Source: raw/Stop_the_Generic_AI_Look_Seedance_2.0_Prompt_Engineering_for_Cinematic_Video_HeyGen.md — YouTube tutorial h8e1GZBKm0k by an unidentified HeyGen-channel creator. The walkthrough is structured around HeyGen’s Avatar Shots feature with Seedance 2.0 as the motion model, layered on top of Avatar 5 for A-roll.

A prompt-engineering framework for HeyGen’s Avatar Shots feature that lifts AI video out of the “generic rubbery look.” Thesis: the model is fine — most users prompt like writers, not like directors of photography. The fix is a 5-element prompt structure (subject / action / environment / camera / style) with one camera move per shot, layered into multi-shot and multi-avatar choreography. Pairs Avatar 5 (long-form A-roll) with Seedance 2.0 (cinematic motion + storytelling).

Key Takeaways

Tool stack and roles

  • Avatar 5 — long-form A-roll. The creator’s entire A-roll in this tutorial is Avatar 5 output.
  • Seedance 2.0 — cinematic motion, environments, storytelling.
  • Avatar Shots (the feature under walkthrough) — combines both: digital twin + fully cinematic Seedance-generated scenes. Unlocks full-body motion, multi-shot cinematic sequences, and up to three avatars in one scene.

The 5-element prompt structure (every shot)

ElementWhat to specifySkip-it failure mode
Subjectwho/what is in the shot, specific identifying details (outfit, identity tags)identity drift across cuts
Actionone intentional action — what they’re doing right nowinconsistent character motion
Environmentlocation + lighting (named lighting: golden hour, warm light, etc.)output falls apart on environment
Cameraexactly ONE camera move per shot (slow push-in, dolly, pan, etc.)over-busy / messy camera work
Styleoverall aesthetic + constraints (shallow depth of field, film grain, stable framing, consistent identity)generic AI look

Hard rule: one camera move per shot. Combining moves degrades output.

Working example — single shot prompt

Professional man in a navy blazer stands in a modern glass office overlooking a city skyline at golden hour. He adjusts his sleeve and looks confidently toward the camera. A slow camera push-in, cinematic shallow depth of field, warm golden light, soft reflections on the glass, subtle film grain, smooth motion, stable framing, natural gestures, consistent identity.

Maps cleanly onto the 5 elements:

  • Subject: professional man in navy blazer
  • Action: adjusts sleeve + looks toward camera
  • Environment: modern glass office overlooking city skyline at golden hour, warm golden light, soft reflections
  • Camera: slow push-in
  • Style: cinematic shallow depth of field, film grain, smooth motion, stable framing, natural gestures, consistent identity

Pre-shot decisions (before writing the prompt)

  • Pick the avatar.
  • Pick output ratio — vertical or horizontal.
  • Pick resolution — recommended 1080p.
  • Decide duration AFTER writing the prompt — re-read your prompt and pick the duration that matches its pacing. This is a non-obvious tip; don’t lock duration first.

Multi-shot sequences (“think in beats”)

  • Each beat = a discrete shot.
  • Give each beat its own:
    • Time window (timestamps inside the prompt)
    • One camera move
    • Purpose (why this beat exists in the story)
  • You can change eras, outfits, worlds in one generation. That’s “multi-clip prompting” — one prompt, multiple scenes, fully directed.

Example beat structure (paraphrased):

Location: 1800s British royal study. Subject: avatar wears powdered wig, formal aristocrat outfit, holding a teacup. Camera: medium shot with slow push-in. Script: …

Then repeat with new beat block for each scene in the sequence.

Multi-avatar choreography (up to 3 avatars per scene)

  • Becomes blocking + choreography, not just dialogue.
  • Define in the prompt:
    • Each avatar’s position in the frame
    • Each avatar’s movement relative to others
    • Camera support for the movement (how camera tracks the action)
  • Skipping these turns the scene “random.” Specifying them makes it feel like a production.

For multi-avatar:

  1. Describe the desired shot, feel, and look at the top.
  2. Then enumerate each character — position + wardrobe.
  3. Then camera motion.
  4. Then beat-by-beat scene description with character moves + camera moves.

