MiniMax H3 LoRA · Reference to Video

The most capable node in this pack, and the fiddliest

Reference-to-video is where omni-modal models actually earn the word. This node takes up to nine reference images, up to three reference videos, up to three reference audio clips, and up to three LoRAs - and asks you to describe the shot those ingredients should produce. Nothing about it is a starting frame. Everything is a casting decision.

That's a different mental model from image-to-video, and it's the one that made H3's reputation. When the open weights landed, the reaction that spread fastest wasn't about image quality - it was motion control: take a reference video of a person moving, reference images of a different person and a different room, and ask for the movement with the new subject. People were doing Kling-style motion transfer with open weights and a prompt. One widely shared thread made the further point that reference-to-video can replace LoRAs for a lot of identity work, which is worth knowing before you spend a weekend training.

Here you don't need the weights or the 33B of VRAM. You need credits and a well-organised set of references.

How it's wired

Three autogrow groups, each growing a socket as you connect something, plus the LoRA input:

• reference_image_urls - up to nine. Each one after the first raises the quote.
• reference_video_urls - up to three, 15 seconds total. Their seconds are billed like output seconds, so a 10-second reference adds 10 seconds' worth of rate on top of your render.
• reference_audio_urls - up to three, for a voice or a sound.
• loras - up to three H3 adapters, each with its own scale.

Images and audio are encoded and uploaded for you (WAV for audio, MP4 remux for video); each distinct upload is cached for twelve hours, so re-running with a new prompt doesn't re-send the same files.

Then the prompt: refer to the references by number and describe the shot, sound included. That's the tooltip's instruction and it's the difference between a coherent clip and soup - the model will not guess which image is the person and which is the room.

Inputs

• prompt - numbered references, described shot, sound.
• aspect_ratio - 16:9 default, through to 9:21.
• resolution - 480p default, then 540p, 768p, 1080p. Rates run about $0.06/s at 480p up to $0.24/s at 1080p - the priciest per-second node in the pack after LTX 2.5.
• duration_seconds - 3 to 15.
• seed - fixed by default.

Outputs: video and task id.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/popnow-ai/popnow-comfyui

Or Manager > PopNow > Install, or comfy node install popnow-comfyui. No dependencies, no model download, no ffmpeg build. Key from popnow.ai/api/console/keys into Settings > PopNow (or POPNOW_API_KEY), restart once, and find it under PopNow > Video > MiniMax.

What actually goes wrong

Reference video seconds are billed. Three reference clips of 5 seconds each is 15 extra seconds of rate on a 5-second output. People fill all three sockets and then wonder why the quote tripled. Use one reference clip unless you genuinely need three.

Too many references, no clear instruction. Nine images don't compose themselves. Two or three references with an unambiguous prompt beat nine with a vague one, every time.

Duration and total budget assumptions. At 1080p, a 15-second output plus 15 seconds of reference video is roughly seven dollars. Run 480p first, always, and set Settings > PopNow > Max credits per run so a badly-scoped quote stops before credits are held.

Upload ceilings. Input video is remuxed to MP4 and capped at 90 MiB, audio at 90 MiB, images at 10 MiB. Long 1080p reference clips hit that ceiling - trim them to the motion you actually want copied.

A refused render. content_rejected and generation_failed are refunded; a job you cancelled in ComfyUI is not. And if an error ends in request_id ..., that's the string support needs - copy it before you clear the console.