A monologue in four angles: one still, one voice, a cut on every pause
A single talking shot reads as an avatar after ten seconds. The same speech cut across four angles of one moment reads as a scene someone filmed with two cameras. The method is from @Arminn_Ai (X, 2026-10-11), who posted the steps and every prompt for a one-minute character monologue:
- "Generated the base shot with Nano Banana 2.1, then 3 more angles of the same moment from it: wide, close-up and a side-angle medium close-up"
- "Cloned the voice with Eleven v4 and generated about a minute of dialogue", "using a different tone in each scene" (this skill designs a voice instead, see section 1)
- "Cut the audio into parts under 15s and fed each part into MiniMax H3 along with its shot"
- "Put the generated clips together in an editing app, and that's it"
Generation runs on varg: REST https://api.varg.ai/v2 (the varg-ai skill) or the MCP server https://mcp.varg.ai/mcp. Price the batch first (POST /v2/estimate or estimate_cost).
1. Ask, then go
- Who speaks and to whom: the character (a photo, a sheet from
varg-character-design, or a description), and whether they look into the lens or at someone beside the camera. - The speech: the topic and the turn. A monologue needs an arc: calm, a shift, a peak, a last line.
- Where: one room, one light, one time of day.
- The voice: a designed voice or the person's own. A real person's voice only with their consent (see
ai-parody-video); an impression is written as a description ("low, tired, dry baritone"), not cloned from recordings. - Length and format: 30 to 90 seconds, 16:9 for a scene, 9:16 for a feed.
2. Write the speech for the ear, with directions in brackets
Six to eight beats, each one tone. Directions go in square brackets before the words they act on; sounds a person makes are directions too.
[low, tired voice, speaking straight to the listener, quiet room late at night] "<line>"
[soft scoff] <line>. It's not magic. It's... [lips smack] <word>.
[measured, like a teacher explaining to a slow student] <line> in the time it takes you to...
[short pause] <button of the beat>.
[tone turning serious, lower] <the shift>.
[quieter, low and intimate] <the detail that lands>.
[a flicker of irritation, slowly building] <line> ATTENTION.
[sharp, raised voice] <line>? [glass set down hard on the table]
[at the absolute peak, every word hammered] <the last line>.
- Ellipses and dashes make pauses; a word in capitals takes the stress.
- Plan the cut points now: a
[short pause]or a sigh every 8 to 14 seconds.
curl -s -X POST https://api.varg.ai/v2/tools/speech/call -H "Authorization: Bearer $VARG_API_KEY" \
-H "Content-Type: application/json" -d '{"model":"eleven_v4","text":"<script with tags>","voice":"<voice id>"}'
Listen to all of it. Regenerate a beat that reads flat rather than fixing it in the picture.
3. Base still, then three more angles of the same moment
Base still on nano_banana_2_1 (or gpt_image_2_5): the medium shot, the character mid-thought, the room readable, practical light named ("warm bedside lamp from frame left").
Then edit that image, one call per angle, with the base as the only reference:
Same scene, same moment, same man, same clothes, same room, same warm lamp light as the
reference. Camera moved to a <wide shot from the foot of the bed, full body, hands on knees /
close-up, head and shoulders, eye level / side-angle medium close-up from his right, 45
degrees>. Nothing in the room changes place. Documentary camera on a tripod, natural skin
texture, natural colour. No text.
Check the four stills together before any video: glasses, hairline, collar, the lamp, what stands on the table. A prop that moves between angles is the first thing viewers catch.
4. Cut the audio at the pauses, one part per angle
ffmpeg -i voice.mp3 -af silencedetect=noise=-35dB:d=0.35 -f null - 2>&1 | grep silence_
# cut in the middle of a silence; every part 5 to 15 s
ffmpeg -i voice.mp3 -ss 0 -to 11.8 -c:a pcm_s16le part1.wav
Assign angles so that no two neighbours match and the size follows the tone: medium to open, medium close-up for the explaining, wide for a gesture or a prop, close-up for the quiet part, medium for the build, close-up for the last line.
