text-to-image-to-video

name	text-to-image-to-video
description	Two-stage pipeline that generates a still image from a text prompt (Flux/SDXL/Imagen) then animates it into a video clip (Kling/Runway/Hailuo/Wan). Use when the user wants tight visual control over the opening frame before motion.

Text → Image → Video Pipeline

Two-stage chain. Stage 1 produces a still you can iterate on cheaply; stage 2 animates the chosen still.

When to use

• The shot needs a specific composition or character pose locked before motion.
• Direct text-to-video drifts off-prompt; this gives you an anchor frame.
• Cost: one image gen + one video gen per accepted take. Iterating on the still is cheap; iterating on motion is not — get the still right first.

Inputs

• A shot brief from scripts/storyboards/NN-*.md (visual prompt seed, duration, character refs).
• Model selection from brief/tools-and-models.md (text-to-image model + image-to-video model).
• Optional: character reference image from characters/<name>.md.

Steps

1. Compose image prompt. Combine the shot's visual seed + character seed prompt + style/vibe from brief/creative-brief.md. Show it to the user before generating.
2. Generate still. Call the configured text-to-image model (typically via Fal or Replicate MCP). Save to generation/text-to-image/NN-shortname-vN.png. Save prompt + model + seed to generation/prompts/NN-shortname-vN.md. Append to logs/production-log.md.
3. Iterate on still. Show the image. If the user wants to refine, loop on step 2 with an incremented vN. Don't proceed until the user approves a still.
4. Compose motion prompt. Describe the motion — camera move, subject action, duration. Pull duration from the shot brief.
5. Animate. Call the configured image-to-video model with the approved still as input. Save to generation/image-to-video/NN-shortname-vN.mp4. Save the motion prompt and parameters to generation/prompts/NN-shortname-vN-motion.md. Log it.
6. Surface result. Show the clip. Offer: accept (copy to clips/raw/ and suggest /promote-take), retry motion, or go back to step 2 (new still).

Common gotchas

• Aspect ratio mismatch. The still must match the target video aspect from brief/creative-brief.md. Set it explicitly in the image gen call; don't rely on the model's default.
• Character drift across shots. Always pass the character's seed prompt verbatim. Consider using a character reference image as a control input if the model supports it.
• Motion model resolution caps. Some image-to-video models downscale. If final output needs 4K, pair with the upscale-and-interpolate skill.