<?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: c58 akf</title>
    <description>The latest articles on DEV Community by c58 akf (@c58_akf_b7c9df8ae6d4c5bc9).</description>
    <link>https://dev.to/c58_akf_b7c9df8ae6d4c5bc9</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%2F3476619%2Fc06a29c9-5c2a-4963-a892-3740cf3ebdf7.png</url>
      <title>DEV Community: c58 akf</title>
      <link>https://dev.to/c58_akf_b7c9df8ae6d4c5bc9</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/c58_akf_b7c9df8ae6d4c5bc9"/>
    <language>en</language>
    <item>
      <title>I tried to remake a 20-second TikTok ad shot by shot with AI. It cost $1.55 and four things broke.</title>
      <dc:creator>c58 akf</dc:creator>
      <pubDate>Wed, 11 Nov 2026 12:57:21 +0000</pubDate>
      <link>https://dev.to/c58_akf_b7c9df8ae6d4c5bc9/i-tried-to-remake-a-20-second-tiktok-ad-shot-by-shot-with-ai-it-cost-155-and-four-things-broke-l15</link>
      <guid>https://dev.to/c58_akf_b7c9df8ae6d4c5bc9/i-tried-to-remake-a-20-second-tiktok-ad-shot-by-shot-with-ai-it-cost-155-and-four-things-broke-l15</guid>
      <description>&lt;p&gt;I'm building &lt;a href="https://rea-videos.com" rel="noopener noreferrer"&gt;REA Videos&lt;/a&gt;, a tool for e-commerce sellers: upload a short viral video, say what to change, add a photo of your product, and get the same video back with your product in it. Same shots, same camera moves, same pacing.&lt;/p&gt;

&lt;p&gt;Before writing any product code, I wanted to know if the pipeline actually works end to end, and what it really costs. So I ran one real test. These are my notes, including the parts that failed.&lt;/p&gt;

&lt;h2&gt;
  
  
  The test
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Input:&lt;/strong&gt; a 20.7-second phone-shot TikTok ad for a portable blender, 480×854.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Change:&lt;/strong&gt; purple cup to a matte dark green one with a cream lid, marble kitchen counter to a wooden table at a lakeside camp, almond milk to oat milk, blueberries to strawberries, and a new caption.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Goal:&lt;/strong&gt; a remake with the same length and the same shot timeline as the original.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  The pipeline
&lt;/h2&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Breakdown.&lt;/strong&gt; I used the open-source &lt;a href="https://github.com/JuneYaooo/rea-video" rel="noopener noreferrer"&gt;rea-video&lt;/a&gt; engine (Apache-2.0) to split the clip into shots. REA Videos is a separate project that builds on it; I'm not affiliated with the rea-video authors. It found 14 shots and gave me start and end times, first and last frames, shot size, subject position, camera move, a caption timeline, and separate voice and music tracks. It ran locally in 61 seconds.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Product reference.&lt;/strong&gt; One image of the new cup, generated once and reused for every shot.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Keyframes.&lt;/strong&gt; For each shot, the original first frame was edited into the new scene with an image model (gpt-image-2 via kie.ai). 15 images, 13 of them in parallel, about 4 minutes in total.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Video clips.&lt;/strong&gt; Each keyframe was animated with Seedance 2.0 mini (also via kie.ai), first-frame mode, 480p, no audio. 14 of 14 clips succeeded.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Assembly.&lt;/strong&gt; ffmpeg trimmed each clip to the original shot length, added the caption layer, and produced two versions: silent, and with the original audio.&lt;/li&gt;
&lt;/ol&gt;

&lt;h2&gt;
  
  
  The numbers
&lt;/h2&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Step&lt;/th&gt;
&lt;th&gt;Time&lt;/th&gt;
&lt;th&gt;Cost&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Breakdown&lt;/td&gt;
&lt;td&gt;61 s (local)&lt;/td&gt;
&lt;td&gt;$0&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Product reference image&lt;/td&gt;
&lt;td&gt;101 s&lt;/td&gt;
&lt;td&gt;$0.03&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;15 keyframes&lt;/td&gt;
&lt;td&gt;~4 min (parallel)&lt;/td&gt;
&lt;td&gt;$0.45&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;14 video clips&lt;/td&gt;
&lt;td&gt;slowest one 160 s&lt;/td&gt;
&lt;td&gt;$1.07&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Assembly&lt;/td&gt;
&lt;td&gt;a few seconds&lt;/td&gt;
&lt;td&gt;$0&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;strong&gt;Total&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;&lt;strong&gt;10.6 min&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;&lt;strong&gt;$1.55&lt;/strong&gt;&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;The final video was 20.70 seconds, the same length as the original. I spot-checked 10 frames against what I asked for; 8 were good.&lt;/p&gt;

