Skip to content

Global-AI-Group-Inc/content-generation-skills / skills/mira-video-prompting

mira-video-prompting

Verified

Which of Mira's video models to pick (the live list is in references/selection.md) and how each one treats a prompt: Seedance 2.5, Runway Gen-4.5, Kling 4.0, Kling 3.0 (and Turbo), Gemini Omni Flash, MiniMax H3, HappyHorse, Wan 3.0. Use it before any generate_video call on Mira, when the user asks "which model", "how long can it be", "does it have sound", "промпт для видео", or pastes a video prompt to improve. NOT for stills (mira-image-prompting) or for the camera and light banks themselves (mira-craft).

references/models/seedance-2-5.md

View on GitHub
<!-- generated from the registry by comfy_study/comfy_video/build_skills.py; edit the registry, not this file -->

# ByteDance Seedance 2.5 (`seedance-2-5`)

| | |
|---|---|
| seconds | 4-30s, any whole second |
| output | 720p; 1080p with quality=hd; 480p draft, finalized to 1080p |
| reference images | 30 |
| audio | yes |
| prompt budget | 6000 chars |
| extend | forward and backward |
| draft | yes: generate_video draft=true renders 480p; finalize_draft re-renders the same take in 1080p within 7 days |
| last frame | yes |
| reference clips | up to 10, 30 s each |
| reference audio | up to 10, bound as @AudioN; needs at least one photo or clip alongside |

## When to pick it

the long-form specialist: up to 30s, heavy referencing (storyboards, keyframe chains, video edit/extend); ~1.6x the price of seedance and pricier than kling-4 or wan3, which also reach 30s - reach for it when length or reference volume decides the clip. ALSO the camera-trick model WITH audio: in live runs (2026-09-06) it executed every phased single-take trick as written - a top-down crash dive, a full 360 orbit returning to the start, a whip pan fusing two locations, a dolly zoom, a camera struck and falling while it records, a locked day-night cycle, a pass through a mirror into a new place - where Runway smeared or split half of them and Kling stopped short of the move; when the MOVE is the concept and the clip needs sound, this is the key

## How it reads a prompt

MODEL: Seedance 2.5 (ByteDance, ModelArk id dreamina-seedance-2-5-260628), 24 fps, 720p, 4-30 seconds, native synchronised audio, native generation in 10+ languages. It is the LONG-FORM model of the line-up - the strongest at a coherent 30-second multi-shot story in a single pass with no stitching (Kling 4.0 and Wan 3.0 also reach 30 seconds) - and ByteDance documents a prompt structure for it that is materially DIFFERENT from Seedance 2.0's, so follow this section and ignore any 2.0 habit. OFFICIAL FOUR-PART STRUCTURE (ByteDance's Seedance 2.5 prompt guide): (1) ONE-SENTENCE SUMMARY first - subject + location + event + genre/style + camera movement, in a single line. (2) DETAILED DESCRIPTION - the subject's concrete appearance, the scene and its props and background, then the camera: shot size, angle, and how steady or handheld the framing is. (3) TIMELINE SEGMENTS - and unlike Seedance 2.0, EXPLICIT TIMECODES ARE OFFICIALLY SUPPORTED AND RECOMMENDED HERE. Write either '0s-3s: ...' / '3s-8s: ...' or 'Shot 1 (0-3s): ...' / 'Shot 2 (3-8s): ...' and cover the exact clip length given in the CLIP SPEC with no gap and no overrun. Each segment names its visuals, its camera movement, the action, any dialogue and any sound effect. Roughly one segment per 3-5 seconds; a 30-second clip may run to eight or nine segments and should tell an actual story rather than repeat one beat. (4) CLOSING NOTES - everything that must stay constant across the whole clip: camera behaviour, depth of field, lighting, ambience, audio bed and overall mood. DIALOGUE: write it as 'Dialogue (the barista): "we roast it here every morning."' - the speaker is named in parentheses and the line is in double quotes, which is ByteDance's own documented form for this model. Keep each line short enough to land inside its segment. NEGATION - IMPORTANT EXCEPTION: for VISUALS keep to positive description, but Seedance 2.5 explicitly supports negative control over subtitles and audio, with 'No subtitles.' and 'No BGM; generate only environmental sounds and action sounds.' as the documented examples. ALWAYS end the prompt with 'No subtitles.' - burned-in captions are the model family's most common defect in portrait framing - and add the BGM clause whenever the clip should carry only diegetic sound. REFERENCE ASSETS: when photos are attached they arrive as numbered assets. Bind each one explicitly in the text - 'the woman in Image 1', 'refer to Image 1 for lighting' - rather than relying on anything drawn or written inside the picture itself, and where an asset is already accurate simply say it should be referenced instead of re-describing the whole scene. GOLD EXAMPLE (verbatim, ByteDance official Seedance 2.5 prompt guide): 'Realistic nature documentary style, natural lighting and shadows. On a warm afternoon, on a grassy slope in the forest, a chubby panda cub rolls down the hill. The panda has fluffy, realistic black-and-white fur, a small round body, and clumsy, adorable movements. The scene is a green forest slope. The camera is a low-angle medium-wide shot with a slight handheld feel. The framing remains mostly stable, keeping the panda in frame at all times. 0s-3s: A panda cub lies on a green grassy slope, its body round and chubby. It rolls sideways down the slope in clumsy tumbles, flattening the grass under its weight. A light breeze passes through, and sunlight filters through the trees from the upper left, creating dappled light and shadow. 3s-8s: The panda rolls toward the lower right of the frame and gradually comes to a stop, shifting from lying on its side to lying on its belly. Its round face turns toward the camera, and its front paws press into the grass. Low camera position, slight handheld feel, subtly following the panda as it moves toward the lower right. Natural depth of field: the foreground grass is slightly blurred, the panda remains clear, and the background forest is softly out of focus. Natural environmental audio only, including wind, rustling grass, and the soft plop of the panda rolling. The overall mood is warm, realistic, and natural.' Do NOT use Seedance 2.0's symbol conventions here (the full-width brackets, the angle-bracket SFX tags): 2.5's guide writes sound as plain prose inside the segments. Do not write resolution or duration codes into the prompt - they are separate parameters. Aim for a rich 200-400 words for a 5-10 second clip and proportionally more for a long one. ASSET REFERENCING (official 2.5 syntax): open the prompt with a short mapping of what each asset is for - '@Image1 is the woman's appearance reference, @Image2 is the product' - before the one-sentence summary. NEGATIVES: 2.5 officially supports negative control ONLY for subtitles ('No subtitles.') and audio ('No BGM; generate only environmental sounds and action sounds.'); all visual constraints stay positive. DOCS 2026-11 (ByteDance Seedance 2.5 guide): give each spoken line as Name (emotion): "line" and never repeat the line's words after it or attach stage directions to single words - both make the model burn the words in as subtitles. When music must stay out, 'No BGM' alone can leak: also name what to drop (music, score, instrumental, melody) once near the start and once at the end. For an edit or extension of an existing clip write only the change - from A to B, with its seconds when known - and close with 'leave the rest unchanged'; do not re-describe what stays.

