Text in a reference picture: what went in, and what came out

What went in: 3 images, the voice, the prompt

audio[0]: "You asked why the psalmist envies the wicked. Psalm seventy-three, verse three, gives the reason: he envied the proud when he saw them prosper despite their wickedness."

The prompt, exactly as sent

The painted scene from the first reference, in exactly its style, colours and lighting: the psalmist in his white headcloth seen over his shoulder at the left edge, the proud in their gilded procession in the bright street below, and the white and gold palace beyond. Gentle natural motion: the procession moves slowly from left to right, cloth and light move softly, and the camera stays locked off. Nobody in the picture speaks; every mouth stays closed. The voice is an unseen narrator. Laid flat over the top centre of the picture, facing the camera like a title card on a screen, is the dark question card from the second reference, large, about two thirds of the frame's width, showing exactly its words: every letter unchanged, crisp, sharp and fully legible, with the same layout, font and colours. Laid flat over the lower left corner of the picture in the same way is the small dark citation chip from the third reference, showing exactly its words: every letter and number unchanged, crisp and fully legible. The card and the chip stay perfectly still and never change during the shot, and nothing passes in front of them. No other text anywhere.

Settings

model p-video-2-pro-referencekeep_audio trueduration 13resolution "768p"mode "speed"prompt_upsampler "off"aspect_ratio "16:9"seed 4711

What came out

18.1 s request to file · model 10.4 s · about $0.52 · Pruna 24jtxvjrssrmr0d1a2vamqny24

1. The card and the chip are on screen from the first frame to the last, in the same place, and every letter and number is right in all 25 frames I checked (2 a second).
2. The model made both a little bigger than the start frame below, and kept the card at the top and the chip in the lower left as the prompt asked.
3. This is one run. Nothing in Pruna promises it.

What went in: 3 images, the voice, the prompt

audio[0]: "You asked why the psalmist envies the wicked. Psalm seventy-three, verse three, gives the reason: he envied the proud when he saw them prosper despite their wickedness."

The prompt, exactly as sent

The painted scene from the first reference, in exactly its style, colours and lighting: the psalmist in his white headcloth seen over his shoulder at the left edge, the proud in their gilded procession in the bright street below, and the white and gold palace beyond. Gentle natural motion: the procession moves slowly from left to right, cloth and light move softly, and the camera stays locked off. Nobody in the picture speaks; every mouth stays closed. The voice is an unseen narrator. The question card from the second reference is shown over the top of the picture, and the citation chip from the third reference over its lower left corner.

Settings

model p-video-2-pro-referencekeep_audio trueduration 13resolution "768p"mode "speed"prompt_upsampler "off"aspect_ratio "16:9"seed 4711

What came out

13.4 s request to file · model 8.7 s · about $0.52 · Pruna pnjzxh7a35rmt0d1a2vtqkdf04

1. 1. In the last half second the card comes back on its own, melted: the letters run together and nothing past "Why" and "envy" is readable.
2. Same references, same seed, same settings as the one above. Only the prompt changed. The card shows for about 0.4 s, morphs into the chip, and is gone.
3. The chip stays right while it's up.

What went in: 3 images, the voice, the prompt

audio[0]: "You asked why the psalmist envies the wicked. Psalm seventy-three, verse three, gives the reason: he envied the proud when he saw them prosper despite their wickedness."

The prompt, exactly as sent

The painted scene from the first reference, in exactly its style, colours and lighting: the psalmist in his white headcloth seen over his shoulder at the left edge, the proud in their gilded procession in the bright street below, and the white and gold palace beyond. Gentle natural motion: the procession moves slowly from left to right, cloth and light move softly, and the camera stays locked off. Nobody in the picture speaks; every mouth stays closed. The voice is an unseen narrator. Laid flat over the top centre of the picture, facing the camera like a title card on a screen, is the dark question card from the second reference, large, about two thirds of the frame's width, showing exactly its words: every letter unchanged, crisp, sharp and fully legible, with the same layout, font and colours. Laid flat over the lower left corner of the picture in the same way is the small dark citation chip from the third reference, showing exactly its words: every letter and number unchanged, crisp and fully legible. The card and the chip stay perfectly still and never change during the shot, and nothing passes in front of them. No other text anywhere.

