1,400 prompts, one model, and a LOCK block for every failure
Every prompt is 8 to 18 thousand characters long, and each model failure became a named LOCK block that was never removed. Multi-shot editing is written inside the text, and the dialogue is a Chukchi-style constructed language in Cyrillic.
Source: the film's public page on Higgsfield. 20m 02s, 4K, published 2026-09-24, produced from 17 Aug to 21 Sep 2026 (5 weeks).
An Arctic drama about love, sacrifice and the cost of letting go. Exiled with her father, whose fading mind mistakes her for his dead wife, a young huntress leads him across the frozen wilderness of 1920s Chukotka to the sacred Whale Bone Alley.
Coverage: complete. 3,353 submissions became 1,400 unique prompts (1,015 video, 385 image). Every prompt was read end to end. No generation was downloaded: this analysis uses only text and metadata. Higgsfield counts 11,961 "generations" because each submission produces 2 to 4 outputs.
A note on quotes. About a third of the quoted lines in this post are condensed from longer passages or joined from nearby text. Read them as close paraphrases of what the prompts say, not always as verbatim strings.
1. In one line
One model (Seedance 2.5), one format (21:9, 1080p, audio on, multi_shots off), and prompts of 8 to 18 thousand characters where every model failure becomes a "LOCK" block that is never taken out again. The multi-shot editing is written inside the text (48% declare HARD CUTs) and not in the tool's parameter. Re-rolls are rare: the prompt is rewritten on each submission (median ×1 in the early scenes, ×3-5 in the last ones).
| Video | 1,015 prompts · 100% seedance_2_5 · 21:9 · 997 at 1080p (18 at 720p, tests for road) |
| Requested duration | 30s (292) › 20s (112) › 10s (86) › 15s (84) › 8s (68) · any value from 4 to 30 |
| Video prompt length | median 7,972 characters · p90 17,639 · maximum 26,922 |
| Image (ASSETS) | 385 · soul_v2 167 · soul_cinematic 115 · seedream_v5_pro 68 · gpt_image_2 14 · nano_banana_2 13 |
| Attachments per video submission | median 5 images (maximum 10) · 406 submissions with an input video · 105 with audio |
| Language | 100% English in the text; dialogue in a Chukchi-style constructed language written in Cyrillic; the notes and assistant preambles left in the prompts are in Russian |
2. The story according to the prompts
The folders are not scenes. They group work sessions: the "9th scene" ends in the Whale Bone Alley, "10th" mixes flashbacks, and the "2nd scene" contains shots labeled SC.3. The real order comes from the internal codes in the prompts (STYLE PREFIX — SC.4Б, SCENE AN-07 / AFTER THE WATER) and from the content.
| # | Stretch | Folder(s) | What happens |
|---|---|---|---|
| 1 | The trap | 1st scene | Polar night. Tyne (18) kills a trapped fox. Her breath is "the only breath left". She sees the glow of the burning camp |
| 2 | The fire | 2nd, 3rd scene | Omryn, lost in dementia, burns the camp laughing. He shouts "Гыронав!" (his dead wife) in front of an oil lamp |
| 3 | The judgment | 4th scene (7 subfolders) | Rintyn demands his death ("there is kele in him"). Tyne steps between him and the spear. The masked shaman arrives with the drum. Sentence: exile, and Omryn's dogs go to Rintyn as payment. Rintyn throws Tyne a single rope (one dog). The shaman tells her about the anerneq and sends her to the Alley |
| 4 | The departure | 5th scene | She packs in the burned yaranga and wakes him by squeezing his hand. She pulls the sled herself next to the dog |
| 5 | The journey | road | Survival montage: bone skate, snowshoes, a burrow in the snow, an ermine the dog brings |
| 6 | The break | 6th scene | He calls her Gyronav, stops the sled and accuses her of lying. She brings him back by singing in her mother's voice and ties his wrists |
| 7 | The ice | 7th, 8th scene | He walks onto the ice as if "called" and falls in. She pulls him out. The dog Tumgy dies: he bites the rope and the pull drags him into the water. In the prompt: "one saved by the other's loss" |
| 8 | The cave | 9th, 10th scene | She dismantles the sled to make a fire, undresses him and greases him. She whispers "Tumgy…". In the morning she puts on the dog's harness herself |
| 9 | The Alley | final-v1, final-v2, root | She arrives with no strength left. She makes fire with a bow and old whale fat. She cuts her palm and puts the knife to her own throat as a sacrifice. He rescues her. In a lucid moment he asks her to make the sacrifice him. He calls her Гыронав again; she sings him her mother's lullaby and kills him off-screen. His "Гыро—" is cut off and "HER BREATH STILL FOGS. His does not". Vigil, aurora ("the only colour in this film"), a fox. She leaves alone |
Anerneq means breath or soul in Inuit languages. This is an inference by the readers; no prompt says it. Breath is the visual motif of the whole film: it opens with her breath over the dead fox and closes with one breath where there were two.