## Doctrine

LONG-FORM: the strongest model for a coherent 30s multi-shot story in ONE pass (kling-4 and wan3 also reach 30s) (5/10/15/30); native audio in 10+ languages; ~1.6x seedance - never for an ordinary short. Structure: @Image1-style asset mapping first, a one-sentence summary, a detailed description, timecoded segments '0s-3s: ...' covering the exact duration with no gaps, then closing notes of what stays constant. Negatives allowed ONLY for subtitles ('No subtitles.') and audio ('No BGM; only environmental sounds').

## Format law the critic applies

Seedance 2.5 format law: whole-second timecodes tiling the duration with no gaps; camera bracket opens each shot; Dialogue (speaker): "line" in double quotes; subtitle/audio negatives allowed ('No subtitles.', 'No BGM...'); asset bindings name every @Image/@Video/@Audio once; edit/extend intents force adaptive ratio.

## Video references

Attached clips reach Seedance 2.5 as numbered reference videos after the photos, in upload order: @Video1, @Video2, @Video3. Bind each one to exactly what it lends - motion and choreography, camera move and pacing, colour grade and texture, or the editing rhythm - 'the woman from Image 1 performs the movement of @Video1', 'keep the handheld camera and cutting rhythm of @Video2'. Name what must not carry over. A clip without a binding is ignored. ModelArk refuses a reference clip that shows a photoreal person who is not one of the account's AI Bloggers registered with Seedance (InputVideoSensitiveContentDetected, live probe 2026-09-08). Blockout figures, mannequins, capsules, stylised or animated characters, products, places, hands and objects go through. For the photoreal motion of someone else use minimax, wan3, omni or kling-motion. When a clip is a 3D blockout or playblast (blockout geometry with mannequin, capsule or 3D-model figures in any colours), treat it as a rough guide: bind it to the camera path, the framing and roughly where the characters stand and move, and say that the figures in the clip, whatever they look like, are only stand-ins whose poses and gestures are NOT copied beat for beat - 'the characters do not repeat the movements of the figures in the clip; they look and act as described, in the chosen style, with natural weight and follow-through'. The style comes from the prompt, not from the clip: live action, anime, cartoon or 3D animation alike. Do not ask for 'the framing at each moment' or 'the position of every character', and do not time the actions to the dummies' movements: that makes the model repeat the blockout frame by frame. The blockout's surfaces and simple shapes are placeholders too: flat colours, faceted or noisy surfaces and flat primitive parts become the real materials and forms the prompt describes - say so, and do not claim the blockout has no textures without checking. Copy the figures' motion only when the user wants that exact choreography. The clip is a staging guide (2026-09-15): handled objects and the characters' business with them (the ball, the drink, the phone), gestures, faces, outfits, crowd, plants, props, materials and textures are absent from it on purpose - the prompt writes all of them with their physics (approach, contact, transfer of force, recovery) at real-time speed.
reverse-engineer-anythingmorluto/reaReverse engineer native, managed, Electron/JavaScript, packaged, firmware, and browser targets with REA, and build agent environments (AgentEnv bundles) of web applications from a live instance or a capture, and of Linux desktop apps with a task from a request. Use shipped-artifact or requested runtime evidence to explain features, compare versions, decompile code, or guide a reconstruction. Skip REA for ordinary source-repository architecture analysis.Stars: 472kUpdated Claude CodeCodexVerifiedInstall
diagnosing-superpowersobra/superpowersUse when a superpowers session went wrong and your human partner wants to know why — repeated work, ignored plans, stumbles, poor results, a skill that didn't fire, "it took too long", "why is it so expensive", "what is it doing" — or wants to build a bug report for the superpowers maintainers, for the current session or a past one identified by id or path, on any harness.Stars: 323kUpdated Claude CodeCodexVerifiedInstall
dispatching-parallel-agentsobra/superpowersUse when facing 2+ independent tasks that can be worked on without shared state or sequential dependenciesStars: 323kUpdated Claude CodeCodexVerifiedInstall
executing-plansobra/superpowersUse when executing an implementation plan in the current session as the implementer yourself — your human partner chose inline execution, or no subagent tool is availableStars: 323kUpdated Claude CodeCodexVerifiedInstall