AI Video Model Prompt Guides
Every video model was trained to read prompts in its own way. These guides show how Kling, Seedance, MiniMax, LTX, Wan, FLUX and Gemini Omni each want a shot written, with a template, a real example and the mistakes to avoid.
Kling 3.0
Subject + Subject movement + Scene, then optionally Camera language + Lighting + Atmosphere (Kling's published text-to-video formula).
Read the guideSeedance 2.5
Subject + action, then Scene, Visual style, Camera movement or cut, Audio. Fill only the slots you need; two or three sentences is the sweet spot.
Read the guideMiniMax H3
Instruction line, blank line, then integrated_multimodal_description + overall_soundscape + non_diegetic_music.
Read the guideLTX 2.5
Shot + Scene + Action + Character + Camera + Audio, written as one flowing paragraph in present tense.
Read the guideWan 3.0
No official formula exists. A working order: Subject and action, then Environment, then one Camera move, then Sound and Dialogue.
Read the guideFLUX 3
[camera] shot of [subject] [action] in [environment]. [supporting visual and motion details]
Read the guideGemini Omni 1.1 Flash
Cinematography + Subject + Action + Context + Style and ambiance (Google's published structure for its video models).
Read the guideHow Each Model Reads a Prompt
The same shot needs a different shape for each model. This is the short version; each guide has the detail.
| Model | Prompt shape | Speaks dialogue |
|---|---|---|
| Kling 3.0Kuaishou | Short declarative sentences: subject, movement, scene, then camera, light and mood. | No |
| Seedance 2.5ByteDance | Two or three sentences in slot order, with sound in bracket channels. | Yes |
| MiniMax H3MiniMax | A strict structure: an instruction line, then labelled picture, soundscape and music fields. | Yes |
| LTX 2.5Lightricks | One flowing present-tense paragraph, no tags, emotion as physical cues. | Yes |
| Wan 3.0Alibaba | Natural prose with spoken lines in quotes; no published guide. | Yes |
| FLUX 3Black Forest Labs | A director's note: camera, subject, action, place, then sounds tied to the picture. | Yes |
| Gemini Omni 1.1 FlashGoogle | Cinematography, subject, action, context, style; always describe the sound. | Yes |
Rather not learn seven dialects? The prompt translator writes one shot in all of them at once.
Or skip the homework: PXLJ writes the prompt for you.
Pick a model in PyxelJam Studio and every shot's prompt is written in that model's own style, from your script.
Write it or import it
Framing, moves, lengths
A drawn frame per shot, same cast every time
Approved frames become video
Trim, order, music, title card
Export or send to your channels
Free to start. No card needed.
Questions
Why does each AI video model need a different prompt?
Each model was trained on captions written in a particular shape, and follows prompts best when they look like those captions. MiniMax H3 expects labelled fields, Seedance reads sound from bracket channels, LTX wants one prose paragraph, and Kling has its own subject-first formula.
Which AI video models can speak dialogue?
Seedance 2.5, MiniMax H3, LTX 2.5, Wan 3.0, FLUX 3, Gemini Omni 1.1 Flash write spoken lines with lip-sync, each in its own format. Kling 3.0 render without speech.
Where do these guides come from?
From each maker's published prompting guide where one exists, cross-checked against the prompts PyxelJam Studio sends these models every day. Each guide lists its sources.