<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0"><channel><title><![CDATA[Andy Tan]]></title><description><![CDATA[Andy Tan]]></description><link>https://andytanlinks.hashnode.dev</link><generator>RSS for Node</generator><lastBuildDate>Wed, 18 Nov 2026 15:04:15 GMT</lastBuildDate><atom:link href="https://andytanlinks.hashnode.dev/rss.xml" rel="self" type="application/rss+xml"/><language><![CDATA[en]]></language><ttl>60</ttl><item><title><![CDATA[Swap the Background, Keep the Product: A Browser Workflow for Consistent Listing Photos]]></title><description><![CDATA[A product listing usually fails for a boring reason. The shoe, bottle, or jacket in shot A is not quite the same object as the one in shot B. The camera…]]></description><link>https://andytanlinks.hashnode.dev/swap-the-background-keep-the-product-a-browser-workflow-for-consistent-listing-photos</link><guid isPermaLink="true">https://andytanlinks.hashnode.dev/swap-the-background-keep-the-product-a-browser-workflow-for-consistent-listing-photos</guid><category><![CDATA[AI]]></category><category><![CDATA[image generation]]></category><category><![CDATA[Photography]]></category><dc:creator><![CDATA[Andy Tan]]></dc:creator><pubDate>Wed, 18 Nov 2026 07:21:17 GMT</pubDate><content:encoded><![CDATA[<p>A product listing usually fails for a boring reason. The shoe, bottle, or jacket in shot A is not quite the same object as the one in shot B. The camera height has shifted. A logo has softened. The white backdrop in the pack shot does not match the lifestyle frame you hoped to add later.</p>
<p>That is rarely a "I need a bigger model" problem. It is an image-to-image problem: you already have a usable photograph of the SKU, and you want a new setting without rebuilding the object from a text prompt.</p>
<p>This is a practical workflow for that job in the browser, before you stand up a local ComfyUI graph.</p>
<h2>Start from the SKU, not from a scene description</h2>
<p>Text-to-image is the wrong first move when the asset is a specific product. The model does not have your last, your cap thread, or the exact placement of a wordmark. If you describe "a matte black bottle of serum on marble, luxury lighting," you will get a handsome bottle. You will not reliably get <em>this</em> bottle.</p>
<p>Look at the source photo and split the job in two:</p>
<ul>
<li><strong>Keep:</strong> silhouette, camera angle, colorway, labels, hardware, the side of the object that faces the lens.</li>
<li><strong>Change:</strong> backdrop, surface, time of day, supporting props, perhaps a cleaner floor or a softer shadow.</li>
</ul>
<p>If the keep list is why you took the photo at all, start from the image. Describing the same object from scratch is how a product becomes "inspired by" rather than itself.</p>
<h2>Prepare the reference as if it were the brief</h2>
<p>The reference is the contract. A slightly boring, sharp pack shot will usually beat a moodier lifestyle frame with clutter in the corners.</p>
<p>Before you generate anything:</p>
<ol>
<li>Crop so the product is obvious. Extra table edge, extra carton, extra hand in the corner are all invitations to drift.</li>
<li>Check labels and edges at 100% zoom. If type is already soft, the generator will not invent a legally accurate label for you.</li>
<li>Note the camera height and the standing plane. A three-quarter view and a straight-on catalog view are different jobs.</li>
<li>Write down the one change you actually want. "New background, same shoe" is a job. "New background, new colorway, new props, and a different focal length" is four jobs stacked on top of each other.</li>
</ol>
<p>Those two lists become the prompt. Everything else is noise.</p>
<h2>Write a keep clause, then a change clause</h2>
<p>A useful prompt for listing work names what is not up for negotiation, then names the setting.</p>
<p>Weak: "Cinematic product photo of a white sneaker on a luxury oak bench at sunrise, 85mm, incredible detail, award-winning."</p>
<p>Stronger: "Keep the same white leather sneaker and the same three-quarter camera angle. Place it on a sunlit oak bench, soft morning side light, no extra props, no change to the shoe itself."</p>
<p>The second version tells the model what must survive. If you also ask it to recolor the leather and invent typography, the object will drift. That is usually an instruction problem, not a sampling problem.</p>
