Vidéo IA à partir d’image : le flux de production complet

Il y a un écart de qualité massif entre quelqu'un qui télécharge une image IA sur Runway et clique sur "générer" et quelqu'un qui suit un workflow de production approprié. La différence se voit dans le produit final: one looks obviously AI-generated, the other could pass for real footage on most platforms.

Cet article décompose le workflow professionnel en 5 phases que j'utilise pour chaque vidéo que je produis. Chaque phase inclut des outils, paramètrès et réglages spécifiques. This isn't theory - it's the exact process behind the content I've been publishing for the past year.

Cette phase prend 15 à 20 minutes mais évite des heures de générations vidéo gaspillées. Sautez-la et vous gaspillerez des crédits à régénérer des clips that fail because the source image had issues.

Agrandissement

Chaque image source devrait être agrandie à au moins 2x its generation resolution before entering the video pipeline. If you generated at 768x1344, upscale to 1536x2688. The reason: video AI models extract detail from the input image to inform the generated frames. More source detail means more stable, higher-quality video output.

Best upscaling tools:

• Topaz Gigapixel AI - $99 one-time. The best quality for photorealistic faces. Use "Standard" mode with "Recover Original Detail" at 50%.
• Real-ESRGAN (via Automatic1111 or ComfyUI) - Free. Use the "4x-UltraSharp" model. Quality is about 85% of Topaz but costs nothing.
• Magnific AI - $39/month. Best for adding detail during upscale. Can actually improve face quality, not just enlarge it. Overkill for most use cases but worth it for hero images.

Correction du rapport d'aspect

Si votre image n'est pas déjà au rapport d'aspect cible, recadrez-la maintenant. Do not rely on the video tool to handle aspect ratio conversion - most either stretch or add ugly letterboxing.

Suppression des artefacts

Parcourez chaque image et corrigez :

• Hand anomalies - Extra fingers, fused fingers, unnatural poses. Use Photoshop's generative fill or SDXL inpainting with a hand-specific LoRA.
• Jewelry distortion - Earrings, necklaces, and rings frequently have AI artifacts. Inpaint or remove them entirely.
• Background inconsistencies - Warped architecture, floating objects, impossible reflections. These get amplified in video.
• Skin texture issues - Over-smoothed skin or plastic-looking areas. Topaz Photo AI's "Recover Face" can help, or use Photoshop's frequency separation technique.

Sélection d'outil par type de plan

Choisissez votre outil en fonction du plan spécifique, pas par fidélité à une seule plateforme :

• Close-up portraits (face fills 40%+ of frame): Runway Gen-3 Alpha. Use "Turbo" mode. Set motion intensity to 3/10.
• Medium shots (waist-up): Kling AI 1.6 or Runway. Kling handles arm gestures better; Runway handles face quality better.
• Full body shots: Kling AI 1.6. No contest here. Set motion mode to "Standard" and motion intensity to 5/10.
• Talking head: HeyGen. Upload image, input script, select voice. 5 minutes max per clip.
• Atmospheric/mood: Luma Dream Machine. The cinematic quality is unmatched for non-dialogue content.

Rédaction de prompts pour chaque outil

Runway Gen-3 Alpha prompts: Keep them short and motion-focused. Runway responds best to prompts under 30 words. Example: "Woman slowly turns head right, natural blink, wind moves hair, soft lighting, static camera, photorealistic." Runway ignores style keywords like "4K" or "cinematic" - it generates at its native quality regardless.

Kling AI 1.6 prompts: Kling handles longer, more descriptive prompts. Include camera movement explicitly. Example: "A woman walks slowly toward the camera on a city sidewalk, natural stride, arms relaxed at sides, slight smile. Camera: slow dolly backward at matching pace. Photorealistic, natural lighting, shallow depth of field." Kling's "Professional" mode adds about 30 seconds to generation time but noticeably improves quality.

Luma Dream Machine prompts: Luma thrives on atmosphere. Example: "Golden hour light wraps around a woman standing on a rooftop, wind moves her dress and hair, city skyline blurred in background, cinematic depth of field, slow camera push-in." Luma automatically applies cinematic color grading, so don't fight it; lean into it.