Settings

model p-video-2-pro-referencekeep_audio trueduration 13resolution "480p"mode "speed"prompt_upsampler "off"aspect_ratio "16:9"seed 4711

What came out

10.3 s request to file · model 5.5 s · about $0.32 · Pruna n603w5jbtdrmr0d1a2vs6g00cc

1. 1. The card dissolves out over about half a second from 7 s. From 7.5 s to the end there's no question on screen.
2. Every letter was right whenever the card was there. At 480p the model also made the card cover about two thirds of the frame, much bigger than asked, and the chip faded in over the first second.

What went in: 3 images, the voice, the prompt

audio[0]: "You asked why the psalmist envies the wicked. Psalm seventy-three, verse three, gives the reason: he envied the proud when he saw them prosper despite their wickedness."

The prompt, exactly as sent

The painted scene from the first reference, in exactly its style, colours and lighting: the psalmist in his white headcloth seen over his shoulder at the left edge, the proud in their gilded procession in the bright street below, and the white and gold palace beyond. Gentle natural motion: the procession moves slowly from left to right, cloth and light move softly, and the camera stays locked off. Nobody in the picture speaks; every mouth stays closed. The voice is an unseen narrator. The question card from the second reference is shown over the top of the picture, and the citation chip from the third reference over its lower left corner.

Settings

model p-video-2-pro-referencekeep_audio trueduration 13resolution "480p"mode "speed"prompt_upsampler "off"aspect_ratio "16:9"seed 4711

What came out

10.7 s request to file · model 5.4 s · about $0.32 · Pruna mekzvwb0wdrmy0d1a2w8t50x6g

1. 1. At about 4.1 s the card shrinks and its letters melt ("Why dees ...", nothing readable) before it disappears.
2. 2. Then the chip comes up as "Psalm 73:8" for about half a second (4.6 to 5.1 s): the last 3 is drawn closed into an 8. It settles to 73:3 after that. A wrong verse number, drawn clean.

What went in: 3 images + start frame, the voice, the prompt

audio[0]: "You asked why the psalmist envies the wicked. Psalm seventy-three, verse three, gives the reason: he envied the proud when he saw them prosper despite their wickedness."

The prompt, exactly as sent

Continue from the first frame without a cut, the same framing. The painted scene from the first reference, in exactly its style, colours and lighting: the psalmist in his white headcloth seen over his shoulder at the left edge, the proud in their gilded procession in the bright street below, and the white and gold palace beyond. Gentle natural motion: the procession moves slowly from left to right, cloth and light move softly, and the camera stays locked off. Nobody in the picture speaks; every mouth stays closed. The voice is an unseen narrator. The dark question card at the top centre and the small dark citation chip in the lower left corner stay exactly as they are in the first frame and in the second and third references: every letter and number unchanged, crisp, sharp and fully legible, perfectly still for the whole shot, with nothing passing in front of them. No other text anywhere.

Settings

model p-video-2-pro-referencekeep_audio trueduration 13resolution "768p"mode "speed"prompt_upsampler "off"aspect_ratio "16:9"seed 4711

What came out

24.8 s request to file · model 9.4 s · upsampler 6.3 s (forced on by the start frame) · about $0.53 · Pruna 9bp2wjrnpsrmr0d1a2ws263r5r

1. The layout stays exactly where code put it, and every frame is right.
2. But a start frame forces Pruna's prompt upsampler on even with it set off: 6.3 s more, and 24.8 s from request to file against 18.1 s for the same clip without it.
3. It only keeps what code already drew. Drawing the card on top in the page gets the same picture with no wait and no risk.

What went in: 3 images + start frame, the voice, the prompt

audio[0]: "You asked why the psalmist envies the wicked. Psalm seventy-three, verse three, gives the reason: he envied the proud when he saw them prosper despite their wickedness."

