HappyHorse Reference to Video

HappyHorse Reference to Video is the character-consistency member of the HappyHorse family. Where the text-to-video node rolls the dice on who shows up, this one takes reference images - up to nine of them - and generates a video that actually features the person or object in those references. The headline capability is the thing the node description leads with: it supports single-character performances and multi-character interactions. You can feed it one person for a solo scene, or several references and have them act together.

It's a cloud Wan-family model through Comfy's API - no local GPU, no model files, billed per call through your Comfy account. It slots above HappyHorse Image to Video (which animates a single first frame) for any job where identity must survive generation, which in practice is most jobs worth paying for.

The inputs

• model - the dynamic combo holding the real controls. Pick happyhorse-1.1-r2v or happyhorse-1.0-r2v and it reveals:
  • prompt - the description of the video. The trick here: refer to your reference characters by identifier, using character1, character2, etc., so the model knows which reference is which in the scene.
  • reference_images - up to 9 images, one per slot.
  • resolution - 720P or 1080P.
  • ratio - the aspect ratio menu (16:9, 9:16, 1:1, 4:3, 3:4, and more on 1.1).
  • duration - 3 to 15 seconds.
• seed - generation seed; the usual caveat that it's not a reproducibility guarantee.
• watermark - toggle the AI-generated watermark.

What comes out

A single VIDEO output.

Gotchas

• Identifiers are the mechanism, not flavor. A prompt that says "a person walks in" while you've supplied three references leaves the model to guess which person - and it may blend them. Use character1, character2 explicitly.
• Reference quality sets the ceiling. Consistent lighting and framing across your reference images is what lets multi-character scenes read as one scene; mismatched references produce floating-identity weirdness.
• Nine images is a lot of context and a lot of cost. More references = pricier render. Trim the set to what the scene actually needs.
• The prompt hides in the model dropdown - same dynamic-combo behavior as every HappyHorse node.

The pattern to build on: generate a couple of clean reference frames per character (locally is fine), then this node with an identifier-tagged prompt turns them into an actual scene - the closest the Wan family gets to "cast your video."