https://www.luyingzhi.cn/en/Recent content in Home on Dream BuilderHugoen© 2026 luyingzhiTue, 27 Oct 2026 07:00:00 +0000https://www.luyingzhi.cn/en/posts/voice-cloning-local-vs-cloud.htmlFri, 23 Oct 2026 02:00:00 +0000https://www.luyingzhi.cn/en/posts/voice-cloning-local-vs-cloud.html<h2 id="start-with-the-decision-table">Start with the decision table</h2> <table> <thead> <tr> <th>Your situation</th> <th>Cloud</th> <th>Local</th> </tr> </thead> <tbody> <tr> <td>Under 100k characters a month</td> <td>✅</td> <td></td> </tr> <tr> <td>Batch generation needed</td> <td></td> <td>✅</td> </tr> <tr> <td>Sensitive content (internal training, unreleased product)</td> <td></td> <td>✅</td> </tr> <tr> <td>Only need stock voices</td> <td>✅</td> <td></td> </tr> <tr> <td>Need to clone a specific voice</td> <td></td> <td>✅</td> </tr> <tr> <td>No GPU</td> <td>✅</td> <td></td> </tr> </tbody> </table> <p>Cloud TTS wins on <strong>convenience</strong>. Local cloning wins on <strong>cost, control and privacy</strong>.</p> <h2 id="cost">Cost</h2> <p>Mainstream cloud TTS runs about $0.015–$0.045 per 1,000 characters. At 500k characters a month that is $7.50–$22.50 monthly, $90–$270 a year.</p>https://www.luyingzhi.cn/en/posts/krea2-workflow-pipelines.htmlThu, 22 Oct 2026 06:00:00 +0000https://www.luyingzhi.cn/en/posts/krea2-workflow-pipelines.html<h2 id="why-three-pipelines-instead-of-one">Why three pipelines instead of one</h2> <p>Most AI image tools take the same approach: one big form, every parameter exposed. Past a certain count the interface becomes a control panel and users no longer know what to adjust.</p> <p>Krea2 splits into three independent pipelines, each exposing only what it genuinely needs:</p> <table> <thead> <tr> <th>Pipeline</th> <th>What it does</th> <th>Parameters you touch</th> </tr> </thead> <tbody> <tr> <td><strong>Photoreal generation</strong></td> <td>Prompt to high-realism photo</td> <td>5–6 (subject, scene, lighting, aspect, sampling strength)</td> </tr> <tr> <td><strong>Character casting</strong></td> <td>Consistent character across images</td> <td>1 reference + 4 (expression, pose, outfit, scene)</td> </tr> <tr> <td><strong>Three-stage upscale</strong></td> <td>Upscale to 2K/4K with detail recovery</td> <td>Target size + 3 refinement strengths</td> </tr> </tbody> </table> <p>The benefit of splitting: <strong>each pipeline&rsquo;s defaults are tuned for its own task.</strong> Photoreal generation and casting do not want the same CFG. Merged into one form, neither gets good defaults.</p>https://www.luyingzhi.cn/en/posts/local-ai-image-generation.htmlTue, 20 Oct 2026 02:00:00 +0000https://www.luyingzhi.cn/en/posts/local-ai-image-generation.html<h2 id="the-short-answer">The short answer</h2> <p><strong>Above roughly 300 images a month, running locally is cheaper and faster.</strong></p> <p>The price is setup friction: you need a decent GPU and you have to manage model files yourself.</p> <p>Here is what six months of my own usage looks like:</p> <table> <thead> <tr> <th>Dimension</th> <th>Cloud service</th> <th>Local desktop</th> </tr> </thead> <tbody> <tr> <td>Cost per image</td> <td>$0.007 – $0.04</td> <td>~$0.0003 (electricity)</td> </tr> <tr> <td>First image latency</td> <td>2–8 s (queue included)</td> <td>3–15 s (model load)</td> </tr> <tr> <td>Subsequent images</td> <td>2–8 s</td> <td><strong>1–3 s</strong></td> </tr> <tr> <td>Privacy</td> <td>Image uploaded to a third party</td> <td><strong>Fully local</strong></td> </tr> <tr> <td>Hardware</td> <td>None</td> <td>8 GB+ VRAM</td> </tr> <tr> <td>Max resolution</td> <td>Vendor cap</td> <td><strong>VRAM bound only</strong></td> </tr> </tbody> </table> <h2 id="why-web-services-feel-slow">Why web services feel slow</h2> <p>A web request breaks down into three parts:</p>