![\"\"](\"https:\/\/crepal.ai\/blog\/wp-content\/uploads\/2026\/05\/image-62.png\")
<\/figure>\n<\/figure>\n\n\n\n\n\n\n\n
What You Need Before You Start<\/h2>\n\n\n\nSource Image Quality and Format<\/h3>\n\n\n\n
This is where most people leave credits on the table. A blurry, heavily-compressed, or oddly-cropped source image will give you a worse output \u2014 no prompt will fix a bad source.<\/p>\n\n\n\n
What works:<\/p>\n\n\n\n
- \n
- Resolution<\/strong>: 1024\u00d71024 minimum on the short edge. Anything under 512px tends to produce noticeable degradation.<\/li>\n\n\n\n
- Format<\/strong>: PNG or high-quality JPG. Compression artifacts in the source get amplified in video. Avoid screenshots of screenshots.<\/li>\n\n\n\n
- Framing<\/strong>: centered subject, clear negative space. If the crop cuts off limbs awkwardly, the model will try to complete them \u2014 and it usually gets it wrong.<\/li>\n\n\n\n
- Lighting consistency<\/strong>: flat or softly directional light generates more stable motion than high-contrast dramatic lighting. The model struggles to maintain shadows across frames.<\/li>\n\n\n\n
- Face clarity<\/strong>: if the face matters to you, it needs to be sharp and facing roughly forward. Profile angles produce drift in most current models.<\/li>\n<\/ul>\n\n\n\n
One thing I learned the hard way \u2014 AI-generated images often work better as source material than photos. They’re already in a style the model understands, and they don’t carry the uncanny-valley tension between photorealistic source and AI-animated output. Stability AI’s research page<\/a> is a decent place to track which underlying models are getting better at this kind of consistency, since most consumer image-to-video tools build on top of research-grade models.<\/p>\n\n\n\n
\n![\"\"](\"data:image\/svg+xml;base64,PHN2ZyB3aWR0aD0iMSIgaGVpZ2h0PSIxIiB4bWxucz0iaHR0cDovL3d3dy53My5vcmcvMjAwMC9zdmciPjwvc3ZnPg==\")
<\/figure>\n<\/figure>\n\n\n\nAccess, Credits, and Moderation Checks<\/h3>\n\n\n\n
Not all platforms that advertise NSFW capability are the same. A few things to verify before burning credits:<\/p>\n\n\n\n
Free and “no sign-up” tiers<\/strong>: Most platforms with genuine NSFW capability don’t offer it on free plans. If you’re looking for a no-sign-up option, you’re mostly looking at limited public APIs or platforms where the moderation toggle is off by default \u2014 and those usually have watermarks and 480p output caps. Usable for testing a workflow, not for final output.<\/p>\n\n\n\n
Age verification<\/strong>: Legitimate platforms gate this behind an age check or account verification step. If a tool has zero verification and full NSFW enabled out of the box, that’s a yellow flag for platform longevity \u2014 those tools tend to disappear without warning.<\/p>\n\n\n\n
Credit consumption<\/strong>: Image-to-video is expensive relative to image generation. Budget 3\u20138 credits per clip depending on resolution and length. Run a low-resolution test before committing full credits to a prompt you haven’t validated.<\/p>\n\n\n\n
\n\n\n\nStep-by-Step Workflow<\/h2>\n\n\n\n
Prepare the Image<\/h3>\n\n\n\n
- \n
- Upscale to at least 1024px if needed \u2014 a dedicated upscaler like Real-ESRGAN handles this better than in-browser tools.<\/li>\n\n\n\n
- Crop to a ratio your target platform accepts (usually 1:1 or 16:9 \u2014 check the platform docs).<\/li>\n\n\n\n
- Run a quick brightness\/contrast pass if the image is very dark. Dark sources produce muddy motion.<\/li>\n\n\n\n
- If the image is AI-generated and you still have the original prompt, keep it. You may want it for the motion prompt later.<\/li>\n<\/ol>\n\n\n\n
Write the Motion Prompt<\/h3>\n\n\n\n
This is the part nobody wants to spend time on, and it’s why most outputs are disappointing.<\/p>\n\n\n\n
The motion prompt describes what moves, not what exists. Your source image already handles the “what exists” part.<\/p>\n\n\n\n
Structure that actually works:<\/strong><\/p>\n\n\n\n
[Subject movement] + [camera or environmental motion] + [pacing\/speed modifier]<\/code><\/pre>\n\n\n\nExamples from my own testing:<\/p>\n\n\n\n
- \n
- Weak: “beautiful woman, sensual”<\/em> \u2192 model mostly just pulsed the image and added grain<\/li>\n\n\n\n
- Better: “slow exhale, subtle chest rise, hair drifts slightly left, soft breeze from right, cinematic hold”<\/em> \u2192 actual motion, face held<\/li>\n\n\n\n
- Specific action: “turns head 15 degrees to right, glances down, lips part slightly, candlelight flickers”<\/em> \u2192 worked, though the candlelight introduced flicker artifacts<\/li>\n<\/ul>\n\n\n\n
Keep prompts under 60 words. Long prompts don’t give the model more instruction \u2014 they give it more to ignore.<\/p>\n\n\n\n
Negative prompts<\/strong> matter more here than in image generation. Standard additions: blurry, warped face, extra limbs, morphing, bad hands, flickering, overexposed<\/p>\n\n\n\n
\n<\/figure>\n<\/figure>\n\n\n\nFix Common Artifacts<\/h3>\n\n\n\n
- Better: “slow exhale, subtle chest rise, hair drifts slightly left, soft breeze from right, cinematic hold”<\/em> \u2192 actual motion, face held<\/li>\n\n\n\n
- Weak: “beautiful woman, sensual”<\/em> \u2192 model mostly just pulsed the image and added grain<\/li>\n\n\n\n
- Format<\/strong>: PNG or high-quality JPG. Compression artifacts in the source get amplified in video. Avoid screenshots of screenshots.<\/li>\n\n\n\n
![\"\"](\"https:\/\/crepal.ai\/blog\/wp-content\/uploads\/2026\/05\/image-66-1024x435.png\")
<\/figure>\n<\/figure>\n\n\n\n![\"\"](\"https:\/\/crepal.ai\/blog\/wp-content\/uploads\/2026\/05\/image-65-1024x559.png\")
<\/figure>\n<\/figure>\n\n\n\n![\"\"](\"https:\/\/crepal.ai\/blog\/wp-content\/uploads\/2026\/05\/image-64-1024x610.png\")
<\/figure>\n<\/figure>\n\n\n\n