Elements — physical references, locations, products, clothing

Once you start adding assets (locations, backgrounds, clothing, products), prompts get shorter:

  • Reference the asset by name in HeyGen.
  • Still use the 5-element structure, just less verbose because the asset injects the visual identity.
  • This is where continuity across shots becomes reliable.

Example shortened prompt (with elements pre-attached):

The avatar faces the camera in a modern office wearing a pastel outfit while holding a mug, speaking with confident and engaging energy.

Workflow recommendation (creator’s stack)

  • Avatar 5 for main message: founder videos, education, product explainers.
  • Avatar Shots + Seedance for: storytelling, transitions, visual depth.
  • Layer Avatar Shots inside an Avatar 5 A-roll project rather than building entire videos in Seedance.

Finishing stage — Topaz upscale settings (added 2026-06-22)

A separate user-curated walkthrough adds the post-generation finishing step this article’s “stop the generic AI look” thesis was missing — upscaling AI-video output in Topaz Video AI (concrete, reusable settings; source: raw/Bringing_AI_characters_into_my_apartment_+_My_Upscale_Settings_steal_this_process.md):

  • Model: use Iris (best for faces); Proteus also good. Avoid Starlight Precise (the default) — “takes forever.”
  • 16:9 → output resolution 3840×2160 (4K), bit rate Dynamic. 21:9 → custom 5040×2160, bit rate Constant 24.
  • Add digital noise = 2–3 (distinct from film grain). Recover Detail = on (modest). Focus Fix = off (AI generation is never out of focus). Input condition “low quality”, tuning advanced/automatic.
  • Codec = H.264 (skip ProRes unless heavy color grading / 10-bit log). Don’t add film grain in Topaz (slows render) — apply grain overlays in the editor instead; if used in Topaz: type gray, amount 70–80, size 1.5–1.8.
  • Finish in a 4K timeline, then add grain + light color correction. (The video’s other half — compositing an AI character into a real still frame — was too vague to capture: the tool is never pinned down and no prompt/settings are given.)

2026-10-30 update: the same discipline for HeyGen Video 1

HeyGen’s walkthrough of its own model, HeyGen Video 1 (raw/I_Tried_HeyGen_Video_1_Here_s_Everything_It_Can_Do.md), teaches a four-part “shot list for a film crew” prompt: subject plus one action, camera, look (mostly light), and sound plus what to leave out (“no music, no logo, no writing on the walls”). Because Video 1 generates audio with the picture, the sound and exclusion line is the extra slot. HeyGen’s Script Enhance fills in camera, light and look but not sound or exclusions. Its advice overlaps with this article’s one-action, one-camera-move rule: one subject in one place doing one thing, with a steady camera.

Open Questions

  • Creator + channel identity — first-party HeyGen account or affiliate creator? Not clear from transcript.
  • Seedance 2.0 official documentation — referenced as a model name; primary docs URL not in transcript.
  • Cost model — generations are HeyGen-plan-gated; tier pricing not in walkthrough.
  • Aspect-ratio + resolution combinations — what’s supported per HeyGen plan tier.
  • Max sequence duration per multi-clip prompt — not stated; relevant for full-production workflows.

Try It

  1. Write your first 5-element prompt — pick a professional or personal scenario, fill all five slots explicitly. Don’t skip any.
  2. One camera move only — if your draft prompt has two camera moves, cut one. Always.
  3. Set ratio + resolution + avatar BEFORE writing the prompt; decide duration AFTER. Re-read your prompt and let pacing dictate length.
  4. Try a multi-shot sequence next — write 3 beats, each with timestamps, one camera move each. Generate as a single multi-clip prompt.
  5. Add multi-avatar choreography last — block positions + movements + camera tracking explicitly. Don’t introduce more than 3 avatars per scene.
  6. Move from prompts to elements — upload your branded product / outfit / location into HeyGen, then reference by name in prompts. Continuity becomes reliable.
  7. Stack with Avatar 5 — use Avatar Shots for storytelling beats inside an Avatar 5 A-roll project, not as the whole video.