Skip to the content
Krizaka Blog

04

From a clip to a film: what we got wrong first

Oct 17–21, 202614 min read

What I faced, A film that looked good and was bad; What we built, a chain that asks questions before it shoots

A film strip hung across an open window at golden hour, a small camera on a tripod below it

On 20 October, I watched our film “The World of Krizaka: l'aube” for the second time, 62 seconds in 4K, made with our birds. My verdict was short: horrible. The voices, the captions, the way the birds interact, all of it was too ugly.

It was not a sloppy film, though. The image and the light were beautiful, I said so myself. It had cost 6.5 hours of an H200 GPU, that is $34.42 at the rental company. But a beautiful image does not make a film, and what I had just watched was not a film.

Here is what was wrong, what we changed, what it cost, and what I learned about the difference between improving a prompt and improving a production chain.

Update, October 2026: “l'aube” and “la tournée” are withdrawn. “Le relais” is the brand film. The steps below are the steps of the work, and they serve both.

Two times I had no loop

The first case is the first batch of clips for Orazaka's showcase. I had given Wan a portrait image cropped to 16:9, and a prompt that described a looping movement. The clips were bad. Each clip costs about $1.20 and 18 minutes of H100, so the five I had to redo come to roughly $6 and an hour and a half of GPU that produced nothing publishable. The fix comes down to two lines: images made at the video's aspect ratio, and a prompt that describes only the movement.

Nobody was looking at the image before launching 18 minutes of GPU. Since then, the generation engine writes to its metrics imageKept, the share of the starting image that the frame kept (a portrait cropped to 16:9 keeps a third of it). That does not say whether the result is good, only that it starts with a handicap, but it warns before you pay.

Here is the same shot before and after, at the middle frame of each clip:

Three images of a cable car in the Alps at dawn. Left, before: a portrait image cropped to 16:9, Wan 2.2 A14B, a blurred clip. Middle, after: an image made in 16:9, LTX-Video in 14 seconds, sharp. Right, after: the same image with Wan 2.2 I2V A14B in 40 steps, 18 minutes.
Before: 18 minutes of H100 for an unusable clip. After: 14 seconds for LTX-Video (roughly $0.01), 18 minutes and $1.20 for Wan.
How this was made
Model
Wan 2.2 A14B (before); LTX-Video 0.9.8 13B distilled and Wan 2.2 I2V A14B (after) · Apache-2.0 and LTX-Video Open Weights License
Engine
RunPod H100 80 GB
Machine
hosted by the provider (not our machine)
Through
frames of the image-to-video bake-off (orazaka Measured runs)
Time
not recorded
Cost
not recorded
Sovkora run
not recorded
Note
Middle frame of each clip. Before: a portrait SD 3.5 still cut to 16:9. After: a FLUX.2 [klein] still made at 16:9, LTX-Video in 14 s, Wan in 18 min.

The second case is the birds. For “Le monde de Krizaka: la tournée”, each keyframe is compared to the drawing of the bird made in the code. Some images were refused: a falcon, an owl and a duck duplicated, an audience of humans, one ghost bird too many.

Two rows of the keyframe review sheet. Left, the reference drawing made in the code. Right, three attempts for each row. Top row: the falcon appears twice in all three attempts. Bottom row: an audience of humans was added behind the birds.
Refused: the duplicated falcon (top row) and the audience of humans (bottom row). On the left, the drawing the image must respect.
How this was made
Model
FLUX.2 [klein] 4B (edit mode) · Apache-2.0
Engine
RunPod H100 80 GB
Machine
hosted by the provider (not our machine)
Through
keyframes of the birds film, review sheet
Time
not recorded
Cost
not recorded
Sovkora run
not recorded
Note
Left: the code render the keyframe must match. Three tries each. Rows: the falcon doubled in every try; an audience of humans. Both rejected and remade.

The vertical keyframe at the end was refused three times. I did not insist: the vertical version shows the horizontal shot on its own blurred background. It is the retry ceiling of my loop, held by hand.

I have a real case for what follows. To test a production chain, I made a short demonstration film of 40 seconds, 10 shots, with Orazaka, my AI engine. The production file gives the time of each step. At the H100's $3.99 an hour, that comes to roughly:

An error found on the keyframe costs one cent. Found in the shot, 36 cents. Found in the edit, 157 times the image. So the loop has to check early, not only at the end.

