视频｜HappyHorse 1.1/1.0 参考生视频 R2V

HappyHorse's "reference-to-video" mode, and it's the one that does something the other HappyHorse nodes can't: fuse up to nine reference images into a single generated video. Same model family (1.1 default, 1.0 for legacy), same cloud API mechanics, but where I2V animates one still, R2V builds a scene out of many - think IP character consistency across shots, or a product shown from multiple angles with the identity held together. The nine-image budget is generous, and it's exactly the input count a character-consistency job wants.

One thing to get straight before you lean on the multi-image fusion: HappyHorse is one of the more obscure models this pack resells, and "reference fusion" on closed models is never as literal as it sounds. It's not ControlNet-style conditioning where image N maps to a precise region. You're handing the model a pile of visual references and a prompt that describes how they combine - so the prompt does a lot of work. The pack's own docs make the move explicit: with multiple references, tell the model what each one is ("image 1 is the person, image 2 is the outfit, image 3 is the setting").

How it works

参考图1 through 参考图9 are optional IMAGE ports (at least one required), plus a 图片URL列表 text field that overrides them - one public URL per line, the recommended route when your images are already hosted because it skips the upload step. A prompt in 生成指令, sync or async in 执行方式, and the usual five outputs come back: 🎬_视频输出, 📁_本地保存路径, 🔗_视频云端链接, 🆔_任务ID, 📄_完整日志.

The inputs that matter

• 参考图1–参考图9 / 图片URL列表 - your visual references. Start with fewer than nine; each extra image costs upload time and can dilute the result.
• 生成指令 - describe the combined scene and each reference's role.
• 执行方式 - sync for single tests, async for batches or long jobs.
• 清晰度 - 1080P default; 720P while iterating.
• 视频时长 - 3–15 seconds.
• 水印 - watermark on or off.
• 随机种子 - fixed default.

Install and first run

cd ComfyUI/custom_nodes
git clone https://github.com/htrert/ComfyUI-Tikpan-Pro

Restart (or Manager → "Tikpan"), paste your sk- key from https://tikpan.com, connect 2–3 images, and write a prompt that names each image. Test with one or two references before you throw all nine at it - you'll learn the model's fusion behavior faster with a minimal case.

Common issues

Uploads are the recurring pain: many images means more chances for one to fail, so compress aggressively or use 图片URL列表 with hosted files. The multi-image input is also where "the model ignored image 3" complaints come from - usually the prompt never said what image 3 was. Timeouts and double-billing behave like every other video node in the pack: keep the 任务ID, use the task-query node to recover, and check the log before resubmitting.