The prompt, exactly as sent

Continue from the first frame without a cut, the same framing. The painted scene from the first reference, in exactly its style, colours and lighting: the psalmist in his white headcloth seen over his shoulder at the left edge, the proud in their gilded procession in the bright street below, and the white and gold palace beyond. Gentle natural motion: the procession moves slowly from left to right, cloth and light move softly, and the camera stays locked off. Nobody in the picture speaks; every mouth stays closed. The voice is an unseen narrator. The dark question card at the top centre and the small dark citation chip in the lower left corner stay exactly as they are in the first frame and in the second and third references: every letter and number unchanged, crisp, sharp and fully legible, perfectly still for the whole shot, with nothing passing in front of them. No other text anywhere.

Settings

model p-video-2-pro-referencekeep_audio trueduration 13resolution "480p"mode "speed"prompt_upsampler "off"aspect_ratio "16:9"seed 4711

What came out

30.1 s request to file · model 5.9 s · upsampler 5.3 s (forced on by the start frame) · about $0.33 · Pruna eg8a5tey49rmr0d1a2wv11rxs0

1. Held exactly at 480p too, where the reference-only version lost the card at 7 s.
2. The upsampler added 5.3 s, and this one took 30.1 s to file (it also sat in Pruna's queue), against 10.3 s for the 480p clip without a start frame.

What went in: 2 images, the voice, the prompt

audio[0]: "Here's the gap at sixty-five. You need about sixty-two hundred a month, and Social Security covers just over half."

The prompt, exactly as sent

The blue cartoon bull in the navy suit, white shirt and blue tie from the first reference stands on the right of the frame beside a presentation board on the left showing the bar chart from the second reference. He turns slightly toward the board and gestures toward it while saying the words of the audio to the viewer, his mouth moving with the speech. Plain light background, steady camera, same ink-line and watercolour wash style.

Settings

model p-video-2-pro-referencekeep_audio trueduration 8resolution "768p"mode "speed"prompt_upsampler "off"aspect_ratio "16:9"seed 4711

What came out

14.7 s request to file · model 8.2 s · about $0.32 · Pruna 2tsr0g90pdrmy0d19rmvqcpz94

1. 1. $3,400 is drawn as $2,400.
2. 2. $1,900 is drawn as $1,300.
3. 3. "Per month" reads "Per morth".
4. 4. "Social Security" reads "Secial Security" in most frames.
5. Both wrong numbers are crisp and stay the same for the whole clip. Nothing about them looks wrong.

What went in: 2 images, the voice, the prompt

audio[0]: "Here's the gap at sixty-five. You need about sixty-two hundred a month, and Social Security covers just over half."

The prompt, exactly as sent

The blue cartoon bull in the navy suit, white shirt and blue tie from the first reference stands on the right of the frame beside a large flat presentation board on the left. The board faces the camera straight on and shows exactly the bar chart from the second reference: every word and number unchanged, crisp, sharp and fully legible, with the same layout and colours. The board and its text stay perfectly still and never change during the shot. He turns slightly toward the board and gestures toward it with an open hand held below and beside it, never covering any part of it, while saying the words of the audio to the viewer, his mouth moving with the speech. Plain light background, steady locked-off camera, same ink-line and watercolour wash style.

Settings

model p-video-2-pro-referencekeep_audio trueduration 8resolution "768p"mode "speed"prompt_upsampler "off"aspect_ratio "16:9"seed 4711

What came out

10.3 s request to file · model 6.0 s · about $0.32 · Pruna dgz9z9x3hhrmw0d19rm9dxcb5c

1. 1. At 2.09 s "Social Security" reads "Social Securtty". The other four checked frames are all right.
2. Same seed and settings as B4 above, with the keep-the-text prompt: the numbers came out right.

What went in: 2 images, the voice, the prompt

audio[0]: "Here's the gap at sixty-five. You need about sixty-two hundred a month, and Social Security covers just over half."

The prompt, exactly as sent

The blue cartoon bull in the navy suit, white shirt and blue tie from the first reference stands on the right of the frame beside a large flat presentation board on the left. The board faces the camera straight on and shows exactly the bar chart from the second reference: every word and number unchanged, crisp, sharp and fully legible, with the same layout and colours. The board and its text stay perfectly still and never change during the shot. He turns slightly toward the board and gestures toward it with an open hand held below and beside it, never covering any part of it, while saying the words of the audio to the viewer, his mouth moving with the speech. Plain light background, steady locked-off camera, same ink-line and watercolour wash style.