3. The template: five eras
The prompts do not share a single skeleton. There are overlapping eras. Weekly evolution is measured below.
A · Long "Revenant" (17 Aug to about 28 Aug, 13-27k characters). It inherits the format of the other films from the same studio:
REFERENCE USAGE (Seedance 2.5): <<<image_1>>> — rol … <<<uuid>>> — rol …
TAGS / CAST & ASSETS · WHO IS WHO (KEY — never confuse or blend)
<locks del plano: HANDS LOCK, BACKLIGHT LOCK, FACE QUARANTINE, SPEAR COUNT…>
STYLE PREFIX — SC.4Б … v3 / 2 SHOTS, 1 HARD CUT, ~15s: (1 ~6s) … → (2 ~8s) … — ANAMORPHIC 21:9 "REVENANT", ANTI-AI-TELL
Style: a real photographed frame of 65mm anamorphic FILM … The Revenant (Iñárritu / Lubezki) … NOBODY can tell this is AI.
COMPOSITION (KEY) · THE SHOT — N SHOTS · SPOKEN LINES
CAMERA — RAW BREATHING HANDHELD (KEY) · LIVING FACE/EYES (KEY) · ANTI-AI-TELL DIRECTIVE (KEY)
LIGHT — NATURAL LIGHT ONLY (KEY) · NO LENS FLARES · HARD ANAMORPHIC "REVENANT" LENS (KEY)
HYPER-REAL CAPTURE (Vision3 500T pushed) · PHYSICS · CHARACTER ACTING (Tigran method) · AUDIO
[TOTAL RUNTIME …] [GEOMETRY ANCHOR … LOCKS absolute: S1…; S2…]
NEGATIVE: <150-300 ítems>
(In the template, rol means role, locks del plano means the shot's locks, and ítems means items. These words are in Spanish in the analyst's notation, not in the prompts.)
B · Compact "Kodak Vision3 500T" (from about 24 Aug, 3-9k). The one used for most of the film (6th, 8th, 9th, final-v1):
LIGHT — … / FILM LOOK — "THE REVENANT", Lubezki, Vision3 500T … Dark but READABLE
WORLD · CAST (<<<uuid>>> de Tyne + @OMRYN @TUMGY en texto)
<LOCKS por problema, el que está fallando arriba de todo con "READ FIRST" / "(KEY)" / "THIS IS THE CHANGE">
THE SHOT o MULTISHOT — 1 (6s) … HARD CUT. 2 (4s) … HARD CUT …
VOICES (conlang cirílico + glosa "never spoken") · AUDIO — DIEGETIC ONLY · CAMERA — REVENANT HANDHELD
KEEP THE LOGIC: … ORDER IS FIXED: A → B → C
FILM LOOK REPRISE · NEGATIVE: <80-150 ítems>
(Notation: "de Tyne" is "for Tyne", "en texto" is "in text", "LOCKS por problema, el que está fallando arriba de todo" is "one LOCK per problem, the one that is failing at the very top", "o" is "or", "glosa" is "gloss", "cirílico" is "Cyrillic".)
