Aurora worfklow

A developer released "Aurora," a three-shot alpine video generated in a single unbroken H3 ref2va generation on an RTX 3090, using Z-Image to draw three storyboard frames that anchor cuts at 00:02.700 and 00:05.400. The 192-frame, 544-pixel-high clip ran in 10.1 minutes on 24GB of VRAM with nine inference steps, a seed of 42, and the lightx2v MiniMax-h3-Turbo LoRA at scale 1. The continuous generation keeps wind, breath, aurora hiss, and a solo-piano score running across the seams.

| | { | | | "cost": | | | { | | | "device": "cuda", | | | "minutes": 10.1, | | | "name": "RTX 3090", | | | "vram gb": 24 | | | } | | | , | | | "description": "\"Aurora\" \u2014 an original three-shot alpine vignette on one unbroken score: a lone hiker stands beneath the first veil of green aurora at deep twilight, is dwarfed by the full dancing sky at the height of the night, and watches it fade into pink-and-gold dawn. The cuts happen inside one H3 ref2va generation Z-Image draws the three boards; <Picture 1 is the first frame, <Picture 2 and <Picture 3 anchor the cuts at 00:02.700 and 00:05.400 , so wind, breath, aurora hiss and the solo-piano score run continuously across the seams. First run 2026-09-09, job 3326d6699bf5.", | | | "id": "Aurora", | | | "seed": 42, | | | "variables": { | | | "board 1 launch": "wide cinematic photograph at dusk on a snowy alpine ridge: a lone hiker in a dark hooded coat walking up the last gentle rise of a snow crest, seen from behind and slightly below as a small dark figure against a deep blue and amber western sky, long blue shadows raking the snow, a plume of breath at the hood, the first sparse stars beginning to show overhead, photographic, sharp focus on the figure, huge sky above", | | | "board 2 gutter": "wide cinematic photograph at deep dusk on a snowy alpine ridge: a small dark hooded figure standing perfectly still on a crest of ice, seen from a distance against a cold lavender-blue sky with the first scattered stars, a thin spindrift of snow streaming off the ridge line, deep indigo shadow filling the valleys below, hushed and patient, photographic, sharp focus on the figure, immense empty sky above", | | | "board 3 shore": "wide cinematic photograph at deep twilight on a snowy alpine ridge: the small dark hooded silhouette of a lone hiker standing on a crest of ice beneath an immense clear starry sky, a faint first veil of green aurora rippling low along the horizon, deep indigo and blue tones, fine snow crystals glittering in the still air, photographic, sharp focus on the figure, huge empty sky above", | | | "height": 544, | | | "image reference type": "diffusers.modular pipelines.minimax h3.MiniMaxH3ImageReference", | | | "lora adapter name": "turbo", | | | "lora model name": "lightx2v/Minimax-h3-Turbo", | | | "lora scale": 1, | | | "lora weight name": "minimax h3 fl2v turbo 8step v1.0 bf16.safetensors", | | | "num frames": 192, | | | "num inference steps": 9, | | | "prompt": "subject definitions:\n<Subject 1 is the lone hiker in <Picture 1 , <Picture 2 and <Picture 3 : a small solitary figure in a dark hooded coat, seen always as a silhouette against the sky, standing on snow, face never visible.\n<Picture 1 is the first frame of Shot 1 : the lone hiker walking up the last gentle rise of a snowy alpine crest at dusk, seen from behind and slightly below, deep blue and amber sky above, long blue shadows on the snow.\n<Picture 2 is a storyboard reference for Shot 2 , defining its wide distant view of the figure standing perfectly still on the crest under a cold lavender-blue sky with the first scattered stars, spindrift streaming off the ridge line.\n<Picture 3 is a storyboard reference for Shot 3 , defining its wide deep-twilight view: the hiker beneath an immense clear starry sky with the first faint veil of green aurora low on the horizon.\n\nsummary:\n keyframe completion + reference generation The target video follows <Subject 1 through an eight-second dusk in three shots: arriving up the last rise exactly as <Picture 1 shows it, standing still into the deepening cold staged as <Picture 2 plans it, and catching the first flicker of aurora staged as <Picture 3 plans it, under one continuous, sparing score.\n\nretention analysis:\n<Subject 1 appears in Shot 1 , Shot 2 , Shot 3 : fully preserved - the small dark hooded silhouette on the snow is kept exactly in every shot, only its posture and the light around it changing.