<p>This is also where a browser <a href="https://imagetoimage.im/?utm_source=hashnode&amp;utm_medium=ugc_article&amp;utm_campaign=imagetoimage">image-to-image AI</a> pass is often enough. Upload the pack shot, keep the product constraints in the prompt, and generate a small set of backdrop variants. You do not need a local GPU for a first pass of "white seamless versus oak bench versus linen studio." Tools in this class sit between a one-tap mobile filter and a full ComfyUI graph: useful when the photograph already holds the hard parts.</p>
<p>Generate two to four variants, not twenty. A short run is easier to compare, and it makes it obvious when the object itself has started to move.</p>
<h2>Judge the output like a catalog, not like a mood board</h2>
<p>Open the best two results next to the source and check the parts a shopper will notice:</p>
<ul>
<li><strong>Silhouette and stance.</strong> Did the toe spring, the bottle shoulder, or the jacket hem change?</li>
<li><strong>Color.</strong> Is the leather, plastic, or fabric still the same colorway, including in shadow?</li>
<li><strong>Labels and hardware.</strong> Are wordmarks still readable, or have they become decorative marks?</li>
<li><strong>Perspective.</strong> Did the camera climb or orbit, even slightly?</li>
<li><strong>Contact shadow.</strong> A floating object on a new floor is usually a failed composite, not a style choice.</li>
</ul>
<p>If the backdrop is right but the product has drifted, tighten the keep clause and run again from the original photo. Do not take the slightly wrong variant and ask the model to "fix the logo." That usually compounds the error.</p>
<p>If the product holds and only the floor or the light is off, a second, narrower pass can be worth it. Change one thing. Leave the rest of the prompt alone.</p>
<h2>Know when to stop and reshoot</h2>
<p>Browser image-to-image is a poor substitute for a missing photograph. Reshoot, or go back to the studio, when:</p>
<ul>
<li>The source is too soft, too tight, or badly lit to be a fair reference.</li>
<li>You need a legally exact label, a registered wordmark, or packaging copy that must match the physical SKU.</li>
<li>You need a different camera height or a true rear view you never photographed.</li>
<li>You are trying to invent a material that was never in the frame.</li>
</ul>
<p>A local stack still earns its keep for batch consistency, ControlNet-level lock, or a long series that has to match. For a handful of listing variants from a decent pack shot, it is often more than you need today.</p>
<h2>A short checklist</h2>
<ul>
<li>Start from the SKU photo when the object is the asset.</li>
<li>Crop, sharpen your read of labels, and write a keep list before you write a scene.</li>
<li>Put the keep clause first in the prompt; change one setting at a time.</li>
<li>Compare variants against the source at 100%, not against a mental mood board.</li>
<li>Reshoot or go local when the source cannot support the claim you want to make.</li>
</ul>
<p>The goal is not a prettier picture in isolation. It is a set of frames in which the same product is still the same product.</p>
]]></content:encoded></item><item><title><![CDATA[Draft Fast, Refine After: A Two-Stage Workflow When a Platform Softens a Legal Prompt]]></title><description><![CDATA[A familiar failure mode in image tools is not a bad sampler. You write a legal, specific brief — a particular costume, a dim street, a literal reading of a…]]></description><link>https://andytanlinks.hashnode.dev/draft-fast-refine-after-a-two-stage-workflow-when-a-platform-softens-a-legal-prompt</link><guid isPermaLink="true">https://andytanlinks.hashnode.dev/draft-fast-refine-after-a-two-stage-workflow-when-a-platform-softens-a-legal-prompt</guid><category><![CDATA[AI]]></category><category><![CDATA[image generation]]></category><category><![CDATA[prompting]]></category><category><![CDATA[workflow]]></category><category><![CDATA[Tutorial]]></category><dc:creator><![CDATA[Andy Tan]]></dc:creator><pubDate>Tue, 17 Nov 2026 07:38:19 GMT</pubDate><content:encoded><![CDATA[<p>A familiar failure mode in image tools is not a bad sampler. You write a legal, specific brief — a particular costume, a dim street, a literal reading of a book-cover scene — and the interface hands back a safer, brighter, vaguer cousin of the idea. The lighting is "nicer." The outfit is less exact. The street has become golden-hour stock photography.</p>
<p>That is a workflow problem. It is also not a license to hunt for a magic bypass. Model providers, host terms, and the law still apply, and they should. The useful goal is narrower: get a first draft that still resembles what you asked for, then refine it.</p>
<h2>Diagnose the softening before you add more tags</h2>
<p>Not every miss is a filter. Before you change tools, decide which of these you are looking at:</p>
<ul>