C · Numbered "ABSOLUTE RULES" / "LAWS" (from 5 Sep). 6-10k characters. Named rules that close with "WRONG… regenerate". In the root folder two new locks appear: "DESCRIBE EVERY MILLIMETRE OF SKIN" and "METAL DOES NOT EXIST".
D · Micro-prompts and "TASK —" (Sep). 7-8s tests to isolate a single problem, surgical V2V edits ("ALREADY CORRECT … CHANGE ONLY THE THREE THINGS LISTED") and conversion to previz.
E · "STRICT CONSISTENCY" with ═══ (14 to 21 Sep, 13-24k). The one in final-v2 and the last prompt of the project:
STRICT CONSISTENCY — ONE LOCATION, ONE LIGHT, ONE FILM STOCK. FIVE SHOTS, HARD CUTS, ~20 SECONDS.
═══ THE SKY — EXACT COLOUR, HIGHEST PRIORITY, DO NOT DRIFT ═══ ← RGB (78,112,150) / #4E7096 + TEMPORAL LOCK
═══ THE LIGHT ═══ ← "EXACTLY ONE LIGHT SOURCE EXISTS IN THIS WORLD AND IT IS THE SKY"
═══ THE LOCATION ═══ ← descrita por FORMA, con SCALE ANCHOR
FILM LOOK · <objetos que deben igualar su referencia> · <personaje con <<<uuid>>> y estado físico>
WIND · CAMERA · FIVE SHOTS, HARD CUTS, ~20s: 1 (4s) … 5 (3s) …
AUDIO — DIEGETIC ONLY · KEEP THE LOGIC (+ contadores: "Total shots: 5. Warm light sources: 0.") · NEGATIVE
(Notation: "descrita por FORMA, con SCALE ANCHOR" is "described by SHAPE, with a SCALE ANCHOR"; "objetos que deben igualar su referencia" is "objects that must match their reference"; "personaje con uuid y estado físico" is "character with uuid and physical state"; "contadores" is "counters".)
| week | prompts | median chars | ═══ | KEEP THE LOGIC | "SHOT IS WRONG" | HARD CUT | <<<audio>>> |
|---|---|---|---|---|---|---|---|
| 17 Aug | 137 | 13,214 | 0% | 24% | 0% | 28% | 0% |
| 24 Aug | 212 | 7,372 | 0% | 54% | 0% | 61% | 5% |
| 31 Aug | 277 | 7,104 | 1% | 56% | 0% | 40% | 17% |
| 7 Sep | 239 | 8,156 | 0% | 49% | 2% | 61% | 5% |
| 14 Sep | 131 | 13,425 | 24% | 76% | 22% | 43% | 0% |
| 21 Sep | 19 | 9,781 | 68% | 95% | 79% | 37% | 5% |
Constants across the five eras: 65mm anamorphic + Kodak Vision3 500T (99%), handheld (98%), "NO MUSIC" (91%) and diegetic-only audio, dialogue in quotes in Cyrillic (89%) and a NEGATIVE at the end (96%).
The prompts read like LLM output pasted in whole. Left inside are assistant preambles in Russian ("Тот же промт, только небо затемнил…", roughly "The same prompt, I only darkened the sky…"), unreplaced placeholders ("[insert LOOK LOCK block here]", "put your tag here"), stray Chinese characters, prompts pasted twice (14k characters, submitted anyway) and a final block wrapped in a code fence. One prompt in final-v2 is a rewrite in natural prose that looks like the output of a prompt optimizer.
4. References: where identity lives
There are two channels. The first scrape saw only one of them.
- Named elements (
reference_elements).tyne(606 prompts),char_anerneq_omyryn,char_anerneq_rintyn, the shaman in 4 versions,loc_anerneq_burning_camp_ext, props (props_buben_v2,props_bone_knife,props_sled). In the text they appear as<<<uuid>>>. If the uuid is deleted from the text, the element stops being attached, which is why there are gaps like "CAST: — 18" or "no , no @OMRYN". @OMRYN, @TUMGY and @NARTY are almost always plain text with no image. - Uploaded attachments (
params.medias): 11,482 images, 406 videos and 105 audio files across 2,714 video submissions. They are keyframes, location plates, previz stills, Tyne's voice (<<<audio_1>>>), the mother's lullaby and the shaman's drum as sound continuity.