On this film, 12 shots were filmed for 10 kept. At $0.36 a shot, that gives $0.44 per kept shot, 1.2 times the price of one run: exactly 1 ÷ p, with p = 10 out of 12. For the film's 35 runs, 5 of them failed, the GPU computation comes to about $7.89. I cannot give a cost per film: the pod bill, $23.62, mixes several jobs, and I do not put it on one film to get a clean figure.

The 40-second film, 10 shots kept out of 12 filmed. The GPU computation of its runs: 7,115 s, $7.89. · 4.1 MB
How this was made
Model
Wan 2.2 I2V A14B (4-step distill) · Apache-2.0
Engine
RunPod H100 80 GB, 720p, SeedVR2 3B to 1080p; keyframes FLUX.2 [klein] 4B; ACE-Step music; Kokoro voice
Machine
hosted by the provider (not our machine)
Through
orazaka cinema, every step an ordinary Studio run
Size
1080×1920, shown here at 540×960
Time
7115 s, as recorded by the run
Cost
US$7.89 (the provider's price)
Sovkora run
not recorded
Note
10 shots (2 reshot), 12 video runs, 35 runs of which 5 failed; GPU time of its runs 7,115 s, $7.89 at $3.99/h.

What I saw, and heard

I wrote my feedback point by point, because an impression cannot be corrected.

The voice was mechanical and flat, a “GPS effect”: no emotion, no link to the image. The music was far too aggressive. There was no sound design: not one wing beat, not one step, no air. The titles had nothing cinematic about them. The birds moved, but they were not alive, and what they did had nothing to do with the subject. And the story, a disaster, carried nobody.

Five Krizaka birds lined up facing the camera in the grass of a mountain slope at sunset, while the caption says “Each bird carries a promise”
The shot I call the lineup: the birds in a row, facing front, waiting for the voice to finish. Nothing happens between them. Image at 36 seconds of “l'aube” (v1).
How this was made
Model
Wan 2.2 I2V A14B (40 steps) · Apache-2.0
Engine
RunPod H200, 720p then SeedVR2 to 4K, shown at 1536 px
Machine
hosted by the provider (not our machine)
Through
frame at 36 s of « Le monde de Krizaka : l'aube » (v1), withdrawn
Time
not recorded
Cost
not recorded
Sovkora run
not recorded
Note
The birds in a line, facing us, while the narrator says « each bird carries a promise »: the shot Oussama's verdict called a lineup.

The defect was not in the prompt

You might think a better prompt was all it needed. Looking again at how the film had been made, the cause was clear: a short brief went straight to a list of shots. Nobody judged the story before shooting. Each character shot came down to “the bird moves”. The voice had no direction, the music one level, the sound one wind bed.

The defect was in the chain. Here are the steps we added to it. They live in our AI engine, because the goal is not to repair one film, but not to make the same mistake on the next.

1. A treatment before any shot

Before a shot is drawn, a short brief becomes a story. A theme, a four-beat arc in order (the hook, the journey, the turn, the resolution), the intent of each character taken from its promise, and for each shot a goal, an action, a look, the physics of the body and what happens between two characters. The narrator's lines are added with their direction (emotion, pace, stressed word, pause), musical cues by story beat, and the list of sounds for each shot.

2. An automatic critique, twice

The treatment is judged by rules that use no model: is the arc in order, does the turn come after the first third, does every shot have a goal, does every sentence fit its shot (at most 2.6 words per second), does the music start low, peak at the turn and resolve. Then by a model that scores six axes from 1 to 5: transport, arc, characters, life, subject, continuity. The story is shot only if both accept it. Otherwise the reasons go back to the author, twice at most.

For “le relais”, the new film, the treatment had 25 shots for 67 seconds. The rules gave it the top mark, and the model judge gave it 5 out of 5 for transport and 3 out of 5 for life.

3. A living, motivated action

This is the point where the birds became actors. The movement of each shot is now written as a small scene: first this, then that, finally that, with the look and its reason, the weight of the body, the effort of a wing beat, a landing, a catch. The phrases “stands there”, “looks around”, “turns its head left then right” are refused by the rule, by name. The shots are shorter, 2.7 to 4 seconds.

Each take is checked afterwards: faithful to the bird's drawing, and alive and on subject (the energy of the movement is measured frame by frame). On the owl scene, where it disputes then recognises the pigeon, the take where it talks moves on 59% of the frames, the one where it laughs on 69%.

