits: one JPEG, PNG, or WebP, and not a huge file.&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight python"&gt;&lt;code&gt;&lt;span class="k"&gt;from&lt;/span&gt; &lt;span class="n"&gt;pathlib&lt;/span&gt; &lt;span class="k"&gt;import&lt;/span&gt; &lt;span class="nc"&gt;Path&lt;/span&gt;

&lt;span class="k"&gt;def&lt;/span&gt; &lt;span class="nf"&gt;check_still&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;path&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;
    &lt;span class="n"&gt;p&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="nc"&gt;Path&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;path&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
    &lt;span class="n"&gt;size&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="n"&gt;p&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;stat&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;st_size&lt;/span&gt;
    &lt;span class="k"&gt;return&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
        &lt;span class="s2"&gt;"format_ok"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="n"&gt;p&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;suffix&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;lower&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="k"&gt;in&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="s2"&gt;".jpg"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="s2"&gt;".jpeg"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="s2"&gt;".png"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="s2"&gt;".webp"&lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
        &lt;span class="s2"&gt;"under_8mb"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="n"&gt;size&lt;/span&gt; &lt;span class="o"&gt;&amp;lt;=&lt;/span&gt; &lt;span class="mi"&gt;8&lt;/span&gt; &lt;span class="o"&gt;*&lt;/span&gt; &lt;span class="mi"&gt;1024&lt;/span&gt; &lt;span class="o"&gt;*&lt;/span&gt; &lt;span class="mi"&gt;1024&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
        &lt;span class="s2"&gt;"bytes"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="n"&gt;size&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
    &lt;span class="p"&gt;}&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;That script does not judge whether a face is sharp. It only catches the boring rejections. I still opened the photo at phone size and asked whether I could tell who the person was in the first second. If I could not, a rainy street was not going to help.&lt;/p&gt;

&lt;p&gt;When a take still went wrong, the fixes were small. If a sign read 2072, I ran it again, or covered the digit in an editor. If the two copies looked like different people, I went back to a plainer, front-facing photo. If a stranger appeared, I ran it again. I also made the file a few days early, so I was not waiting on a render at 11:55 p.m.&lt;/p&gt;

&lt;p&gt;A short checklist I now keep next to the prompts:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;One clear, front-facing still of a consenting adult, body and outfit visible, no heavy filters.&lt;/li&gt;
&lt;li&gt;Confirm the file is JPEG, PNG, or WebP and under 8 MB.&lt;/li&gt;
&lt;li&gt;Build the still first. Read both signs. Confirm there are exactly two copies of the same person.&lt;/li&gt;
&lt;li&gt;Animate with a slow camera and a timed action list, not a single mood word.&lt;/li&gt;
&lt;li&gt;Watch the finished clip for years, faces, hands, and extra people before you keep it.&lt;/li&gt;
&lt;li&gt;If you share it, label it as AI-generated when the platform offers that option.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;I still get a bad take. The difference is that I can see which stage failed. If the signs are wrong, I fix the still. If the faces smear, I slow the camera down. I no longer ask one video pass to invent the street, the years, and both people at once.&lt;/p&gt;

</description>
      <category>ai</category>
      <category>tutorial</category>
      <category>python</category>
      <category>productivity</category>
    </item>
  </channel>
</rss>