Every reference carries an explicit role and a list of what NOT to take. This is the most consistent technique in the project:
- "The keyframe is a COMPOSITION, LIGHT and PALETTE map ONLY … textures come from the assets at FULL quality, never smoothed to match the keyframe."
- "BODY REFERENCE ONLY — DO NOT USE HIS FACE" · "OBJECT ONLY, ignore the gallery room" · "PALETTE ONLY — NOT COMPOSITION" · "The reference image defines the light ONLY."
- "REFERENCE SCOPE (KEY): the references supply faces, wardrobe, the dog and the sled ONLY. IGNORE their backdrops, ground, horizon, sky, weather, lighting direction, exposure and colour grade."
- "CAVEATS TO THE KEYFRAME (they OVERRIDE anything read differently in the still)."
- "the keyframe is a STARTING POSITION, not a freeze" · "THE FIRST FRAME IS ALREADY IN MOTION — NO held tableau", because the clip started frozen on the reference.
- Chaining: "KEYFRAME (the last frame of the previous clip): SHOT 1 STARTS EXACTLY ON THIS COMPOSITION."
- Continuity audio: "the first beat lands exactly where the reference's next beat would fall; the voice is the SAME throat continuing mid-breath."
- Voice: "TYNE'S VOICE … use THIS voice for EVERY line she speaks, in the close-ups AND on the wide … only the space around it changes."
Reverse previz pipeline (4th scene). A video that already came out well is converted into colored mannequins (Rintyn yellow, Omryn blue, Tyne orange) and then repainted with the real assets, using a COLOR LEGEND (mapping only) … COPY NOTHING VISUAL from the previz; the previz gives ONLY motion, positions and timing. This separates motion from identity. They also use Blender stills with the same color code and V2V with "DO NOT BREAK THE SOURCE VIDEO — SUPREME RULE".
5. Techniques by axis (worth copying)
Light and color: the whole project is a fight against orange. Seedance puts warm fire into night scenes and lifts the sky. The fix arrived in layers:
- Declared exposure: "EXPOSED FOR THE DEEP BLUE DUSK, NOT for the fire … If her whole face reads orange — WRONG."
- A single source: "EXACTLY ONE LIGHT SOURCE EXISTS IN THIS WORLD AND IT IS THE SKY. NO key, NO rim, NO backlight."
- Color allowed without its light: "frostbite … is PIGMENT IN THE FLESH, NOT LIGHT FALLING ON HER". This lets warm skin, blood or ivory in without opening the door to a warm light source.
- Numeric sky with a temporal lock. The value drifted between versions: RGB (62,88,115) → #6F8A9E → #5E7D95 → (45,62,88) → #4E7096 / RGB (78,112,150) "MID-TONE", the final one. It comes with "IDENTICAL IN THE FIRST FRAME AND THE LAST FRAME" and "KEEP THE HORIZON OUT OF FRAME".
- Fixed value scale: "snow → face → sky → fur → bone"; "THE BLUE CHANNEL IS THE STRONGEST AND THE RED CHANNEL IS THE WEAKEST IN EVERY PART OF THE IMAGE".
- Tungsten look: "3200K FILM STOCK IN DAYLIGHT, WITHOUT AN 85 CORRECTION FILTER — only the white balance is wrong."
- Fire off-screen, only as an effect: "WE DO NOT SEE THE FIRE. What we see is THE LIGHT ON HER CHANGE." Distant light in physical order: snow → emerging shadow → skin → face, half a stop, irregular pulse.
- Light per cheek: "THE FIRE IS BEHIND THE CAMERA: the lit side of every face is the side TURNED TOWARD THE LENS."
- Anti-flash: "snow 1 metre away and 300 metres away are EQUALLY dim … a bright foreground fading into dark distance is a flash-photo signature."