\n<Picture 1 Shot 1 first frame : fully preserved - the shot opens on its framing, the walking figure seen from behind, the dusk sky and the long blue shadows.\n<Picture 2 Shot 2 storyboard : partially preserved - its wide distant view, the still figure, the cold lavender sky and the spindrift are kept, with the light newly deepening around the waiting figure.\n<Picture 3 Shot 3 storyboard : partially preserved - its wide deep-twilight view, the immense starry sky, the faint veil low on the horizon and the hiker's place on the snow are kept, with the veil newly brightening and the figure newly still.\n\ndetailed description:\nThe target video is in a realistic cinematic style, shot wide with the sky as the subject and the human figure kept small, moving from amber dusk blue through cold indigo to the first green of the aurora at deep twilight.\n Shot 1 A wide shot opens exactly on <Picture 1 : a lone hiker in a dark hooded coat walks up the last gentle rise of a snowy alpine crest at dusk, seen from behind and slightly below, long blue shadows raking the snow, a plume of breath at the hood, the western sky amber fading into deep blue. With slow, even steps the figure crests the rise and stops, boots planting, the breath plume rising and dispersing, a pack strap settling against the coat. The camera holds a slow tracking shot that settles to stillness as the figure stops, the sky darkening almost imperceptibly around them.\n Shot 2 At 00:02.700, the camera cuts to the wide distant view of <Picture 2 : the hiker stands perfectly still on the crest, small against a cold lavender-blue sky where the first scattered stars are appearing, a thin spindrift of snow streaming off the ridge line, the valleys below filling with indigo shadow. The figure does not move; only the breath plume, the drifting snow and the deepening of the light mark the passing of time. The camera holds a near-static wide shot through the long middle of the piece, patient and unhurried, as dusk settles into cold.\n Shot 3 At 00:05.400, the camera cuts to the wide deep-twilight view of <Picture 3 : the hiker stands beneath an immense clear starry sky, the horizon a deep indigo line - and low along the northern horizon a faint first veil of green aurora appears, rippling gently upward. The figure's head lifts slightly as they notice it; the veil brightens and spreads just a little, holding its low place above the ridge. The camera holds the wide shot nearly still, drifting in very slightly toward the figure and the awakening sky, as the first light of the night begins.\n\noverall soundscape:\n Shot 1 opens on the crunch of boots on crusted snow and the tap of a pole fading as the figure crests the rise and stops, a strap settling, breath arriving as a steady mist, a low wind over the crest. Shot 2 is the long hush of the wait: wind breathing over the ridge, spindrift hissing thinly across the ice, the faint tick of cooling fabric, near silence. Shot 3 the wind falls away entirely and a faint high keening rises from the horizon with the first green veil, the hiker's slow exhale the only human sound.\n\nnon diegetic music:\nA sparse prelude for solo piano and soft sustained strings, quiet and unresolved, carries continuously across all three shots without pause: slow single notes over deep pedal tones that wait as much as they play, growing no louder as the sky darkens, and ending on one open, unresolved held chord as the first veil of green rises - a chord left deliberately unfinished, as if inviting what comes next.", | | | "reference image short edge": 1024, | | | "weights dtype": "{int4}", | | | "width": 960 | | | }, | | | "steps": | | | { | | | "name": "board 1 launch", | | | "pipeline": { | | | "arguments": { | | | "guidance scale": 0, | | | "height": 544, | | | "num inference steps": 9, | | | "prompt": "variable:board 1 launch", | | | "width": 960 | | | }, | | | "configuration": { | | | "component type": "ZImagePipeline", | | | "offload": "sequential" | | | }, | | | "from pretrained arguments": { | | | "low cpu mem usage": true, | | | "model name": "Tongyi-MAI/Z-Image-Turbo", | | | "torch dtype": "torch.bfloat16" | | | } | | | }, | | | "result": { | | | "content type": "image/jpeg", | | | "embed metadata": true | | | } | | | }, | | | { | | | "name": "board 2 gutter", | | | "pipeline": { | | | "arguments": { | | | "guidance scale": 0, | | | "height": 544, | | | "num inference steps": 9, | | | "prompt": "variable:board 2 gutter", | | | "width": 960 | | | }, | | | "configuration": { | | | "component type": "ZImagePipeline", | | | "offload": "sequential" | | | }, | | | "from pretrained arguments": { | | | "low cpu mem usage": true, | | | "model name": "Tongyi-MAI/Z-Image-Turbo", | | | "torch dtype": "torch.bfloat16" | | | } | | | }, | | | "result": { | | | "content type": "image/jpeg", | | | "embed metadata": true | | | } | | | }, | | | { | | | "name": "board 3 shore", | | | "pipeline": { | | | "arguments": { | | | "guidance scale": 0, | | | "height": 544, | | | "num inference steps": 9, | | | "prompt": "variable:board 3 shore", | | | "width": 960 | | | }, | | | "configuration": { | | | "component type": "ZImagePipeline", | | | "offload": "sequential" | | | }, | | | "from pretrained arguments": { | | | "low cpu mem usage": true, | | | "model name": "Tongyi-MAI/Z-Image-Turbo", | | | "torch dtype": "torch.bfloat16" | | | } | | | }, | | | "release pipeline": true, | | | "result": { | | | "content type": "image/jpeg", | | | "embed metadata": true | | | } | | | }, | | | { | | | "name": "voyage", | | | "pipeline": { | | | "arguments": { | | | "height": "variable:height", | | | "num frames": "variable:num frames", | | | "num inference steps": "variable:num inference steps", | | | "output": | | | "videos", | | | "audio", | | | "sampling rate" | | | , | | | "prompt": "variable:prompt", | | | "references": | | | { | | | "from previous result": "board 1 launch", | | | "reference type": "variable:image reference type" | | | }, | | | { | | | "from previous result": "board 2 gutter", | | | "reference type": "variable:image reference type" | | | }, | | | { | | | "from previous result": "board 3 shore", | | | "reference type": "variable:image reference type" | | | } | | | , | | | "width": "variable:width" | | | }, | | | "configuration": { | | | "component type": "ModularPipeline", | | | "components": { | | | "audio vae": { | | | "device": "cuda", | | | "residency": "on demand" | | | }, | | | "text encoder": { | | | "remove modules": | | | "lm head" | | | | | | }, | | | "text encoder.model": { | | | "group offload": { | | | "offload type": "leaf level" | | | }, | | | "truncate layers": { | | | "language model.layers": 51 | | | } | | | }, | | | "transformer ref": { | | | "group offload": { | | | "low cpu mem usage": true, | | | "num blocks per group": 1, | | | "offload type": "block level", | | | "record stream": true, | | | "use stream": true | | | } | | | }, | | | "vae": { | | | "device": "cuda", | | | "residency": "on demand" | | | } | | | }, | | | "configs": { | | | "reference image short edge": "variable:reference image short edge" | | | }, | | | "load components": { | | | "dtype": "torch.bfloat16", | | | "quantization config": { | | | "text encoder": { | | | "arguments": { | | | "dequantize fp32": false, | | | "modules to not convert": | | | ".model.visual", | | | "lm head" | | | , | | | "quantization device": "cuda", | | | "return device": "cpu", | | | "weights dtype": "variable:weights dtype" | | | }, | | | "configuration": { | | | "config type": "sdnq.SDNQConfig" | | | } | | | }, | | | "transformer ref": { | | | "arguments": { | | | "dequantize fp32": false, | | | "modules to not convert": | | | "proj in", | | | "audio proj in", | | | "context embedder", | | | "time embedder", | | | "time proj", | | | "token refiner", | | | "norm out", | | | "proj out", | | | "audio proj out" | | | , | | | "quantization device": "cuda", | | | "return device": "cpu", | | | "use quantized matmul": true, | | | "weights dtype": "variable:weights dtype" | | | }, | | | "configuration": { | | | "config type": "sdnq.SDNQConfig" | | | } | | | }, | | | "vae": { | | | "arguments": { | | | "dequantize fp32": false, | | | "quant conv": true, | | | "quantization device": "cuda", | | | "return device": "cpu", | | | "use quantized matmul conv": true, | | | "weights dtype": "{int8}" | | | }, | | | "configuration": { | | | "config type": "sdnq.SDNQConfig" | | | } | | | } | | | } | | | }, | | | "pre load modules": | | | "sdnq" | | | | | | }, | | | "from pretrained arguments": { | | | "model name": "MiniMaxAI/MiniMax-H3", | | | "workflow": "ref2va" | | | }, | | | "loras": | | | { | | | "adapter name": "variable:lora adapter name", | | | "model name": "variable:lora model name", | | | "scale": "variable:lora scale", | | | "weight name": "variable:lora weight name" | | | } | | | | | | }, | | | "result": { | | | "content type": "video/mp4", | | | "fps": 24 | | | } | | | } | | | | | | } |