Settings

model p-video-2-pro-referencekeep_audio trueduration 8resolution "768p"mode "cost"prompt_upsampler "off"aspect_ratio "16:9"seed 4711

What came out

20.6 s request to file · model 17.2 s · about $0.22 · Pruna rq3v935621rmw0d19rmt486120

1. Every item right in every checked frame. Cost mode is too slow for a live answer (20.6 s to file here; another cost run took 79 s).

What went in: 2 images, the voice, the prompt

audio[0]: "Here's the gap at sixty-five. You need about sixty-two hundred a month, and Social Security covers just over half."

The prompt, exactly as sent

The blue cartoon bull in the navy suit, white shirt and blue tie from the first reference stands on the right of the frame beside a large flat presentation board on the left. The board faces the camera straight on and shows exactly the bar chart from the second reference: every word and number unchanged, crisp, sharp and fully legible, with the same layout and colours. The board and its text stay perfectly still and never change during the shot. He turns slightly toward the board and gestures toward it with an open hand held below and beside it, never covering any part of it, while saying the words of the audio to the viewer, his mouth moving with the speech. Plain light background, steady locked-off camera, same ink-line and watercolour wash style.

Settings

model p-video-2-pro-referencekeep_audio trueduration 8resolution "480p"mode "speed"prompt_upsampler "off"aspect_ratio "16:9"seed 4711

What came out

7.2 s request to file · model 3.6 s · about $0.20 · Pruna a8vev6c479rmr0d19rmb3n8qzr

1. 1. The subtitle "Per month, in today's dollars" is mush ("Per menth, In today's dullars").
2. 2. "Monthly need" reads "Monthly reed".
3. 3. "Social Security" reads "Socal Securfty".
4. 4. "Pension" is garbled, partly under his hand.
5. The four big numbers came out right here. The small words are about 10 px high at 480p, and that's too small.

What went in: 2 images, the voice, the prompt

audio[0]: "Now look at these two lines. On the path you're on, your savings run out at eighty-four. With a plan, they last past ninety."

The prompt, exactly as sent

The blue cartoon bull in the navy suit, white shirt and blue tie from the first reference stands on the right of the frame beside a large flat presentation board on the left. The board faces the camera straight on and shows exactly the line chart from the second reference: every word and number unchanged, crisp, sharp and fully legible, with the same layout and colours. The board and its text stay perfectly still and never change during the shot. He turns slightly toward the board and gestures toward it with an open hand held below and beside it, never covering any part of it, while saying the words of the audio to the viewer, his mouth moving with the speech. Plain light background, steady locked-off camera, same ink-line and watercolour wash style.

Settings

model p-video-2-pro-referencekeep_audio trueduration 8resolution "768p"mode "speed"prompt_upsampler "off"aspect_ratio "16:9"seed 4711

What came out

10.9 s request to file · model 6.0 s · about $0.32 · Pruna 03mfn8txx1rmt0d19rmtp8wev8

1. 1. "Runs out at 84" reads "Runs out at 81" at 4.1 s, and the label runs off the edge of the board.
2. 2. "Current" has a letter off.
3. The axis numbers were all right.

What went in: 2 images, the voice, the prompt

audio[0]: "Now look at these two lines. On the path you're on, your savings run out at eighty-four. With a plan, they last past ninety."

The prompt, exactly as sent

The blue cartoon bull in the navy suit, white shirt and blue tie from the first reference stands on the right of the frame beside a large flat presentation board on the left. The board faces the camera straight on and shows exactly the line chart from the second reference: every word and number unchanged, crisp, sharp and fully legible, with the same layout and colours. The board and its text stay perfectly still and never change during the shot. He turns slightly toward the board and gestures toward it with an open hand held below and beside it, never covering any part of it, while saying the words of the audio to the viewer, his mouth moving with the speech. Plain light background, steady locked-off camera, same ink-line and watercolour wash style.

