Bad. Very Bad.
A Byzantine haggler tells an arms dealer his legendary sword is "bad, bad, very bad", knocks the price from 500 to 60 gold, pays like it physically hurts, and walks off deadpan. A 3D cube turn later he bursts into his home holding it over his head, grinning like a lunatic. He was lying. Pictures were made on Aaron's laptop (Qwen-Image 2.1 stills, a Blender previs, one 13-segment MiniMax-H3 latent chain), every sound came from ElevenLabs, and the edit is a frame-exact Python compositor.
- Length
- 29.9 s, landscape
- Picture
- 1920×1080, 60 fps, H.264
- Sound
- ElevenLabs only, −14 LUFS
- GPU
- 108 min on the laptop
- ElevenLabs
- 1,691 credits
- Wall clock
- about 2 h 40 min
What the edit adds
Left: the upscaled MiniMax-H3 frame the shot uses at that moment. Right: the same moment of the final cut. The edit adds the camera moves, text behind the subject, impact frames, the HUD, the RPG price card, the grades and the 2.5D background layer; the drawing itself is H3's.
The cut
Nineteen frames from the final 60 fps file.
What I learned
- One shot per continuation plus an end keyframe on every segment held 13 H3 segments on one look. No drift at 1344×768, and the chain never spilled out of 8 GB.
- H3's own audio is the lip-sync map. Its voice-band bursts are syllable-accurate, and a piecewise time map puts the ElevenLabs words on the mouths.
- Read every strip for H3's extra cuts. Segments 9 and 10 each held two shots; a vertical step in the time map jumps across the cut.
- A word Whisper can't read on its own is a bad take, not a mix problem. '[horrified gasp] Sixty?!' and a sulky 'Fine.' were re-voiced for 40 credits once the voices were saved. The buried 'Very bad.' was a level problem, fixed by a voice leveler and deeper ducking.
- Put a hit on the beat by moving the cut, not the hit. The flick shot opens a 16th early so the finger touches the blade on the downbeat.
- isnet-anime mattes are good until a busy textile sits behind a close-up. Cut it out per time range, and keep the white outline off where the matte is ragged; check outlined shots at full size.
- When a music model ignores section lengths, build the score from parts: a 30 s groove plus a separate drop conditioned on it, cut on a measured 104.94 BPM grid.
Time and cost
| Stage | Measured | Notes |
|---|---|---|
| Qwen-Image 2.1 stills | 28 jobs, 1,649 s GPU | sheets 36-42 s, plates 28 s, keyframe edits 43-87 s, 2048×1152 hero 116-136 s |
| MiniMax-H3 latent chain | 13 segments, 4,636 s GPU | 310 s first, 301-453 s per continuation, peak 7.6 GB, one lease per clip |
| Anime upscale (realesr-animevideov3 ×4) | 1,022 frames, 195 s GPU | 0.17 s a frame, to 2688×1512 |
| Blender 5.2 previs | 16 shots, headless | not timed |
| Compositor render | 1,794 frames, 399 s CPU | 8 workers, x264 slow CRF 16 |
| ElevenLabs | 1,691 credits | music 894, SFX 384, voice design 334, TTS 79 |
| Whole job, wall clock | about 2 h 40 min | brief written 12:36, final render 15:17; the H3 chain alone 13:33-14:52 |
The H3 latent chain
The haggle is ONE MiniMax-H3 latent chain (Aaron's workflows 31/32 through h3-film's film.py): every segment continues the saved AV latent of the one before, so the look, faces and light carry across the cuts. Cuts mode, 1344x768, a hard cut at about 0.9 s into each segment, every segment pinned to its keyframe. H3's own audio was discarded; its speech bursts were only used to find where the mouths move.
Segment 1
First frame = the repaired market plate: a locked-off wide street with doves. The edit adds the slow push and the title from 0.3 s.