Shape: describe by shape, not by name. The model drew "dog bones" in the whale graveyard. In a final-v2 prompt: "DESCRIBE THEM BY SHAPE, NOT BY NAME": the ribs become "a SINGLE LONG SMOOTH CONTINUOUS SHAFT … like a giant scythe blade … no knob, no ball", with a SCALE ANCHOR of 3-5 m and the rule "IF ANY SHAPE HAS A ROUNDED KNOB AT EACH END, THE SHOT IS WRONG". Side effect: they later had to negate "a scythe prop". Other cases: "IT IS ONE PRONG, NOT TWO … NOT paddle-shaped" and "HOW TO DRAW A JAWBONE SO IT IS NOT A TAIL".
Skin and texture.
- Back and forth between "slight softness" and "tack-SHARP, every pore". The second produced crust and mottling, and they closed it with "GRAIN LOCK — THE GRAIN LIVES IN THE FILM, NEVER ON HER FACE", "SURFACE LOCK: cold reads as COLOUR, never as pattern" and "grain laid OVER a sharp image — never replacing it".
- Hands: "EACH FINGER IS ONE UNBROKEN LENGTH OF SKIN FROM KNUCKLE TO FINGERTIP … NAILS NORMAL". Against rings that kept reappearing: "NOTHING ON HER HANDS CATCHES OR REFLECTS LIGHT". Filling the whole surface with description leaves the model no room to invent accessories.
Camera: handheld with a body behind it.
- "a smooth or parked frame is WRONG" · "a perfectly smooth stabilized drift is an AI-tell" · "realism is SUBTRACTIVE".
- Moves are described as steps: "the camera WALKS IN on real feet — trembling steps, footstep bounce, slight overshoot, rough settle"; "widens by walking, not by zooming"; "a human lean, not a dolly move".
- Movement described by what is seen: "the face stays locked … it is the BACKGROUND that moves, like a car window". "TWO MOVEMENTS AND THEY NEVER HAPPEN AT THE SAME TIME": camera and subject take turns.
- Optics as characterization: the shaman on a telephoto for the compression; Rintyn at 24-28mm "at arm's length", described by its signatures ("hood LOOMS … WORLD FALLS AWAY … barrel curve").
- Geometry in screen coordinates: "HIS OWN LEFT = the viewer's SCREEN-RIGHT"; DIRECTION MAP "pixel truth" (where the irises and the whites of the eyes sit); PROOF OF HEIGHT (which part is closest to the lens, to force the high angle); COMPASS LOCK and SUN AXIS with the side of the sun in each shot; SCALE LOCK "1/12 OF THE FRAME HEIGHT".
- Cuts: "CUT ON MOVEMENT — EVERY CUT IS A MATCH-CUT … never repeat the action across the cut"; "SHE NEVER MOVES INTO THE NEXT POSE ON SCREEN — the cut does it"; "FACING — DESCRIBED AS TWO STATES, NOT AS A SPIN".
Acting: getting the "AI reaction" face out of Seedance.
- Method: "CHARACTER ACTING (Tigran method)" with super-objective, tactic, obstacle and involuntary; later, Stanislavski.
- "ACTING LAW": "The face is ASYMMETRIC and LATE … the BODY moves a half-beat BEFORE the face understands."
- "MOTION BUDGET: exactly THREE facial events … the anger lives in the SET of the face"; "GODFATHER-QUIET … ONE microscopic NOD"; "RINTYN'S FACE DOES NOT OVERACT. NO snorting, NO flared nostrils".
- Dementia without a mad face: "HOW HIS EYES CHANGE — SHOW IT, DO NOT ACT IT … THE EYES STOP TRACKING HER."
- Tears: "They WELL AND SIT in the lower lid … fill and do not spill"; "IF TEARS RUN DOWN HER FACE … IT IS WRONG". The tears swung back and forth six times between versions.