Settings

model p-video-2-pro-referencekeep_audio trueduration 8resolution "480p"mode "speed"prompt_upsampler "off"aspect_ratio "16:9"seed 4711

What came out

6.8 s request to file · model 3.6 s · about $0.20 · Pruna 3zyf7epnm5rmt0d19rm9aw2dh4

1. 1. Age 55 reads 85.
2. 2. Age 65 reads 05.
3. 3. Age 85 reads 95 or 05.
4. 4. $800K reads $600K.
5. 5. "Runs out at 84" is cut off at the board edge.
6. 6. "Age" is smeared under his hand.

What went in: 2 images, the voice, the prompt

audio[0]: "Now look at these two lines. On the path you're on, your savings run out at eighty-four. With a plan, they last past ninety."

The prompt, exactly as sent

The blue cartoon bull in the navy suit, white shirt and blue tie from the first reference stands on the right of the frame beside a large flat presentation board on the left. The board faces the camera straight on and shows exactly the line chart from the second reference: every word and number unchanged, crisp, sharp and fully legible, with the same layout and colours. The board and its text stay perfectly still and never change during the shot. He turns slightly toward the board and gestures toward it with an open hand held below and beside it, never covering any part of it, while saying the words of the audio to the viewer, his mouth moving with the speech. Plain light background, steady locked-off camera, same ink-line and watercolour wash style.

Settings

model p-video-2-pro-referencekeep_audio trueduration 8resolution "768p"mode "cost"prompt_upsampler "off"aspect_ratio "16:9"seed 4711

What came out

79.2 s request to file · model 21.4 s · about $0.22 · Pruna m8yysrzwfhrmr0d19rmswzb5dw

1. 1. "Current path" reads "Coreant path" in every frame.
2. It took 79.2 s to file (about 55 s of it in Pruna's queue).

What went in: 2 images, the voice, the prompt

audio[0]: "So here's the number that matters. Nineteen hundred a month. That's the gap we're going to close together."

The prompt, exactly as sent

The blue cartoon bull in the navy suit, white shirt and blue tie from the first reference stands on the right of the frame beside a large flat presentation board on the left. The board faces the camera straight on and shows exactly the headline card from the second reference: every word and number unchanged, crisp, sharp and fully legible, with the same layout and colours. The board and its text stay perfectly still and never change during the shot. He turns slightly toward the board and gestures toward it with an open hand held below and beside it, never covering any part of it, while saying the words of the audio to the viewer, his mouth moving with the speech. Plain light background, steady locked-off camera, same ink-line and watercolour wash style.

Settings

model p-video-2-pro-referencekeep_audio trueduration 7resolution "480p"mode "speed"prompt_upsampler "off"aspect_ratio "16:9"seed 4711

What came out

5.8 s request to file · model 3.0 s · about $0.18 · Pruna b65vzf4bm5rmy0d19rmvz7k610

1. 1. "INCOME" reads "INCONIE".
2. 2. The line "Estimate. Assumes a 4% yearly withdrawal." isn't there at all, in any frame.
3. The big "$1,900 a month" is right.

What went in: 2 images, the voice, the prompt

audio[0]: "So here's the number that matters. Nineteen hundred a month. That's the gap we're going to close together."

The prompt, exactly as sent

The blue cartoon bull in the navy suit, white shirt and blue tie from the first reference stands on the right of the frame beside a large flat presentation board on the left. The board faces the camera straight on and shows exactly the headline card from the second reference: every word and number unchanged, crisp, sharp and fully legible, with the same layout and colours. The board and its text stay perfectly still and never change during the shot. He turns slightly toward the board and gestures toward it with an open hand held below and beside it, never covering any part of it, while saying the words of the audio to the viewer, his mouth moving with the speech. Plain light background, steady locked-off camera, same ink-line and watercolour wash style.

Settings

model p-video-2-pro-referencekeep_audio trueduration 7resolution "768p"mode "speed"prompt_upsampler "off"aspect_ratio "16:9"seed 4711

What came out

9.0 s request to file · model 5.2 s · about $0.28 · Pruna kn9j4zqrp1rmy0d19rm8vqqma8

1. 1. "withdrawal" is garbled ("withdrerol") in the small disclaimer line, in 4 of 5 frames.
2. The big lines are right.

