Video summary
How I Make Ultra Realistic AI Videos using Seedance 2.5
Main summary
Key takeaways
Seedance 2.5 workflow for “ultra-realistic” AI videos
- The creator argues realism comes from more than image/video sharpness. It’s largely about avoiding “fake” artifacts in:
- Motion
- Facial behavior
- Dialogue and performance
- They outline a step-by-step approach using Seedance 2.5 plus supporting tools/model choices.
Why AI looks “fake”: Uncanny Valley concept
- The video explains the uncanny valley: overly human or “too perfect” renders can feel “off” or creepy.
- To avoid it, prompts should include human imperfection, such as:
- Flaws and asymmetry
- Slight inaccuracies
- Natural movement
- Imperfection should be intentional—use models/prompts that produce believable irregularities rather than “polished” output.
Model selection for character generation (imperfection > perfection)
- The creator compares different models and claims the best results for “cinematic human” faces come from Hicksfield Soul.
- They emphasize that this model produces:
- Facial imperfections (e.g., wrinkles, bags under eyes, imperfect hairlines, not fully sharp details)
- A more human/cinematic look for portraits
Character sheet (reference) as the baseline
For generating videos, they recommend building a character sheet first as a reference.
Suggested pipeline
- Generate a character sheet with SeaArt 5.0 Pro
- Can generate up to 2K
- Example aspect ratio: 16:9
- Avoid tiny distracting accessories (example: removing a small hat)
- Very fine details may harm realism/consistency in videos.
- Upload the character sheet into the prompt box
- Also include an “original character” image for identity reference.
- Add constraints such as:
- “do not recast”
- “do not beautify”
- “do not use as an image background”
- Add camera/lens realism cues (example provided):
- 35mm lens
- f/1.8
- shallow depth of field
- fine film grain
- muted/desaturated color grade
Why this matters
- Investing effort in the reference frame helps prevent wasting time/credits during generation.
Dialogue is a major realism factor
- The creator claims dialogue strongly determines whether AI videos feel real.
- They compare:
- Better dialogue prompts (more dramatic, emotionally timed, acted-like)
- “Average AI” dialogue (generic and less performance-accurate)
“Prompt like a director” (scene blocking + tone + timing)
They break down an effective dialogue prompt structure:
- Describe where characters are and how the scene progresses (e.g., bedroom → living room)
- Specify the main action beats
- Assign:
- Who speaks first
- Interruptions/cuts
- The emotional state
- Include pauses/timing implicitly through staging language so the AI “acts out” the scene.
Accents and speaker styling via prompt structure
For accents, they stress including:
- Speaker identity/name
- Region and accent type (example: Kingston / Jamaican vibe)
- Delivery style (e.g., unhurried vs shock-based sudden changes)
They also recommend specifying emotional/physical reactions that change vocal delivery (e.g., a voice jump when shock hits).
Symbol-based prompting for dialogue/text and sounds
- They mention using bracket-like symbols to cue the model, including:
- Which parts are spoken text by the character
- How to structure sound cues (using different symbols for different audio categories)
Acting performance words for physical realism
- They use detailed performance/action words to control body language and micro-behaviors.
- Example concept: a multi-step acting sequence such as:
- Smile changes → finger tightening → reluctant looking back → freezing/screaming → frantic pedaling → hiding → shaky breath release
- They note prompts can be long (e.g., four pages) when packed with action-beat descriptors.
Emotion prompting + triggers
- Recommend describing emotion intensity from subtle to explosive.
- Include triggers that cause emotion/action, and pair emotion changes with visual cues.
Camera direction is required for realism
The creator stresses that realistic outputs require basic filmmaking knowledge, especially:
- Camera focus
- Main subject vs secondary subject
- Use rack focus and focus-change triggers
- Camera movement
- Example movement types: crash zoom, handheld, tracking shot, orbit/reveal, etc.
- They direct viewers to their website for camera movement prompt examples
- Camera angles / shot planning
- Use establishing shots, then tighten framing
- Maintain readability (avoid distracting cuts during key action)
- Escalate intensity with more aggressive camera behavior
Sound prompting: diegetic vs non-diegetic
They describe two audio categories:
- Diegetic sounds (characters would hear)
- Non-diegetic sounds (music/extra drama cues the audience hears)
Example prompting format
- Detailed footstep sound design (e.g., door hiss, steel slam, squeak, wet splash)
- Non-diegetic dramatic cues aligned to visual timing (e.g., sub-bass hit when a light dies)
Editing pass to fix “AI artifacts”
If the generated clip isn’t believable, they recommend an editing pass:
- Trim/cut out unrealistic slow-motion segments
- Add zoom-ins/zoom-outs to guide attention (e.g., eye focus)
- Add subtle realism effects:
- Noise overlay (small intensity; key for believability)
- Edge blur (camera-like softness)
- Cinema bars (example: ~10–15% crop on top/bottom)
Key “how-to” items highlighted
- Create character sheet references first (SeaArt 5.0 Pro), then upload into Seedance 2.5
- Use imperfect realism prompts; avoid “beautify” / “recast”
- Prompt dialogue like a director: beats, tone shifts, interruptions, emotional timing
- Use speaker details for accents and specify vocal delivery changes under stress
- Use bracket/symbol cues for spoken content and sound instructions
- Add detailed acting-performance words for body language continuity
- Direct camera like filmmaking: focus, movement, angles, shot readability
- Prompt sound with diegetic vs non-diegetic separation
- Post-edit with trimming, zooms, and subtle noise/blur/cinema bars
Main speakers/sources
- Main speaker: The video creator (self-narrating walkthrough of “How I Make Ultra Realistic AI Videos using Seedance 2.5”)
- Sponsored source: Hicksfield (used to explore “Seedance 2.5” ecosystem tools and models)
- Mentioned AI models/tools:
- Hicksfield Soul
- SeaArt 5.0 Pro
- “SORA 2.5” (referenced for comparison/next video)