&lt;p&gt;One number surprised me: &lt;strong&gt;58 billed seconds for a 20.7-second video.&lt;/strong&gt; The video model generates at least 4 seconds per clip, and this ad has 14 short shots. You pay for 4 seconds and then trim most of it away. Billed seconds came out about 2.8 times the video length. Merging neighbouring short shots into one generation is the obvious fix, and I haven't tested it yet.&lt;/p&gt;

&lt;h2&gt;
  
  
  What broke
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;1. One reference image hijacked five shots.&lt;/strong&gt; I passed the finished keyframe of shot 1 as an extra reference for the other shots, hoping it would keep the scene consistent. Instead, five keyframes copied its composition. Shot 12, "pour the smoothie into a glass jar", turned into "pour oat milk into the cup". The fix: pass only the original shot's first frame plus the product photo, and describe the new scene in text.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;2. "Filmed or AI?" got it wrong.&lt;/strong&gt; Without a vision model, the breakdown called the phone footage 100% AI, at 0.598 confidence. My guess is that platform re-encoding wipes out the sensor noise the check relies on. For the real product, a multimodal model has to look at each shot.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;3. Caption OCR picked up everything except the captions.&lt;/strong&gt; It read the numbers on a clock and the word "Almond" on a carton as captions, and filed the actual caption under "label". So now the user confirms the captions before generating, and the caption layer is rendered by us, not copied from OCR.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;4. Two submits, two bills.&lt;/strong&gt; The cache was only written after a job finished, so submitting the same job twice paid twice. The fix is boring and necessary: write a "submitted" state and the task ID before calling any paid model.&lt;/p&gt;

&lt;p&gt;Smaller issues: describing the product in words only made its shape drift (always send the real photo), one shot still had blueberries, and a still keyframe lost a moving detail (a countdown on the blender's display).&lt;/p&gt;

&lt;h2&gt;
  
  
  Caveats
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;This was one clip, rendered at 480p. I haven't tested 720p yet, and its price per second is roughly double.&lt;/li&gt;
&lt;li&gt;I wrote the per-shot action descriptions by hand. In the real product a vision model writes them.&lt;/li&gt;
&lt;li&gt;"8 out of 10 frames" is my own spot check on one video, not a benchmark.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  Architecture that came out of it
&lt;/h2&gt;

&lt;p&gt;Cloudflare Workers can't run this: the breakdown needs Python, OpenCV and onnxruntime, assembly needs ffmpeg, and a job runs for minutes. So the plan is to split it:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Workers + D1&lt;/strong&gt; handle users, credits and job state.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;A small VPS&lt;/strong&gt; polls for jobs, runs the breakdown, calls the image and video models, runs ffmpeg, and reports back. It never touches money; credits are charged and refunded only on the Workers side.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;R2&lt;/strong&gt; holds every file. The models read signed URLs, and files are deleted after 30 days.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  Where REA Videos is now
&lt;/h2&gt;

&lt;p&gt;The site is in pre-launch. The pages and pricing are public, but the tool itself isn't open yet: the button adds you to a waitlist. No users, no testimonials, nothing to brag about yet. If you sell physical products and keep saving competitors' ads, I'd like to hear what you'd want from a remake.&lt;/p&gt;

&lt;p&gt;If you've built video pipelines with first-frame models, I'm curious how you handle short shots. Merge them, or eat the minimum clip length?&lt;/p&gt;