5. Lip sync each part with a performance prompt
One job per part: the angle still, the audio part, and a prompt that directs the body, not only the mouth. The template, shortened from the posted prompts:
<static camera.> @Image 1 speaks the dialogue from @Audio 1, matching its voice, tone, timing
and pauses exactly. <Shot size> on a tripod with the faint natural drift of a documentary
camera operator. Framing, lighting, wardrobe and the room stay as in @Image 1. <Light>, natural
skin texture, natural colour.
He is <who>, talking to someone <just beside the lens / straight into the lens>. <Mood of this
part>. His head and upper body move naturally with his speech: small tilts, a slight nod on
stressed words, his shoulders rising and falling with each breath. His hands <rest below the
frame and gesture now and then, visible as movement in his shoulders / are on his knees>.
"<line 1>" <what the eyes, brows, mouth and head do on it>. "<line 2>" <the action: leans in,
eyes narrow, looks over the rim of his glasses>. "<line 3>" <...>. After the last word he stops
speaking and holds the look, breathing quietly.
Irregular natural blinking, small eye movements, slight asymmetry in the face, a swallow
between lines, lips and jaw moving precisely with each syllable.
- Quote every line of the part in order and put its action right after it. The lines tell the model where in the audio each gesture belongs.
- End every prompt with the hold ("stops speaking and holds the look"). It gives the edit a clean out.
- A physical action (clears his throat into a fist, sets a glass down) goes in the wide shot, where hands are in frame.
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": "minimax_h3_max",
"prompt": "<performance prompt>",
"files": [{"url":"ANGLE_STILL"},{"url":"PART_WAV"}],
"resolution": "720p", "duration": 12}'
Tested on varg on 2026-11-12 with a 14 s speech cut into two parts:
minimax_h3_maxwith the still and the audio part infileskeeps your audio as it is (the pauses in the output land on the same milliseconds) and follows the prompt: asked to "turn his eyes to the lens" on the last line, he did. A 6.4 s part at 720p cost $0.56. The output runs about a second longer than the audio, which is the hold you asked for; trim it in the edit.minimax_h3_max_lipsynctakes one image and audio from 5 s up to 15 min and ignores direction: the lips are right, the head barely moves and the eyes stay where the still put them. A 7.8 s part at 480p cost $0.48. Good for a draft of the whole speech on one angle, not for the performance.estimate_costshows the worst case for the lip sync model ($3.60 at 480p); the real charge follows the audio length.- ElevenLabs v4 does not read the bracketed tags aloud;
[sighs]and[whispers]came out as a sigh and a whisper.
Check the model page for current inputs (https://varg.ai/models/llms.txt).
6. Join and check
printf "file '%s'\n" shot1.mp4 shot2.mp4 shot3.mp4 shot4.mp4 shot5.mp4 shot6.mp4 > list.txt
ffmpeg -f concat -safe 0 -i list.txt -c:v libx264 -crf 17 -c:a aac scene.mp4
# then lay the original voice.mp3 over the whole cut so the level and tone are one piece
ffmpeg -i scene.mp4 -i voice.mp3 -map 0:v -map 1:a -c:v copy -c:a aac -shortest final.mp4
| Check | How | Fix |
|---|---|---|
| Lips on the words | Watch each cut point at half speed | Rerun that part; shorten it if it is over 12 s |
| The face is the same person in all four angles | Put a frame from each shot in one row | Regenerate the angle still from the base, not from another angle |
| Props and light hold | Same row of frames | Name the prop and its place in the angle prompt |
| The cut jumps | A cut between two similar sizes | Swap one of them for the wide or the close-up |
| Dead eyes, no body movement | The prompt only described speech | Add the per-line actions and the breathing |
| Room tone drops between parts | Listen on headphones | Use the single voice file over the cut (above); add quiet room tone with video-sound |
Cost of a 60-second monologue (estimate before you spend)
| Step | Count | About |
|---|---|---|
| Base still and three angles, with retries | 6 to 8 images | under $1 |
| Voice, two takes | 2 | under $1 |
Six lip-synced parts of 10 s at 720p on minimax_h3_max, one retry each |
12 runs | about $10 (measured: $0.56 for 6.4 s) |
Related skills
ai-parody-video (the joke, the format, the rules for real people and known characters), scene-recast-video, video-script, varg-character-design, character-passport, video-sound, caption-styles, varg-character-keynote, varg-podcast-video, media-judge, varg-ai.