310.0 s GPU · peak 7343 MB
Prompt
This is an establishing shot of a Byzantine market street; the townspeople are small figures in the background. A covered market street in Constantinople in the year 1043 at golden hour: pale stone arcades with round arches, colourful striped silk awnings, hanging brass lamps, stalls of spices and copper pots, and the golden dome of the Hagia Sophia at the far end of the street. No modern objects. A modern Japanese TV anime come to life: bold clean black ink outlines, flat cel shading with crisp hard-edged shadows, vivid saturated colours, richly painted anime backgrounds, warm golden late-afternoon light. Not 3D, not CGI, not a photo. The camera pushes in slowly and steadily down the street; one continuous shot, no cuts. [0s-5s] Wide establishing shot of the market street at golden hour: townspeople stroll between the stalls, the striped awnings ripple in the breeze and a flock of white pigeons bursts up into the glowing orange sky toward the golden dome at the far end of the street. Audio: a lively market crowd murmur, footsteps on stone, pigeons flapping their wings, no music.
Segment 2
Nikos's old hand strokes the velvet beside the sword. The edit tracks a sparkle onto the blade's brightest point.
306.0 s GPU · peak 7615 MB
Prompt
Continue the same moment. Only the old arms dealer's hand and sleeve are in this shot, no face. The sword is the golden sword: a long straight polished steel blade, a golden cross-guard with curled ends, a red leather grip and a golden pommel with one big ruby. A modern Japanese TV anime come to life: bold clean black ink outlines, flat cel shading with crisp hard-edged shadows, vivid saturated colours, richly painted anime backgrounds, warm golden late-afternoon light. Not 3D, not CGI, not a photo. Completely still, locked-off static camera in every shot: no camera movement at all, no pan, no zoom, no push in; shots change only by hard cuts. [0s-1s] The previous shot holds, unchanged. [1s-4s] At exactly the 1 second mark, hard cut to an extreme close-up from above of the golden sword lying diagonally on the dark red velvet of the counter: the bony old hand of the arms dealer in a mustard-yellow sleeve sweeps slowly over the blade in a grand presenting gesture, and a bright star-shaped glint of light travels along the polished blade from the guard to the tip. This one static shot fills the whole rest of the clip, from the 1 second mark to the end; the wide shot does not come back. Audio: a lively market crowd murmur, a soft magical metallic shimmer, no music.
Segment 3
The low-angle entrance, arms crossed, cloak in the wind; H3 made its hard cut at 0.83 s (asked for 1 s). BASIL goes behind him on the matte.
301.4 s GPU · peak 7433 MB
Prompt
Continue the same moment. The man is Basil, a big burly Byzantine man about 45 years old: a huge thick square black beard, heavy black eyebrows, a big hooked nose, a round brown felt cap, a deep-blue knee-length tunic with gold embroidered bands at the collar and cuffs, and a crimson cloak pinned at his right shoulder with a round gold brooch. A modern Japanese TV anime come to life: bold clean black ink outlines, flat cel shading with crisp hard-edged shadows, vivid saturated colours, richly painted anime backgrounds, warm golden late-afternoon light. Not 3D, not CGI, not a photo. A static low-angle camera with only a very slight slow push in. [0s-1s] The previous shot holds, unchanged. [1s-4s] At exactly the 1 second mark, hard cut to a dramatic low-angle shot looking up at Basil, who stands with his arms crossed and towers over the camera against the orange sunset sky: his red cloak blows in the wind, and he slowly narrows his eyes and frowns down at the camera, unimpressed. This one static shot fills the whole rest of the clip, from the 1 second mark to the end; the wide shot does not come back. Audio: wind gusts, a heavy cloak flapping, a lively market crowd murmur, no music.
Segment 4
'Bad. Bad. Very bad.' at the stall. H3's audio gave three clean speech bursts, so a time map put each ElevenLabs word on the mouth. Used twice, around the flinch cut-in.
