A short film with Kling 4: stills first, 30-second takes, cut like live action

Kling 4 animates well and badly frames shots it invents. So the image model does the cinematography and Kling 4 does the acting, motion, voices and the in-between. This is how fal's Kling 4 Halloween music video was made (YouTube, 2026-10-30): about 18 key stills on GPT Image 2.5, mostly 30-second Kling 4 takes at 1080p 21:9, then an edit that kept the best seconds of each. Prompt rules and the failure table live in kling-4-prompting.

Generation runs on varg: REST https://api.varg.ai/v2 (the varg-ai skill) or the MCP server https://mcp.varg.ai/mcp. Price every take first (POST /v2/estimate or estimate_cost with "model": "kling_v4").

1. Ask, then go

• Story: 3–6 beats, one sentence each (setup, joke or turn, action, payoff, button line). A 60–90 s film is 3–6 takes.
• Cast: a photo or sheet per character (varg-character-design), the outfit written as a list of details.
• Place: a real, named location if possible ("Washington Square Park", "the Brooklyn Banks"). GPT Image 2.5 researches real places well, which Kling 4 text-to-video does not.
• Look: aspect ratio (21:9 for cinema, 9:16 for social), lens (ultra-wide fisheye, 35 mm, telephoto), film stock and grade, grain, motion blur. One look image (a frame you like) used as "reference only for color grade".
• Budget: a 30 s take at 1080p ≈ $7.05 (705 credits); at 720p about two thirds of that. Plan 2–4 takes per beat.

2. Key stills on GPT Image 2.5

One still per moment Kling must hit: the opening of each beat, any shot with a specific angle (over-the-shoulder, overhead), the peak of an action, the end state. Template:

Cinematic <21:9> film still, <angle>, <lens and distortion>.
Use the <N> reference images: Image 1 is <person> to match exactly: <face, hair, glasses,
cap, top with text, trousers, shoes>. Image 2 is <person> to match exactly: <…>.
Image 3 is a reference only for <the bike / the prop> (do NOT show <what to leave out>).
Image 4 is a reference only for color grade, <film stock> tone, grain and motion blur; do not
include its people.

Scene: <who is where doing what, in the present tense, mid-motion, with blur on what moves>.
<Background people, blurred, at the frame edges>. <Light>.
<Film look> like Image 4: <highlight color>, <shadow color>, deep blacks, coarse grain, motion
blur. Avoid text overlays, clean digital look, <wrong props>, <wrong time of day>.

curl -s -X POST https://api.varg.ai/v2/tools/image/call -H "Authorization: Bearer $VARG_API_KEY" \
  -H "Content-Type: application/json" -d '{"model":"gpt_image_2_5","prompt":"<still prompt>",
  "files":[{"url":"CAST_1"},{"url":"CAST_2"},{"url":"PROP"},{"url":"LOOK"}],"aspect_ratio":"21:9"}'

• Check each still before you animate: hands, the bike or board, faces against the cast photos. A warped object in a still becomes the worst part of the take.
• Block the scene across stills so it makes sense in time: if he walks past her in still 1, he is on her other side in still 2.
• Make the second angle of a dialogue (the over-the-shoulder) as a still too; do not leave it to Kling.

3. One take per beat

Pin the reference route (with plain kling_v4, two files become start and end frame and a third is not used):

curl -s -X POST https://api.varg.ai/v2/tools/video/call -H "Authorization: Bearer $VARG_API_KEY" \
  -H "Content-Type: application/json" -d '{
  "model": "fal:fal-ai/kling-video/v4/reference-to-video",
  "prompt": "<beat prompt>",
  "files": [{"url":"STILL_A"},{"url":"STILL_B"},{"url":"STILL_C"}],
  "aspect_ratio": "21:9", "resolution": "1080p", "duration": 30, "audio": true}'

The beat prompt is short: what happens in order, the line in quotes, the camera event, the look, the length.

21:9 the man walks by the woman as she places a white pumpkin on the table. the man says
"what does the white one do?" she smiles and smashes the white pumpkin on the top of the man's
head with both hands. The camera shakes and falls to the ground and fades to black. handheld
shaky 16mm documentary style footage. 15 seconds.

• Order matters when you need it: use provider_options.fal.multi_prompt with seconds per shot (see kling-4-prompting). Keep each take in one place; a location change is a new take, cut in the edit.
• Dialogue: Kling 4 makes the voice itself. For the same voice in every take, make an element with a voice_id (front photo + 1–3 more angles) and write @Element1 says "…".
• Action: add "No slow motion, everything moves at real speed, motion blur."
• Draft at 720p; rerun the chosen take at 1080p with the same stills and prompt.

4. Cut it like live action

A 30 s take holds several cuts, retries of the same move and a ramp-up. Watch it at full speed, mark the 2–6 s pieces that work (a clean landing, a good line read, a new angle you did not ask for) and drop the rest. From the film: "I ended up using the ending part of it", "we used that double tail whip… that's the crash we used", "this shot made the final cut".

ffmpeg -ss 12.4 -to 17.9 -i take3.mp4 -c:v libx264 -crf 16 -c:a aac piece_07.mp4
printf "file '%s'\n" piece_*.mp4 > list.txt && ffmpeg -f concat -safe 0 -i list.txt -c copy film.mp4

Lay the music last over the whole cut (varg-music-video for beat-synced cuts).

5. Fix pass

6. Kling 4 or Seedance 2.5?

With the same pre-made stills, references and lighting, they are "very comparable"; Kling 4 is sharper and does 30 s takes with dialogue, Seedance 2.5 does better with only sheets and a long prompt. Pick Kling 4 when you have (or will make) the stills; pick Seedance 2.5 when you want the video model to invent the shots.

Example plan (60 s, 21:9)

About 13 stills (≈ $3) and 4 takes at 1080p (≈ $26), more for retries; keep 60 s.

Related skills

kling-4-prompting (inputs, prompt rules, failure table), kling-4-action-montage, varg-character-design, varg-consistent-characters, varg-music-video, varg-ai.