- Subtext in one line: "she is not deciding anything, she already knows"; "THE CONFLICT IS NOT WHETHER — IT IS HER OWN BODY REFUSING TO OBEY A DECISION SHE HAS ALREADY MADE".
- Physical: "Every movement is PARTIAL, UGLY, LATE AND PAID FOR"; "EVERY GRAB FAILS ONCE BEFORE IT HOLDS"; "A CORPSE, PLAYED WITH COMPLETE DEADWEIGHT".
- Closed verb vocabulary: "Verbs allowed: HAUL, DRAG, GRIND, STRAIN, TRUDGE. FORBIDDEN: run, jog, stride, glide"; for undressing him, "LOOSEN, EASE, WORK FREE, SLIDE".
What is not shown.
- "THE EVENT MUST READ, THE ACT MUST NOT BE SHOWN." · "The frame edge is the moral line of the film, and it is never crossed." · "HER OWN BODY IS THE LINE."
- Death in a single physical fact: "HER BREATH STILL FOGS. His does not. Do not use anything else to say it."
- Abstract states as visible phenomena: "WETNESS — DESCRIBED AS VISIBLE PHENOMENA, NOT AS A WORD".
- State of the world before the shot: "THE STATE OF THINGS — ALREADY DONE"; "TIMELINE LOCK — BEFORE/AFTER THE ICE BROKE" (dog alive or dead, sled whole or chopped up).
Audio and language.
- Dialogue in a Chukchi-style constructed language (part of it is real Chukchi) written in Cyrillic, with "(gloss, for meaning only, never spoken or shown)", a phonetic guide and even IPA. There are explicit "NO ENGLISH" and "NO Russian" lines. The parenthetical glosses still leaked at least once: an English "no no", which they corrected in a later prompt of final-v1.
- "AUDIO — RAW LOCATION SOUND, NOT DESIGNED: one production microphone near the fire … NO cinematic whoosh, NO chimes, NO AI-clean tones." "NO ambient bed, NO room tone, NO stinger."
- The lullaby from attached audio: "A REAL UNTRAINED HUMAN VOICE … never swells, never resolves … ENDS BY BREAKING OFF HALFWAY THROUGH A NOTE."
- A voice that is hard to handle in post: "SPOKEN LINES: NONE … steam pulses are its visual placeholder." Voice-over without lip movement: put the line over a shot with no faces.
- DRUM MAP with timecodes and the cuts landing on the beats; "the WHIP of the rope carrying ACROSS THE CUT".
Contractual closings.
KEEP THE LOGIC:recaps the shot and fixes the order ("THE ORDER IS DIG → SET THE KIT → DRILL → EMBER").- From September it adds counters: "Total shots: 2. Hard cuts: 1. Completed knife thrusts: 1. Falling tears: 0. Warm light sources: 0. Dogs: 0."
IF … THE SHOT IS WRONG/WRONG — regenerate: verifiable conditionals, not adjectives.- Priority markers on the block that is failing: "READ THIS FIRST, IT IS THE MOST COMMON FAILURE", "(HIGHEST PRIORITY, DO NOT OMIT)", "READ FIVE TIMES", "repeated because it keeps failing".
6. What they chased (and how they solved it)
The NEGATIVE works as a logbook: each failure is added at the front and almost never deleted. Sometimes a negative denies exactly what the previous prompt asked for, which marks a change of direction and not a correction.