453.4 s GPU · peak 7433 MB
Prompt
Continue the same moment. The man is Basil, a big burly Byzantine man about 45 years old: a huge thick square black beard, heavy black eyebrows, a big hooked nose, a round brown felt cap, a deep-blue knee-length tunic with gold embroidered bands at the collar and cuffs, and a crimson cloak pinned at his right shoulder with a round gold brooch. A modern Japanese TV anime come to life: bold clean black ink outlines, flat cel shading with crisp hard-edged shadows, vivid saturated colours, richly painted anime backgrounds, warm golden late-afternoon light. Not 3D, not CGI, not a photo. Completely still, locked-off static camera in every shot: no camera movement at all, no pan, no zoom, no push in; shots change only by hard cuts. [0s-1s] The previous shot holds, unchanged. [1s-4.5s] At exactly the 1 second mark, hard cut to a medium close-up of Basil at the stall, turned toward the right: he shakes his head slowly in disgust and says 'Bad.', then again 'Bad.', then waves one big hand dismissively and scoffs 'Very bad.', his eyes half closed and his nose wrinkled. This one static shot fills the whole rest of the clip, from the 1 second mark to the end; the wide shot does not come back. Audio: a lively market crowd murmur; a deep gruff man's voice with a thick Greek accent says 'Bad. Bad. Very bad.' with disgust, no music.
Segment 5
Opened on the back of Nikos's head and a stray white blade for about a second, then settled into the keyframe; the shot starts at clip frame 49.
395.5 s GPU · peak 7433 MB
Prompt
Continue the same moment. The old man is Old Nikos, a skinny old arms dealer about 70: a long pointed white beard, a bald head with white wisps above big ears, a long hooked nose, a gap-toothed grin, a mustard-yellow robe and a small dark-green cap. A modern Japanese TV anime come to life: bold clean black ink outlines, flat cel shading with crisp hard-edged shadows, vivid saturated colours, richly painted anime backgrounds, warm golden late-afternoon light. Not 3D, not CGI, not a photo. Completely still, locked-off static camera in every shot: no camera movement at all, no pan, no zoom, no push in; shots change only by hard cuts. [0s-1s] The previous shot holds, unchanged. [1s-3.5s] At exactly the 1 second mark, hard cut to a close-up of Old Nikos behind his counter, turned toward the left: he flinches backwards as if struck, his eyes wide, his grin freezing into a nervous grimace as a big drop of sweat runs down his bald head. This one static shot fills the whole rest of the clip, from the 1 second mark to the end; the wide shot does not come back. Audio: a lively market crowd murmur, a sharp nervous gasp, no music.
Segment 6
'You charge money for THIS?' with a finger on the blade. The mouth is mapped onto 'You' and 'THIS'; the punch-in lands on the beat.
374.5 s GPU · peak 7433 MB
Prompt
Continue the same moment. The man is Basil, a big burly Byzantine man about 45 years old: a huge thick square black beard, heavy black eyebrows, a big hooked nose, a round brown felt cap, a deep-blue knee-length tunic with gold embroidered bands at the collar and cuffs, and a crimson cloak pinned at his right shoulder with a round gold brooch. The sword is the golden sword: a long straight polished steel blade, a golden cross-guard with curled ends, a red leather grip and a golden pommel with one big ruby. A modern Japanese TV anime come to life: bold clean black ink outlines, flat cel shading with crisp hard-edged shadows, vivid saturated colours, richly painted anime backgrounds, warm golden late-afternoon light. Not 3D, not CGI, not a photo. Completely still, locked-off static camera in every shot: no camera movement at all, no pan, no zoom, no push in; shots change only by hard cuts. [0s-1s] The previous shot holds, unchanged. [1s-4s] At exactly the 1 second mark, hard cut to a medium shot from behind the counter: Basil leans forward over the counter toward the camera and jabs his index finger down at the golden sword lying on the velvet in the foreground, his eyebrows raised high in disbelief, and shouts 'You charge money for THIS?' This one static shot fills the whole rest of the clip, from the 1 second mark to the end; the wide shot does not come back. Audio: a lively market crowd murmur; a deep gruff man's voice with a thick Greek accent shouts 'You charge money for this?', no music.