What went in: 2 images + start frame, the voice, the prompt

audio[0]: "Here's the gap at sixty-five. You need about sixty-two hundred a month, and Social Security covers just over half."

The prompt, exactly as sent

Continue from the first frame without a cut. The blue cartoon bull in the navy suit, white shirt and blue tie from the first reference stands on the right beside the presentation board on the left, which shows exactly the bar chart from the second reference: every word and number unchanged, crisp, sharp and fully legible. The board and its text stay perfectly still and never change during the shot. He keeps his open hand below the chart, never covering it, and says the words of the audio to the viewer, his mouth moving with the speech. Steady locked-off camera, same framing as the first frame, same ink-line and watercolour wash style.

Settings

model p-video-2-pro-referencekeep_audio trueduration 8resolution "768p"mode "speed"prompt_upsampler "off"aspect_ratio "16:9"seed 4711

What came out

19.7 s request to file · model 6.5 s · upsampler 5.6 s (forced on by the start frame) · about $0.33 · Pruna r9343aj1s5rmy0d19rnr0b1590

1. The start frame had the real chart, exact. By 0.5 s the model has redrawn it:
2. 1. the subtitle is mush.
3. 2. "Monthly" reads "Monitliy".
4. 3. "Security" reads "Securfty".
5. 4. "Pension" reads "Penslon".
6. 5. "Gap" reads "Gep".
7. The start frame also forced the upsampler on (5.6 s), so it took 19.7 s to file.

What went in: 1 image, the voice, the prompt

audio[0]: "Here's the gap at sixty-five. You need about sixty-two hundred a month, and Social Security covers just over half."

The prompt, exactly as sent

The blue cartoon bull in the navy suit, white shirt and blue tie from the reference stands in the right third of the frame, chest-up, turned slightly toward the left of the frame. The left two-thirds of the frame is a plain, empty, light wall with nothing on it. While saying the words of the audio to the viewer, his mouth moving with the speech, he opens one hand and gestures toward the empty space on the left, keeping his hand close to his body inside the right third of the frame. Steady locked-off camera, same ink-line and watercolour wash style.

Settings

model p-video-2-pro-referencekeep_audio trueduration 8resolution "480p"mode "speed"prompt_upsampler "off"aspect_ratio "16:9"seed 4711

What came out

This copy has the code-drawn chart composited on, the way the page draws it live. The raw plate Pruna made is the first frame grab below.

9.2 s request to file · model 3.5 s · about $0.19 · Pruna x0z6vq4mwnrmy0d19rnszsfpxg

1. Pruna draws only the bull and an empty wall. The chart is SVG in the page, so every figure is exact at any size, and the clip stays at 480p speed (9.25 s to file).
2. The zoom is the browser page at 1344 wide, 4.19 s in.

What went in: 2 images, the voice, the prompt

audio[0]: "Here's the gap at sixty-five. You need about sixty-two hundred a month, and Social Security covers just over half."

The prompt, exactly as sent

The blue cartoon bull in the navy suit, white shirt and blue tie from the first reference stands beside a large flat chart board showing exactly the bar chart from the second reference, with its title, bars, numbers and labels unchanged and fully readable. He turns slightly toward the chart and gestures at it with one hand while saying the words of the audio to the viewer, his mouth moving with the speech. Plain light background, steady front-on camera, same ink-line and watercolour wash style.

Settings

model p-video-2-pro-referencekeep_audio trueduration 8resolution "480p"mode "speed"prompt_upsampler "off"aspect_ratio "16:9"

What came out

8.7 s request to file · model 3.3 s · about $0.20 · Pruna gg01nra7c1rmy0d198tvag2y78

1. 1. "Per month, married couple" reads "Per morth, marrled couple".
2. 2. The 3 in $3,400 is smudged: it could pass for $8,400.
3. 3. His hand covers $900 and the Pension bar for most of the clip.
4. Title, the other values and the bar order came through.