The owl of “l'aube” (v1): it talks. · 0.3 MB
How this was made
Model
Wan 2.2 I2V A14B · Apache-2.0
Engine
RunPod H200, 40 steps, 720p then 4K, shown at 960 px
Machine
hosted by the provider (not our machine)
Through
the owl of « Le monde de Krizaka : l'aube » (v1), same promise line
Time
not recorded
Cost
not recorded
Sovkora run
not recorded
Note
For comparison with the v2 scene. Muted here.
The owl of “le relais”: it disputes, then recognises the pigeon. 5.4 seconds. · 0.3 MB
How this was made
Model
Wan 2.2 I2V A14B · Apache-2.0
Engine
RunPod H100, 40 steps, 720p take shown at 960 px; ACE-Step cue, synthesized foley, Zonos narrator
Machine
hosted by the provider (not our machine)
Through
segment proof of « le relais »: the owl argues, then recognises the pigeon (5.4 s)
Time
not recorded
Cost
not recorded
Sovkora run
not recorded
Note
Muted here. The proof page carries the sound.
Sound
-14.0 LUFS integrated, true peak -1.9 dBTP

Keeping a bird faithful to its drawing took a detour. A free-form image edit had redrawn a brand bird (the falcon had lost its visor). The keyframe now starts from the rendering of the character itself, placed on an empty set, then retouched at a strength of 0.60 for a character and 0.80 for a group: a group of five was drifting at 0.60.

4. A directed voice, chosen blind

We pitted speech synthesis engines against each other, licence first, with the same synthetic reference voice (never a real person's voice). For each candidate, we measured the movement of the voice's pitch in semitones, how well the words were understood with Whisper, and whether the requested pace was kept.

  • Kokoro, the voice of the first version: a single delivery, from 8.6 to 11 semitones.
  • Chatterbox: very expressive in English (17 to 19 semitones), but it ignores a slow pace, and its French dropped words (an error of 0.67 on one sentence).
  • Zonos: an emotion vector, a rate and a pitch variation per sentence, in English and in French. From 11.8 to 19.4 semitones, every word understood, the requested pace kept. Chosen.
  • F5-TTS and XTTS-v2: refused, their weights are non-commercial.

We also tried having an audio model judge the voice, Qwen2-Audio. After calibration, we refused it: it gave about 4 out of 5 to everyone, including the flat voice of the first version, and when asked to choose between two takes it answered “A” 36 times out of 36, in both orders. A judge with a position bias judges nothing. So the voice is measured instead: understanding, pitch movement, pauses, pace. Measurement decides.

We also compared two candidates blind, the voice of the first version and Zonos: each sentence twice, in random order, the answer hidden. The sentences are the birds' promises, such as “The falcon watches: safety first” or “The pigeon delivers, every time”. On the same sentences, the median pitch movement goes from 9.95 semitones for the first voice to 13.6 for the second, and each take of the second is chosen from six at most, first for its accuracy (mean error of 0.033). So it was not an ear that chose, it was the measure.

5. Music that follows the story, and fades under the voice

The first version had one piece of music, at the same level from start to finish, too loud. ACE-Step now makes one piece per story beat: a dark intro, a warm bed, a rise, a resolution. A mixer sets them at a level that depends on the intensity (from −24 dB for 0 to −9 dB for 1, never full), chains them with fades, and lowers the music under the voice (ducking by the voice, 4 to 1, 60 ms attack, 900 ms release). The final result is brought to −14 LUFS within 0.5, with a peak below −1 dBTP, measured a second time and refused if it falls outside.

On the medina test scene, the measurements give −14.0 LUFS and −1.9 dBTP in English, −14.1 LUFS and −1.3 dBTP in French.

6. A sound effect per shot

This is the step where we ran hardest into licences. Every model that makes sound effects from a text or an image had an obstacle: MMAudio is non-commercial, Stable Audio Open asks for a licence acceptance that is an account decision, and HunyuanVideo's excludes the European Union, the United Kingdom and South Korea. So we chose not to use a model. The sound is synthesised from noise, resonators and envelopes: 12 ambiences (rain, storm, wind, lake, waves, a room, dawn, a distant city) and 20 events (wing beats, steps, rustling, paper, an object set down, a clink, bells, bird calls, splashes, a room, a parcel, thunder). A check makes sure no shot is silent and that every character shot has its event.

7. Film titles

The titles of the first version did not look like cinema. They are now written with a subtitle engine: the name in spaced capitals whose spacing slowly opens while it stays on screen, and the subtitle in fine serif italic, arriving 0.45 s later. Fade in 1.1 s, fade out 0.8 s. The captions sit higher than before (16% of the height in 16:9, 26% in vertical), with a soft halo instead of a box, two balanced lines of 42 characters at most, never in the top left corner where the viewer's own “AI-generated” badge goes. Everything sits inside the safe zone (10% margin).