Segment 7
Sighting down the blade. The HUD finds the bright, unsaturated steel in every frame, so its brackets follow the blade as he lifts it.
376.5 s GPU · peak 7473 MB
Prompt
Continue the same moment. The man is Basil, a big burly Byzantine man about 45 years old: a huge thick square black beard, heavy black eyebrows, a big hooked nose, a round brown felt cap, a deep-blue knee-length tunic with gold embroidered bands at the collar and cuffs, and a crimson cloak pinned at his right shoulder with a round gold brooch. The sword is the golden sword: a long straight polished steel blade, a golden cross-guard with curled ends, a red leather grip and a golden pommel with one big ruby. A modern Japanese TV anime come to life: bold clean black ink outlines, flat cel shading with crisp hard-edged shadows, vivid saturated colours, richly painted anime backgrounds, warm golden late-afternoon light. Not 3D, not CGI, not a photo. Completely still, locked-off static camera in every shot: no camera movement at all, no pan, no zoom, no push in; shots change only by hard cuts. [0s-1s] The previous shot holds, unchanged. [1s-4s] At exactly the 1 second mark, hard cut to a close-up of Basil's face turned toward the right: he lifts the golden sword horizontally to the height of his eyes and slowly sights along the gleaming blade with one eye shut and the other narrowed in suspicion, as a sharp line of light glides along the steel. This one static shot fills the whole rest of the clip, from the 1 second mark to the end; the wide shot does not come back. Audio: a soft metallic slide, a deep gruff man's voice with a thick Greek accent hums a suspicious 'Hmm...', a lively market crowd murmur, no music.
Segment 8
The finger touches the blade at 1.08 s and the glint follows; H3's own audio rang at 1.06 s. The shot opens a 16th early so the touch lands on the downbeat.
372.5 s GPU · peak 7433 MB
Prompt
Continue the same moment. Only his hand and the sword are sharp; his bearded face is a soft blur in the background. The sword is the golden sword: a long straight polished steel blade, a golden cross-guard with curled ends, a red leather grip and a golden pommel with one big ruby. A modern Japanese TV anime come to life: bold clean black ink outlines, flat cel shading with crisp hard-edged shadows, vivid saturated colours, richly painted anime backgrounds, warm golden late-afternoon light. Not 3D, not CGI, not a photo. Completely still, locked-off static camera in every shot: no camera movement at all, no pan, no zoom, no push in; shots change only by hard cuts. [0s-1s] The previous shot holds, unchanged. [1s-4s] At exactly the 1 second mark, hard cut to an extreme close-up of the polished blade crossing the frame: at the 1.5 second mark a thick finger flicks the flat of the blade with its fingernail; the blade quivers and a bright ring of white light flashes along the steel with tiny sparkles. This one static shot fills the whole rest of the clip, from the 1 second mark to the end; the wide shot does not come back. Audio: one bright pure metallic ring like a bell with a long shimmering sustain, no music.
Segment 9
H3 made its own extra cut: a close-up for 'No, no', then the raised-finger medium of the keyframe for 'Not good at all', four syllables, four bursts. The close-up's matte swallowed a tapestry, so the white outline is off there.
345.1 s GPU · peak 7617 MB
Prompt
Continue the same moment. The man is Basil, a big burly Byzantine man about 45 years old: a huge thick square black beard, heavy black eyebrows, a big hooked nose, a round brown felt cap, a deep-blue knee-length tunic with gold embroidered bands at the collar and cuffs, and a crimson cloak pinned at his right shoulder with a round gold brooch. A modern Japanese TV anime come to life: bold clean black ink outlines, flat cel shading with crisp hard-edged shadows, vivid saturated colours, richly painted anime backgrounds, warm golden late-afternoon light. Not 3D, not CGI, not a photo. Completely still, locked-off static camera in every shot: no camera movement at all, no pan, no zoom, no push in; shots change only by hard cuts. [0s-1s] The previous shot holds, unchanged. [1s-4.5s] At exactly the 1 second mark, hard cut to a close-up of Basil's face and shoulders turned toward the right: he sneers with theatrical disgust, shakes his head and says 'No, no.', then wags his index finger beside his face and says 'Not good at all.' This one static shot fills the whole rest of the clip, from the 1 second mark to the end; the wide shot does not come back. Audio: a lively market crowd murmur; a deep gruff man's voice with a thick Greek accent says 'No, no. Not good at all.', no music.