<li><strong>Platform rewrite.</strong> The text in the box is no longer the text you typed, or the result is consistently "helped" toward a generic beauty default.</li>
<li><strong>Model limit.</strong> The same prompt fails in a local UI or on another host, or the refusal is explicit.</li>
<li><strong>Overlong brief.</strong> A paragraph of style tags gave the system more surface area to "improve."</li>
</ul>
<p>If the work is adult, disallowed, or illegal, stop and change the brief. This article is about ordinary concept work: stylized portraits, wardrobe tests, tabletop tokens, period interiors, or a night scene that keeps being turned into daylight.</p>
<p>A short test that saves time: run the same short, literal prompt once on the current host and once in a setting with fewer platform-level rewrites. If both refuse, you are at a model or policy limit. If only the first host wanders, you are dealing with a rewrite layer.</p>
<h2>Stage A — draft until the idea is recognizable</h2>
<p>Treat the first generation as a sketch, not a delivery file.</p>
<ul>
<li>Start short and literal. Name the subject, the setting, and one constraint. Confirm that the costume, era, or layout survived before you add lenses and film stocks.</li>
<li>Write lighting as a constraint, not as a mood word. "Overcast, no beauty dish, visible sodium street lamps" holds better than "cinematic."</li>
<li>Save the first draft that is directionally right, even if the hands are bad or the background is messy. That file is now your reference. The original prompt is not.</li>
<li>If the output keeps sliding toward a generic beauty-lighting default, reduce adjectives and restate the one detail that keeps disappearing.</li>
</ul>
<p>For draft exploration with fewer platform-level filter rewrites, a browser <a href="https://norestrictionsai.com/?utm_source=hashnode&amp;utm_medium=ugc_article&amp;utm_campaign=norestrictionsai">uncensored AI image generator</a> can sit next to open-source UIs as a concept pass. In practice this means the default model is not putting an extra prompt-rewrite or keyword-screening step in front of you. It is not a promise that every model will honor every request, and it is not a way around provider policies or the law. Use it to see whether the brief holds; then move on.</p>
<p>A note on language: "uncensored" or "unrestricted" here means the <em>platform is not rewriting the brief by default</em>. The underlying model can still refuse, distort, or decline. If a result comes back empty or clearly policy-blocked, that is a stop sign, not a prompt-engineering challenge.</p>
<h2>Stage B — refine from the keeper, not from the original text</h2>
<p>Once you have a draft that matches intent, stop regenerating from the prompt alone. Use the draft as a reference and make smaller edits:</p>
<ul>
<li>Fix composition: crop, straighten the stance, remove one distracting object.</li>
<li>Lock style: push it toward illustration, product still, or photographic grade without changing the subject.</li>
<li>Repair local issues with a tighter inpaint or a second image-to-image pass at a lower change strength.</li>
<li>Only then upscale, if you actually need print resolution.</li>
</ul>
<p>This is where people often undo their own progress. They take a good draft, go back to a longer prompt, and ask the tool to "make it better." The subject drifts, the costume softens, and the night street is golden hour again. Refine the picture you already have.</p>
<h2>Guardrails that should stay in place</h2>
<p>A less-filtered draft stage is not an "anything goes" pipeline.</p>
<ul>
<li>Stay inside the law and the terms of the tools you use.</li>
<li>Do not try to recreate a living person's face for a deceptive use, and do not treat a generated logo or typography as official.</li>
<li>If a model or a host blocks a request, accept the block. Switching tools to evade a safety limit is a different project from recovering a softened legal brief.</li>
</ul>
<p>The point of Stage A is prompt fidelity for work you are allowed to do. The point of Stage B is craft. Keep those jobs separate and the work gets faster, not looser.</p>
<h2>A short loop you can reuse</h2>
<ol>
<li>Write a short literal prompt. Generate one or two drafts.</li>
<li>If the idea is recognizable, stop text-only generation.</li>
<li>Refine from the keeper: composition, then style, then local repairs.</li>
<li>Do a human check on faces, type, and anything a client would treat as a fact.</li>
</ol>
<p>If Stage A never produces a recognizable draft, the brief may be the problem, or you may be at a real policy limit. Either way, adding twenty style tags rarely fixes it.</p>
]]></content:encoded></item></channel></rss>