</description>
      <category>ai</category>
      <category>showdev</category>
      <category>video</category>
      <category>python</category>
    </item>
    <item>
      <title>A simple web simulator about chance, mortality, and waiting</title>
      <dc:creator>c58 akf</dc:creator>
      <pubDate>Mon, 20 Oct 2025 02:36:53 +0000</pubDate>
      <link>https://dev.to/c58_akf_b7c9df8ae6d4c5bc9/a-simple-web-simulator-about-chance-mortality-and-waiting-1dk</link>
      <guid>https://dev.to/c58_akf_b7c9df8ae6d4c5bc9/a-simple-web-simulator-about-chance-mortality-and-waiting-1dk</guid>
      <description>&lt;p&gt;Last week I built a small browser game — or maybe a thought experiment.&lt;br&gt;
It started as a curiosity about how probability feels when you live it instead of just calculating it.&lt;/p&gt;

&lt;p&gt;You start a simulated life. Every second, a random event might end it.&lt;br&gt;
There’s no strategy, no goal, no progress bar. Just a silent question:&lt;/p&gt;

&lt;p&gt;how long would you last if life were reduced to chance?&lt;/p&gt;

&lt;p&gt;It was loosely inspired by that classic Quora scenario:&lt;/p&gt;

&lt;p&gt;“You press a button and receive $100 million, but a random person dies instantly.”&lt;br&gt;
Except this time, you’re the random person.&lt;/p&gt;

&lt;p&gt;After a few runs, I realized I wasn’t really playing anymore — I was just watching, waiting.&lt;br&gt;
And that waiting became the point.&lt;/p&gt;

&lt;p&gt;Sartre once said we are “condemned to be free.”&lt;br&gt;
This simulation doesn’t offer choices, only outcomes — but somehow it still feels like freedom.&lt;/p&gt;

&lt;p&gt;You can try it here:&lt;br&gt;
👉 &lt;a href="https://probability.games/games/death-simulator" rel="noopener noreferrer"&gt;https://probability.games/games/death-simulator&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;It’s not meant to be morbid, just an experiment in what randomness feels like when it’s personal.&lt;/p&gt;

</description>
      <category>sideprojects</category>
      <category>showdev</category>
      <category>discuss</category>
      <category>webdev</category>
    </item>
    <item>
      <title>BGMI Name Generator</title>
      <dc:creator>c58 akf</dc:creator>
      <pubDate>Sat, 11 Oct 2025 12:10:52 +0000</pubDate>
      <link>https://dev.to/c58_akf_b7c9df8ae6d4c5bc9/bgmi-name-generator-2lgo</link>
      <guid>https://dev.to/c58_akf_b7c9df8ae6d4c5bc9/bgmi-name-generator-2lgo</guid>
      <description>&lt;p&gt;In the fast-paced, high-stakes world of Battlegrounds Mobile India (BGMI), your in-game name is more than just a label; it's the first glimpse of your persona, a declaration of your intent, and the name that will echo in the memories of your defeated opponents. A powerful name can be intimidating, memorable, and a crucial part of your gaming identity. This is where BGMI-Name.Online steps in, your ultimate ally in crafting the perfect moniker to represent you on the battlefield.&lt;/p&gt;

&lt;p&gt;BGMI-Name.Online is a comprehensive, user-friendly platform dedicated to helping every BGMI player discover and create a unique and stylish in-game name. We understand that a name is a reflection of a player's style – whether you're a lone wolf, a strategic mastermind, or a fearless aggressor, your name should resonate with your approach to the game. Our extensive and ever-growing database of names is meticulously categorized to cater to every preference. From cool and edgy names that strike fear into the hearts of your rivals, to clever and humorous options that showcase your personality, we have it all.&lt;/p&gt;

&lt;p&gt;Our powerful name generation tool is at the heart of the BGMI-Name.Online experience. This intuitive feature allows you to input keywords, select specific themes, and even incorporate special characters and symbols to generate a truly one-of-a-kind identity. Are you looking for a name that includes your favorite number, a mythical creature, or a word that embodies your clan's spirit? Our generator can seamlessly weave these elements into a list of creative and impactful suggestions. This name generator might help you spice up your game with unique nicknames—check it out! &lt;a href="https://bgmi-name.online/" rel="noopener noreferrer"&gt;https://bgmi-name.online/&lt;/a&gt;&lt;/p&gt;