Segment 10
Medium, push-in, yelling close-up, then a hard cut to the low-angle point under a sunset. 'Very' plays on the close-up, 'bad' on the point, with the impact frames.
371.6 s GPU · peak 7433 MB
Prompt
Continue the same moment. The man is Basil, a big burly Byzantine man about 45 years old: a huge thick square black beard, heavy black eyebrows, a big hooked nose, a round brown felt cap, a deep-blue knee-length tunic with gold embroidered bands at the collar and cuffs, and a crimson cloak pinned at his right shoulder with a round gold brooch. A modern Japanese TV anime come to life: bold clean black ink outlines, flat cel shading with crisp hard-edged shadows, vivid saturated colours, richly painted anime backgrounds, warm golden late-afternoon light. Not 3D, not CGI, not a photo. The camera pushes in fast toward his face and shakes strongly at the moment he points. [0s-1s] The previous shot holds, unchanged. [1s-3.5s] At exactly the 1 second mark, hard cut to a dramatic low-angle close-up: Basil thrusts his arm straight at the camera, points his index finger right into the lens and growls 'Very bad.', his furious face glaring over his huge hand while his beard and cloak whip in a sudden gust of wind. This one static shot fills the whole rest of the clip, from the 1 second mark to the end; the wide shot does not come back. Audio: a sudden gust of wind; a deep gruff man's voice with a thick Greek accent growls 'Very bad.', no music.
Segment 11
A nervous smile, then the wail with spraying anime tears. The shot starts as the mouth opens; graded grey for the despair beat.
349.4 s GPU · peak 7585 MB
Prompt
Continue the same moment. The old man is Old Nikos, a skinny old arms dealer about 70: a long pointed white beard, a bald head with white wisps above big ears, a long hooked nose, a gap-toothed grin, a mustard-yellow robe and a small dark-green cap. A modern Japanese TV anime come to life: bold clean black ink outlines, flat cel shading with crisp hard-edged shadows, vivid saturated colours, richly painted anime backgrounds, warm golden late-afternoon light. Not 3D, not CGI, not a photo. Completely still, locked-off static camera in every shot: no camera movement at all, no pan, no zoom, no push in; shots change only by hard cuts. [0s-1s] The previous shot holds, unchanged. [1s-3.5s] At exactly the 1 second mark, hard cut to a close-up of Old Nikos: his eyes turn into blank white circles of shock, his mouth falls open into a wailing howl of despair, he clutches his chest with both bony hands and comical tears burst from his eyes as he gasps 'Sixty?!' This one static shot fills the whole rest of the clip, from the 1 second mark to the end; the wide shot does not come back. Audio: a lively market crowd murmur; a thin raspy old man's voice with a Greek accent gasps 'Sixty?!', no music.
Segment 12
The coin hand is animated on threes, three slams exactly 1.0 s apart. Slowed to 0.64-0.82x, two slams land between the lines and on the cut.
370.4 s GPU · peak 7433 MB
Prompt
Continue the same moment. Only Basil's two big hands and the sleeves of his deep-blue tunic with gold embroidered cuffs are in this shot. The sword is the golden sword: a long straight polished steel blade, a golden cross-guard with curled ends, a red leather grip and a golden pommel with one big ruby. A modern Japanese TV anime come to life: bold clean black ink outlines, flat cel shading with crisp hard-edged shadows, vivid saturated colours, richly painted anime backgrounds, warm golden late-afternoon light. Not 3D, not CGI, not a photo. Completely still, locked-off static camera in every shot: no camera movement at all, no pan, no zoom, no push in; shots change only by hard cuts. [0s-1s] The previous shot holds, unchanged. [1s-4s] At exactly the 1 second mark, hard cut to a close-up from above of the red velvet counter: Basil's big hand slaps gold coins down one at a time, slowly and reluctantly - one, then another, then a third - onto a small heap of gold coins, while his other hand drags the golden sword toward himself. This one static shot fills the whole rest of the clip, from the 1 second mark to the end; the wide shot does not come back. Audio: three heavy gold coins slapped onto a wooden counter one at a time, a reluctant grumbling sigh, no music.
