References
A reference is not a frame of the clip but an answer to “who” and “what”. You can shoot the same character in different scenes and they stay themselves.
New 2.1 features are in beta. Check availability and the final price in the order panel.
Contents
- 01. What reference mode is
- 02. How to send an order
- 03. How to refer to references
- 04. The full prompt: six sections
- 05. subject_definitions - who is who
- 06. summary - what happens
- 07. retention_analysis - what is kept
- 08. detailed_description - the scene itself
- 09. The two sound sections
- 10. Reference precision
What reference mode is
In /ref, a picture defines a character, object or style instead of the first frame. Send one to three pictures and describe what should happen. With MP on, plain language is enough; the full prompt format is for manual control. References use Navi Flow.
How to send an order
/ref The character from the first image walks through a city at night. No music.
Attach 1–3 pictures as an album. Duration is 4, 6, 8 or 15 seconds depending on your plan: 4 seconds without a plan, 6 seconds with any Ani Video plan or AniWorld, 8 and 15 seconds with Study or AniWorld. Separate add-ons are unavailable with a video reference.
To follow a source clip
Add one clip up to 20 seconds (Draft: any Ani Video plan, Cinema: Study). It supplies actions, shots, camera, setting and rhythm; pictures supply character appearance and the desired style. For multiple characters, say who replaces whom and what should stay.
/ref The character from image 1 replaces the person on the left; image 2 replaces the person on the right. Follow the clip’s actions and camera. Keep its background and text.
Check the picture numbers in Sources. Video-reference MP uses Spark with low imagination; the button is on/off only. Quality is Draft or Cinema. Beyond 10 seconds, the price grows faster than duration: base prices are 358/448 tokens for 10 s and 1185/1508 for 20 s. The panel shows your total.
Results depend on the sources: complex motion, faces and text may not transfer accurately.
How to refer to references
In the prompt a picture is referred to by number, not by description.
The first picture of the album is <Picture 1>, the second is <Picture 2> and so on.
Characters and objects get their own numbers <Subject 1>, <Subject 2>, and a separate section says where each of them came from.
The full prompt: six sections
18+For manual prompts: expand
subject_definitions: <Subject 1> is the girl in <Picture 1>, with long blonde hair and a black bow. <Subject 2> is the room in <Picture 2>, with a wooden desk and a window. summary: [reference generation] <Subject 1> sits at the desk in <Subject 2> and turns toward the camera. retention_analysis: <Subject 1> (appears in [Shot 1]): fully_preserved - her face, hair and clothing are retained. <Subject 2> (appears in [Shot 1]): fully_preserved - the desk and window are retained. detailed_description: The target video is 2D-animated with soft daylight. [Shot 1] A medium shot of <Subject 1> seated at the desk. She slowly turns her head toward the camera and smiles. Her hands stay on the desk. The camera stays completely still, no zoom. overall_soundscape: Quiet room tone with faint street noise outside the window. non_diegetic_music: N/A
A walk-through of the sections is below.
subject_definitions - who is who
18+For manual prompts: expand
One line per source. You say what each number means and which features have to be kept.
<Subject 1> is the girl in <Picture 1>, with long blonde hair and a black bow.
One character can be assembled from several pictures:
<Subject 1> is the woman whose appearance comes from <Picture 1> and whose outfit comes from <Picture 2>.
summary - what happens
18+For manual prompts: expand
One paragraph, starting with the task type in square brackets:
[reference generation] <Subject 1> turns toward the camera and smiles.
Types can be added together with a plus, but for the usual case with pictures [reference generation] is enough.
Only those already declared in subject_definitions.
retention_analysis - what is kept
18+For manual prompts: expand
One line per number: where it appears and how carefully it is carried over.
<Subject 1> (appears in [Shot 1]): fully_preserved - her face, hair and clothing are retained.
| Mark | What it means |
|---|---|
fully_preserved | Carried over completely |
partially_preserved | Used, but some features are changed |
attribute_transfer | Features carried over onto another character |
weak_reference | Only the general likeness of style or mood is kept |
Do not lower the mark because of them.
detailed_description - the scene itself
18+For manual prompts: expand
The same as integrated_multimodal_description in regular mode, with two differences.
First: the style is declared before the shot marker
The target video is 2D-animated with soft daylight. [Shot 1] A medium shot of <Subject 1>…
Second: where a character appears, their number goes with them
Later in the text the number is repeated, and there is no need to explain again who it is.
Everything else is as in the regular guide: shot markers, timecodes At 00:02.500, speech through (S1) and <d>[Russian] …</d>, holds for calm scenes and no holds for scenes with motion.
If the character speaks, the character number and the voice number are written together:
<Subject 1> (S1) turns toward the camera and says: <d>[Russian] Привет.</d>
The two sound sections
18+For manual prompts: expand
The same as in regular mode.
N/A for full silence only.N/A if there is no music.Reference precision
Photo references always use precise mode. There is no separate switch in the panel.