UGC Model Swap

• When to use
• User shares a UGC video and wants the same video with a different person

User describes a UGC scenario (challenge, reaction, review, try-on, unboxing) and wants a specific character type User wants multiple variants of the same clip with different character types Model Always use Seedance 2.0 (seedance_2_0). Do NOT use Cinematic Studio 2.5 Motion Control for this — it fails silently with reference images in this context. Do NOT use Kling 3.0 — also fails silently with reference images here.

Step 1 — Analyze the source video Run video_analyze(video_source=, category="analysis_templates") to extract:

• Setting (location, lighting, background)
• Character appearance and outfit

Action sequence — described as ONE continuous scene, not split into numbered scenes Props (what they are, how they're used) Audio / dialogue Step 2 — Character reference image (optional but recommended) If the user provides a photo of the replacement character:

Upload via higgsfield_upload or use the attached media ref directly

Pass as medias: [{"role": "image", "data": {"id": "...", "type": "media_input"}}] Reference with @Image1 at the start of the prompt If no photo provided — describe the character in text only (e.g. "young attractive white woman with long blonde hair").

Step 3 — Build the prompt Structure (single continuous scene — NO "Scene 1 / Scene 2" splitting)

[CHARACTER DESCRIPTION] [SETTING] [OPENING SHOT — camera angle, what's in foreground] [CONTINUOUS ACTION — everything that happens, in order, as one flow] [CAMERA RULES — face lock, never tilts, etc.] [AUDIO] Key rules

1. Single scene, no numbered splits. Describe the entire video as one continuous flow. Splitting into "Scene 1 / Scene 2" confuses the model and breaks continuity.

2. Face always in frame — state it repeatedly and explicitly. If the action involves the character looking down, reaching, placing something — the camera must NOT follow. Use:

"HER FACE IS ALWAYS IN FRAME. Camera locked on her face. Never tilts down. Never follows her hand." State this both inline in the action description AND as a separate closing rule.

3. Props — be hyper-specific. If there's one prop (e.g. one ice cube), explicitly forbid variations:

"just one cube, no bucket, no bowl, no container, no pile — only ONE standalone [prop] resting in the palm"

4. Physical actions — describe the body mechanics clearly. Vague actions get wrong results. Instead of "she puts it between her knees", say:

"she deliberately reaches her hand down and tucks the [prop] inside the waistband of her jeans, pushing it down — her hand moves below frame, the prop disappears from view" The more specific the body mechanic, the more accurate the output.

5. Audio — always include. Always add "generate_audio": true and an Audio line in the prompt:

Audio: "[line 1]" / "[line 2]" / [sound description] / "[line 3]" Natural room acoustics.

Prompt template

[Character description: age, ethnicity, hair, complexion, outfit — be specific]. INTERIOR [setting] — [lighting description], [background details].

Opening: [shot type] — [what's in foreground and how]. [Character] stands in [setting] looking at [prop/camera] with a neutral expression. HER FACE IS ALWAYS IN FRAME.

[She/He] [smiles/reacts] and [action 1 — e.g. reaches out to take the prop].

Camera stays tight on [her/his] FACE in close-up throughout the entire video —

[action 2 — detailed body mechanics, e.g. "she deliberately reaches her hand down and tucks the ice cube inside the waistband of her jeans"]. [Prop/object] disappears out of the bottom of the frame. [Her/His] face fills the shot — [reaction sequence: expression shift → gasp → laugh / shock / etc.].

FACE IS ALWAYS CENTER FRAME. Camera locked on [her/his] face. Never tilts down. Never follows [her/his] hand. [Skin/lighting note]. Handheld smartphone UGC feel.

Audio: "[line]" / "[line]" / [sound] / "[line]" Natural room acoustics. Step 4 — Generate json

higgsfield_generate({ "requests": [{ "type": "generation", "model": "seedance_2_0", "media_type": "video", "params": { "prompt": "", "aspect_ratio": "9:16", "duration": 10, "generate_audio": true, "medias": [ { "role": "image", "data": {"id": "<media_id>", "type": "media_input", "url": ""} } ] } }] }) Omit medias entirely if no reference image is provided.

Batch variants

To generate multiple character variants of the same clip, submit all in one higgsfield_generate call with multiple requests (up to 8 concurrent). Change only the character description in each prompt — keep everything else identical.

Example variant types that work well:

• Different ethnicities / ages
• Different outfits (latex suit, cheerleader uniform, formal wear, etc.)
• With/without reference photo
• Pitfalls

Problem Fix Model generates a bucket/pile instead of one prop Add explicit negative: "no bucket, no bowl, no container, no pile — only ONE standalone [prop]" Camera tilts down and follows the hand Repeat face-lock rule both inline AND at the end of the prompt Character drops the prop instead of placing it deliberately Describe full body mechanics: "deliberately reaches her hand down and tucks... her hand moves below frame" Kling 3.0 / Cinematic Studio 2.5 fails silently Use Seedance 2.0 — it's the only model that reliably handles this workflow Reference image character not matching Ensure role: "image" (not start_image) and reference with @Image1 at prompt start Video opens in wrong location (outdoor instead of kitchen) Add "INTERIOR [room]" explicitly at the top of the prompt before any action description Video splits into disconnected scenes Remove any "Scene N" labels — write as one continuous paragraph