How to work through it
Create a small manifest for every image, video, and audio file: stable label, owner, permission, intended role, material facts it may support, transformations allowed, and expiry or territory restrictions. The official H3 repository documents separate input conditions for FL2VA and Ref2VA, so choose the task family before assembling the pack. Remove decorative references that do not change a visible, audible, temporal, or identity decision.
State that the subject appearance comes from image A, the camera movement comes from video B, the performance rhythm comes from video C, and the voice or musical timing comes from audio D. Describe how these ingredients should combine in the target shot. Avoid vague instructions such as use all references. Context-IR is designed to reason across modalities, but an explicit relationship remains easier for a human reviewer to audit and correct.
Reference packs often disagree about wardrobe, lighting, geometry, timing, camera direction, ambience, or emotional tone. Decide which source governs each dimension and which details may change. If a product source is authoritative, say that its shape, label, color, and proportions override stylistic material. If a motion clip includes an unwanted camera move, ask only for the subject motion and define the intended camera separately.
After generation, compare the result against each role in the manifest. Check identity and product fidelity, source leakage, dialogue and sound timing, unwanted imitation, visible trademarks, copyrighted style risk, and whether any generated detail could be mistaken for a verified fact. Save the request, reference hashes, model path, output, reviewer, date, and disposition so another editor can reproduce the decision without access to private brainstorming notes.
