References and Source Media

References and source media

Prompt-driven references

Give media stable aliases, then mention only what a scene needs:

@hero         picture identity or appearance
@performance  motion video
@voice        audio reference

Tagged Ref2VA activates the aliases present in the current scene prompt and compiles them to compact native H3 labels. A scene that does not mention the tag does not consume that reference.

Use:

• Tagged Picture Ref for identity, appearance, or a visual object;
• Tagged Motion Ref for action and performance without requesting the source identity, wardrobe, setting, or composition;
• Tagged Audio Ref for voice, cadence, ambience, or other sound guidance;
• Tagged Video Ref when the whole video reference—not just motion—is the intended control.

Timed semantic picture anchors

#hero[2.50s] presents a tagged picture to Qwen again at an approximate scene-local time. It is semantic reinforcement, not a frame lock, pose lock, mask, ControlNet, or continuation boundary.

Start with two or three sparse anchors. Too many repeated pictures can resist the source video's changing pose.

Source Timeline

Register long source picture and sound once:

Load Video / Load Audio → Source Timeline → Preflight → Loop Start
                                              ↓
                                         Current Shot
                                              ↓
                                      active scene window

Path-backed media stays lazy. Only the active overlap-aware scene window is decoded, while the descriptor and fingerprints remain available for resume, Run Manager recovery, and assembly.

Use Source Timeline for:

• an exact source soundtrack;
• a long motion-reference video;
• continuing an existing clip;
• source-aligned masked editing;
• recovery without repeating full decoded AUDIO wires.

Do not return a Current Shot audio-slice fingerprint to Plan; that creates a graph cycle. Structured scene dependencies already record the exact source window used for generation.

Compliance modes

Mode	Behavior
Strict	Compiles valid active aliases and stops on an unknown, inactive, or invalid tagged reference.
Soft	Compiles valid aliases, warns about unresolved prompt tags, and continues.
Disabled	Passes prompt text unchanged and makes scheduler-owned validation warning-only.

Errors outside reference scheduling remain real errors, including model/VAE wiring, sampling, continuation geometry, and checkpoint integrity failures.

Native numbering

Active pictures, videos, and audio each receive compact numbering for the current scene. A later active @hero may become <Picture 1> when an earlier picture is inactive. This is why stable aliases are safer than handwritten native numbers when scene schedules change.

Picture storyboard mode

Tagged Ref2VA can compile #picture[time] anchors as separate Qwen-only pictures with approximate timing instructions. This preserves more still-image detail than the default timestamped-video presentation but costs more Qwen visual tokens. Start at 512 or 768 resolution and use 1024/1280 selectively.

Legacy scheduled references

Numeric selectors remain available when explicit ranges are useful:

Selector	Active scenes
blank, all, or *	Every scene
3	Scene 3
1:5	Scenes 1 through 5
1,3,5:8	Scenes 1, 3, and 5 through 8

Disjoint selectors are valid for references; disjoint generation ranges are not valid for a continued loop.

Resume fingerprints

Static loader media, aliases, and selectors should participate in the Plan's generation fingerprint. Source Timeline and structured scene dependencies record scene-local media windows automatically. A changed reference invalidates only the accepted scenes that used it.

Next: Plan and scene authoring, Advanced workflows, or Node reference.