| Recurring failure | Where | Final solution |
|---|---|---|
| Fire orange invading everything | whole project | Take the fire out of frame, expose for blue, "PIGMENT NOT LIGHT", dominant blue channel. In final-v2, "there is no fire" |
| Sky that lightens or changes between cuts | final-v1/v2, 9th | HEX + RGB + TEMPORAL LOCK + horizon out of frame |
| Extras' faces on protagonists and clones in the crowd | 4th scene | Portraits → "COMPLETELY FACELESS" → FACE QUARANTINE → take the character out of frame with a low angle |
| Clip that starts frozen on the keyframe | all | "starting position, not a freeze" / "FIRST FRAME ALREADY IN MOTION" |
| Smooth or crane camera | 2nd, 4th, 5th | "CAMERA DOCTRINE: EVERYTHING IS RAW HANDHELD … NO stable shot exists in this film" |
| Rings on fingers, male hands | 1st, 5th, final | Describe every centimeter of the hand; "NOTHING … CATCHES OR REFLECTS LIGHT" |
| Dog bones, "chew bones" | ASSETS, final-v2 | Describe by shape + SCALE ANCHOR + the "rounded knob" rule |
| The shaman drops the drum | 4th/version_of_shaman | Physical solution: "the DRUM hangs on a SHORT LEATHER STRAP looped around the LEFT WRIST — CANNOT fall" |
| Rintyn moves his lips during the voice-over | 4th/5th shot | The line moves to a shot with no faces |
| Intruding objects (sled, metal) | final-v2 | Stop naming them: "NO MAN-MADE OBJECTS OF ANY KIND" |
| Romantic reading of the father-daughter pair | 5th, 6th | "a daughter's", "NOT sexualised in any way" |
| Swapped characters | 2nd, 4th | Identity tied to wardrobe and prop: "THE MAN HOLDING THE TORCH IS ALWAYS IN THE DARK SEALSKIN" |
Longest arcs: the euthanasia scene (26 prompts in about 6 hours in final-v2; 16 versions in final-v1), the knife disarming (21 prompts), Omryn in the water (21 prompts in 5 hours), undressing him (11 prompts in 2 hours) and the sky of the Alley. The most re-rolled prompt of the project is in final-v2: ×21, a continuous take of about 35s.
7. ASSETS: the 385 images
- 43% were not written by the author. The 167
text2image_soul_v2prompts are the automatic expansion by Soul's enhancer: 30 characters in Russian ("polar bear and 2 cubs") became three 1,200-character captions that invent things ("Azimuthal focus", forests in Chukotka). - There is little character material. Rintyn has 4 character sheets in gpt_image_2 and Tyne a single sheet (18 years old, blue tattoos, her dead mother's parka, deliberately too big). Omryn has no asset in this folder, so his elements were made elsewhere.
- Most of it is locations and props.
- Locations: the tundra plate in about 35 light versions; the burning camp in about 30, with states (before, during, 3 hours after, morning); the burned interior in about 20; the Alley, the bowl and the skull in about 45.
- Props on seamless gray, with numbered parts: SNARE, YARAR (drum), spear, harness, knife and horned mask, with a mirrored rear view.
- Reusable editing techniques:
- Temporal state of a location from its own asset: "keep the SAME place … but hours later, changed by time".
- Reverse shot from a reference.
- A photochemical regrade that touches nothing else.
- Generate clean, then "wound" with numbered defects.
- A plate that does not anticipate the event: the hole in the ice "has to appear in the scene, not be waiting for it in the plate".
- The teal-green grade was abandoned because the snow came out green.
8. Non-obvious findings
multi_shotswas never used (1,015 of 1,015 set tofalse). All multi-shot structure is written in the text, with a duration per shot and HARD CUT. The written duration often exceeds the requested one (42-52s of script in 30s generations).- Naming a uuid attaches the reference; deleting it detaches it. That is why the prompts have grammatical gaps. One case uses the uuid to exclude the character: "NO <<
>>. She is gone." - They rewrite instead of re-rolling in the early scenes (×1.0-1.2), and re-roll more at the end (×3-5, up to ×21) when the template has stabilized.
- Contradictions between scenes were not resolved. An iron trap in scene 1 against "METAL DOES NOT EXIST" in scene 5. "cold blue-black" against "a VIOLET film, NOT a blue one". The wife's name appears with seven spellings (Гыронав, Гуренав, Гуренова…). The readers found no sign of which version made it into the final cut.
- Other models, almost none. There is one
cinematic_studio_video_4_0submission with a<<<movement:uuid>>>token; a minute later they go back to Seedance. - Truncation. One prompt is cut at exactly 13,000 characters and loses LIGHT, AUDIO and NEGATIVE. It is probably a UI limit.