</description>
    </item>
    <item>
      <title>AI Saree — Product Details</title>
      <dc:creator>c58 akf</dc:creator>
      <pubDate>Mon, 15 Sep 2025 11:18:18 +0000</pubDate>
      <link>https://dev.to/c58_akf_b7c9df8ae6d4c5bc9/ai-saree-product-details-1lll</link>
      <guid>https://dev.to/c58_akf_b7c9df8ae6d4c5bc9/ai-saree-product-details-1lll</guid>
      <description>&lt;p&gt;AI Saree is a web-based portrait generator that turns a single photo into cinematic, retro saree portraits in seconds—no prompts required. Built around a reference-guided pipeline, it preserves identity while applying era-correct lighting, color, and texture inspired by classic studio photography from the 1970s–1990s. New users receive 200 free credits (enough for two high-quality generations) with no card required.&lt;/p&gt;

&lt;p&gt;How It Works (3 Steps)&lt;br&gt;
    1.  Upload a Photo: Use a clear, front-facing portrait with even lighting and an unobstructed face for best fidelity.&lt;br&gt;
    2.  Pick a Retro Style: Choose decade, saree color, and backdrop, or start with curated presets matching the current Instagram “vintage portrait” trend.&lt;br&gt;
    3.  Download in HD: Generate in seconds and export watermark-free, high-resolution files suitable for social and print.&lt;/p&gt;

&lt;p&gt;Core Technology&lt;br&gt;
AI Saree blends portrait reference analysis (facial structure, lighting, and skin tone) with a vintage studio pipeline that adds soft light, backdrop toning, film grain, and era-specific color palettes. The system prioritizes likeness retention and composition consistency, so multiple outputs of the same person remain coherent and print-ready. Compared with prompt engineering in general-purpose models, AI Saree reduces trial-and-error and delivers reliable, repeatable aesthetics.&lt;/p&gt;

&lt;p&gt;Key Features&lt;br&gt;
    • Authentic Vintage Look: Era-correct film character, soft lenses, and classic studio backdrops.&lt;br&gt;
    • Saree Styles &amp;amp; Drapes: Select colors and drape variations to match mood and decade.&lt;br&gt;
    • Simple &amp;amp; Fast: Three-step, no-prompt flow; results in seconds.&lt;br&gt;
    • HD &amp;amp; Print-Ready: Balanced contrast and color for high-quality sharing and printing.&lt;br&gt;
    • Privacy &amp;amp; Control: Your photos are used to generate your art; download, archive, or delete anytime.&lt;/p&gt;

&lt;p&gt;Best-Practice Input Tips&lt;br&gt;
Shoot near a window or in diffused daylight. Keep a neutral background and a relaxed expression. Frame head-and-shoulders or half-body and avoid cropping the drape edges too tightly. If deciding between looks, try two contrasting palettes (e.g., warm 70s vs. cooler 80s) and pick the strongest result.&lt;/p&gt;

&lt;p&gt;Export &amp;amp; Presentation&lt;br&gt;
    • Social: Square for profiles, 4:5 for feeds, 9:16 for stories.&lt;br&gt;
    • Print: Matte or satin papers preserve detail without glare; canvas adds texture that complements vintage tones.&lt;br&gt;
    • Framing: Natural wood fits 70s warmth; slim black suits 80s urban palettes; gold accents echo 90s studio glam.&lt;/p&gt;

&lt;p&gt;Plans &amp;amp; Pricing&lt;br&gt;
    • Free: $0. Includes 200 credits, access to all retro styles, no watermark.&lt;br&gt;
    • Pro: $3.9/month. Includes 5,000 credits/month (~50 generations), premium styles and backdrops, priority processing, optional top-ups, and cancel anytime.&lt;/p&gt;

&lt;p&gt;Who It’s For&lt;br&gt;
Creators, photographers, and anyone who wants polished, retro saree portraits for social posts, profile images, gifts, or wall art—without prompt tinkering or complex photo editing. Start free, explore multiple eras, and share a timeless portrait within minutes at &lt;a href="https://aisaree.org" rel="noopener noreferrer"&gt;https://aisaree.org&lt;/a&gt;  .&lt;/p&gt;

</description>
    </item>
  </channel>
</rss>