Segment 13
Wide street, Basil walks off toward the dome with the sword under his arm while Nikos sobs in the foreground. The cube turn starts from this shot.
309.4 s GPU · peak 7521 MB
Prompt
Continue the same moment. Basil walks away seen from behind; Old Nikos weeps in the foreground. The man is Basil, a big burly Byzantine man about 45 years old: a huge thick square black beard, heavy black eyebrows, a big hooked nose, a round brown felt cap, a deep-blue knee-length tunic with gold embroidered bands at the collar and cuffs, and a crimson cloak pinned at his right shoulder with a round gold brooch. The old man is Old Nikos, a skinny old arms dealer about 70: a long pointed white beard, a bald head with white wisps above big ears, a long hooked nose, a gap-toothed grin, a mustard-yellow robe and a small dark-green cap. A covered market street in Constantinople in the year 1043 at golden hour: pale stone arcades with round arches, colourful striped silk awnings, hanging brass lamps, stalls of spices and copper pots, and the golden dome of the Hagia Sophia at the far end of the street. No modern objects. A modern Japanese TV anime come to life: bold clean black ink outlines, flat cel shading with crisp hard-edged shadows, vivid saturated colours, richly painted anime backgrounds, warm golden late-afternoon light. Not 3D, not CGI, not a photo. Completely still, locked-off static camera in every shot: no camera movement at all, no pan, no zoom, no push in; shots change only by hard cuts. [0s-1s] The previous shot holds, unchanged. [1s-4s] At exactly the 1 second mark, hard cut to a wide shot down the market street: Basil walks away from the camera along the middle of the street toward the sunset, seen from behind, the golden sword tucked under his arm and his red cloak swaying with each step, while in the lower right foreground Old Nikos slumps over his counter, weeping. This one static shot fills the whole rest of the clip, from the 1 second mark to the end, with no other shot in between. Audio: a lively market crowd murmur, footsteps, an old man sobbing, no music.
Sound (ElevenLabs only)
Everything you hear came from ElevenLabs: two designed voices (Voice Design v3), three re-voiced or extra lines (v3 TTS), the haggle groove and the triumphant drop (Eleven Music), and 20 sound effects, 18 of which made the mix. The groove's beat grid (104.94 BPM) drives every cut. Whisper (medium) heard every line correctly in the final mix after two lines were re-voiced.
Cast and item
Two new characters and one prop, designed for this film. Each sheet is one Qwen-Image text-to-image job; single panels were cropped out as references so Qwen would not copy one pose into every shot.
Plates
Location plates for the market street, the stall, the reverse angle and the home. Qwen put people into two 'empty' plates and an Ottoman minaret on the 1043 Hagia Sophia; one removal edit each fixed them.
Previs
A grey-box Blender 5.2 previs of the whole film: the stall at true scale, stand-ins for Basil, Nikos and the sword, one camera per shot keyed to the beat grid of the music, cut on the real scratch mix. It fixed the shot list and timing before any video was rendered; its guide frames set the framing for the H3 end keyframes.
H3 end keyframes
One Qwen end frame per H3 continuation (the shot each segment must land on), all at 1344x768 from the character crops plus a location plate. Nikos's first 'crushed' frame read as laughing, so it was redone with the mouth corners pulled down.
The home still
The still the whole joke lands on. Seed 215 won: both arms up, the whole sword held flat over his head. For the 2.5D move it was split into a clean room plate and a Basil+sword layer.