Paramètrès de contrôle du mouvement

Générez 2 à 3 versions de chaque clip. Your success rate at intensity 3-4 is about 80%. At intensity 6+, it drops to 40-50%. Budget your credits accordingly.

Montage : Couper et organiser

Importez tous les clips générés dans votre éditeur. I use DaVinci Resolve for anything longer than 30 seconds and CapCut for quick Reels/TikToks. First pass:

1. Trim the first 0.3-0.5 seconds from every clip (the "morph-in" artifact)
2. Trim the last 0.3-0.5 seconds (degradation zone)
3. Arrange clips in narrative order
4. Add 0.3-0.5 second cross-dissolve transitions between clips

Étalonnage couleur

Les outils vidéo IA produisent des températures de couleur incohérentes entre les clips. Even consecutive generations from the same tool can look different. In DaVinci Resolve:

1. Pick your "hero" clip - the one with the best color
2. Use "Shot Match" to match every other clip to the hero clip's grade
3. Fine-tune: boost shadows slightly (Lift: +0.02), reduce highlights (Gain: -0.03), and add a subtle S-curve to the Lum vs. Sat curve for a polished look
4. Apply a consistent LUT if you have a brand look. FilmConvert and Dehancer have popular presets.

In CapCut, the built-in "Filters" are a faster approximation. The "Film" and "Retro" categories have several options that apply consistent grading across all clips.

Stabilisation

Certains clips générés par IA ont un léger tremblement, especially at higher motion intensities. Apply stabilization in DaVinci Resolve (Edit page > Inspector > Stabilization) with "Translation" mode and smoothness at 0.5. Don't over-stabilize - it creates a floaty, unnatural look.

Enregistrement et génération de voix off

For AI influencer content, you have two options:

• AI voiceover (ElevenLabs): Use the Turbo v2.5 model. Settings: Stability 0.50, Similarity Boost 0.75, Style 0.00 (keep style at zero for natural speech). Export as WAV for best quality. Cost: roughly $0.01-0.02 per sentence.
• Human voiceover: Hire from Fiverr ($15-50 per video). More natural but adds cost and turnaround time. Some creators use their own voice - this works if you're comfortable with the AI influencer having "your" voice.

Sélection musicale

Layer music under voice at -15 to -20 dB relative to the voiceover. For videos without voice, music sits at -6 to -10 dB. Match the BPM to your edit cuts - if you cut every 3 seconds, a 100 BPM track gives you a natural beat to cut on.

Sources: Suno v4 for custom generation, Epidemic Sound ($15/month) for professional library tracks, or Artlist ($17/month) for both music and sound effects.

Design sonore

Three layers make content feel polished:

1. Ambient bed - Room tone, outdoor ambience, or location-specific sound. -20 to -25 dB. Constant throughout the clip.
2. Foley effects - Footsteps, clothing rustle, door sounds, glass clinks. -10 to -15 dB. Sync to on-screen action.
3. Transition effects - Whoosh sounds on cuts, bass drops on reveals. -8 to -12 dB. Use sparingly.

Paramètrès d'export par plateforme

Exportez toujours des fichiers séparés pour chaque plateforme. Never rely on the platform's built-in cropping. TikTok compresses more aggressively than Instagram, so I actually export TikTok versions with slightly higher sharpening (+10-15 in DaVinci Resolve's output sharpening) to compensate.

Optimisation de la taille des fichiers

Instagram recommends files under 250MB. TikTok under 287MB. For 15-30 second videos, you won't hit these limits at the bitrates above. For longer content, use variable bitrate (VBR) with 2-pass encoding in DaVinci Resolve or HandBrake for tighter compression without visible quality loss.

Miniature / Image de couverture

Instagram et TikTok vous permettent de sélectionner une image de couverture. Pick the most visually striking frame in your video - usually the most flattering angle of your AI influencer with the best lighting. On Instagram, you can also upload a custom cover image. Generate a dedicated cover using your image AI tool; it doesn't need to be a frame from the video.