The title “The World of Krizaka” in spaced capitals over an orange horizon at sunset, with “The Relay” in light italics below, above a beach and a string of warm lights
The title test for “le relais”: the spaced name, the italic subtitle arriving after it. This is not the film yet.
How this was made
Model
Title test over a plate (no model for the type)
Engine
libass, Plus Jakarta Sans and Cormorant Garamond (OFL)
Machine
hosted by the provider (not our machine)
Through
title proof of « le relais », frame at 3 s, 16:9
Size
1280×720
Time
not recorded
Cost
not recorded
Sovkora run
not recorded
Note
Not a render of the film: the title test, to judge the type set like a film (a tracked name, an italic serif subtitle arriving after it).

8. 4K, and what it costs

The shots are filmed at 720p with Wan 2.2 I2V A14B in 40 steps, then upscaled to 4K with SeedVR2. The landscapes come from LTX-Video, much faster. The keyframes are made with FLUX.2 [klein] on a Mac, at zero dollars of GPU.

The numbers, in context

For “l'aube”: 23 takes, of which 4 were refused or shortened. The compute time of the runs themselves was 5.5 h ($29.00); the rest is set-up and waiting. The keyframes, the music and the narration were made on a Mac, at $0.

What we measured on an H200 with 141 GB: a Wan 2.2 shot of 81 frames in 720p, 40 steps, takes 1,567 seconds, about 26 minutes, 36 seconds per step. On an H100, it is of the order of 18 minutes per 5-second clip. LTX-Video makes a clip in 11 to 20 seconds. For 4K, SeedVR2 took about 17 minutes per shot with a batch of 5, and 290 seconds with a batch of 21: which makes it possible to run it alongside the generation instead of after it.

For “le relais”, the GPU cost $35.27 in total (measured), under a $40 ceiling, for 68 seconds of 4K film: $31.12 per finished minute, on an H200. About $4.1 of that total was lost to set-up, on two servers that the guard stopped. The tests alone (two sessions: the voice, the titles, one full scene) had cost $6.01 before the full render. “L'aube” had cost $34.42.

What the GPU taught us about money

A GPU you forget costs as much as a GPU that works. So the machine stops by itself: after 20 minutes without work, or after 6 hours at most. It was this guard that cut “l'aube” at the sixth hour, and the same server had to be picked up once more, for about 30 minutes, for one shot retake and three 4K upscales.

A 200 GB network disk, to keep the models at hand, costs about $14 a month doing nothing. We decided to delete it after the last films and to recreate it on demand: redoing the downloads takes a few minutes of GPU, which costs less than keeping it. And on a 94 GB GPU, generation and 4K together ran out of memory: they run one after the other.

The film

“The World of Krizaka: le relais” is 68 seconds in 4K. A parcel, a question: will it arrive? The parcel passes from bird to bird, from city to city, and the music changes genre at each place. You can watch it on krizaka.com, in French and in English.

“Le relais”, 68 s, no subtitles, silent here. The full film has sound. · 6.6 MB
How this was made
Model
Wan 2.2 I2V A14B · Apache-2.0
Engine
RunPod H200, 40 steps, 720p then SeedVR2 to 4K; ACE-Step cues, synthesised foley, Zonos narrator; shown at 960 px, muted here
Machine
hosted by the provider (not our machine)
Through
« Le monde de Krizaka : le relais », the site cut without captions
Time
not recorded
Cost
US$35.27 (the provider's price)
Sovkora run
not recorded
Note
The reference film since 2026-10-21. 68 s in 4K for $35.27 of GPU, $31.12 per finished minute. Everything else ran on a Mac at $0.

Each film is labelled as AI-generated, in the file and in the player, which shows its own badge.

The line not to cross

These films evoke genres: a rap rhythm, a dinner, a match. We evoke them by original means: light, framing, a piece composed for the occasion (here, for example, a rap track with oud and darbuka). No name or face of a real person, no existing song, no game, no broadcast. The characters are original, the voice starts from a synthetic reference, the sound design uses no model under a restrictive licence. The rules are in the chain's data, and a check verifies them before shooting, not after.

To read a licence without making a mistake, I wrote this post. The licences were read at source on 20 October 2026. They change: check again before relying on them. I am not a lawyer: I read these texts as an engineer.

Chapter 04

Results

“L'aube”, GPU billed

$34.42

What I saw, and heard

per finished minute, “Le relais”

$31.12

The numbers, in context