realistic-video-prompting

Realistic Video Prompt Generator

Transform a vague idea from the user ("I want a video with the everyday feel of a Korean neighborhood") into high-density, executable, de-AI-ified structured video prompts that look like real footage.

Core Philosophy

The essence of a high-quality video prompt is answering one question:Who filmed this footage, with what device, in what era, and for what purpose?

Once the "source identity of the footage" is determined, all details can be derived. A DV casually filmed by a friend in 2005 will shake, lose focus, and suddenly shut off; a blogger's vertical phone video in 2024 will have a beauty-filter feel and feature talking to the camera; a 1990s surveillance camera has no sound, a top-down angle, and a low frame rate.

The second key insight:AI models have strong default aesthetics(cinematic camera movement, perfect composition, commercial color grading, clean image). To achieve realism, you must actively turn off these default behaviors with negative constraints and implant "imperfect events" into the timeline.

Workflow

Step One: Determine the Four Elements (none may be missing)

If the user has not made it clear, first ask these four questions (they can be combined into one prompt):

Filmer identity: Who is filming? A friend, family member, passerby, Vlogger, surveillance, documentary crew, or an unattended fixed camera?
Device and era: DV, VHS, 8mm film, iPhone, GoPro, surveillance, professional cinema camera? Which era?
Subject and scene: Whom to film? Where? What time period?
Emotional goal: Warm nostalgia, lonely detachment, tense unease, calm healing?

If the user has already provided enough information, infer directly and confirm your understanding in one sentence before outputting.

Step Two: Generate paragraph by paragraph according to the seven-part template

Output strictly in the following order, with each section as an independent block and no information mixed between sections:

主要角色 → 地点 → 视觉风格 → 摄像风格 → 时间轴 → 音频 → 目标

Writing rules for each section:

1. Main characters

Write only visible concrete features: age range, makeup, color/material/fit of each piece of clothing, shoes, accessories, hairstyle details (including small features such as stray hairs/bangs)
Do not use abstract adjectives (fashionable, good-looking, stylish)
Add one sentence of skin/texture description (e.g., "realistic skin texture")
Add one sentence of personality temperament (used to drive body language)
Must end with a consistency lock sentence: "Maintain consistent identity, clothing, hairstyle, and appearance throughout the video"

2. Location

Start with time period + region + scene type
List 8 to 12 specific environmental elements (potted plants, clotheslines, utility poles, tree shadows, etc.); the elements must mutually corroborate the same world
Add dynamic light and shadow description (tree shadows moving, cloud shadows passing)
End with negative constraints: Exclude commercial elements that models love to add (shops, advertisements, crowds, cafes)

3. Visual style

Use 4 to 6 phrases to define the realism level: such as "hyperrealist documentary realism, genuine improvised behavior, natural body language, an unscripted slice-of-daily-life feel, strong environmental authenticity"

4. Camera style

This is the core paragraph for removing the AI feel. Consult references/camera-aesthetics.md, select the "defect parameter pack" for the corresponding device and write it in completely
Principle: translate the device aesthetics into 6 to 10 specific physical phenomena (focus hunting, exposure fluctuations, rolling shutter, compression artifacts, etc.)
Must end with a triple negative constraint: no stabilization, no cinematic camera movement, no modern color grading (adjust according to the device)

5. Timeline

Total length 10 to 15 seconds, cut 5 to 7 shots, each shot 2 to 3 seconds, use the 00:00–00:02 format
Each shot contains three layers: subject action + environmental response (wind, light, animals, passersby) + one imperfect event
Select imperfect events from the list in references/anti-ai-checklist.md , and do not use the same type more than once across the 7 shots
There must be a chain of everyday logic between actions (sitting → walking into an alley → encountering a cat → hanging laundry), do not write it as a collage of unrelated shots
The ending must have a closing method: sudden cut to black (like turning off the device), walking out of frame, camera drooping down to the ground, etc., matching the identity of the person filming

6. Audio

First declare the mode (ambient natural sound only / live sound capture + ambience)
List 6 to 9 specific sound elements, corresponding one-to-one with the scene elements (if there is a cat, there is meowing; if there is a clothesline, there is the sound of fabric)
End with a triple negation: no music, no sound design, no narration (adjust as needed)

7. Goal

A one-sentence soul anchor, format: "Capture XX, as if it were a XX of XX, ending with 4 to 6 adjectives"
This sentence is the basis for the model's trade-offs when details conflict, and must be self-consistent with all preceding paragraphs

Step 3: Self-check (mandatory before output)

Check item by item against references/anti-ai-checklist.md :

Does the character section have a consistency lock sentence?
Does the location section exclude commercial elements?
Does the camera section have at least 6 physical defect descriptions + 3 "no XX" statements?
Does every shot in the timeline have an imperfect event? Are there any repeats?
Does the ending have a closing method?
Do the sound elements in the audio section correspond to the visual elements?
Has any abstract adjective slipped into the full text (beautiful, cinematic, high-end)?
Is the goal sentence self-consistent with the identity of the person filming / the device / the era?

Step 4: Delivery

Output the complete prompt (in Chinese or English according to user needs; when feeding it to the model, English usually works more reliably, so you may proactively provide an English version)
Attach a brief explanation: what the "source material identity" setting of this prompt is, and which knobs the user should adjust when they want to make changes (device pack, time period, emotion words, closing method)

Reference files

references/camera-aesthetics.md: defect parameter packs for 8 types of devices/media, required reading when writing the camera style section
references/atmosphere-dictionary.md: a translation table from mood words to specific visual phenomena, consult when writing the location and timeline
references/anti-ai-checklist.md: a list of sources of the AI feel, a library of negative constraint sentences, and a library of imperfect events, required reading during self-check

Variant handling

The user wants a non-realistic style(animation, cyberpunk, commercial film): the seven-part structure remains unchanged, but the camera section is changed to the aesthetic parameters corresponding to the medium (animation frame rate, the camera movement language of commercials), and the negative constraints are changed to exclude elements unwanted in that style
The user wants multiple characters: each character gets an independent paragraph of description, each with its own consistency lock sentence, and the interactions between characters are clearly specified in the timeline
The user only wants to optimize the existing prompt: first diagnose according to the seven-part structure which sections are missing and which sections have insufficient density, provide a diagnosis table, and then rewrite
Long videos exceeding 15 seconds: it is recommended to split them into multiple prompts for segmented generation, with each keeping the character section completely consistent to maintain continuity