MiniMax H3 Image-to-Video

Most image-to-video nodes take one start frame and hope the model invents a plausible continuation. MiniMax H3 Image-to-Video does that, and then it adds a genuinely rarer party trick: you can also give it a last frame, and the model interpolates between the two. It's the difference between "animate this photo" and "animate a transition from frame A to frame B," and it makes this the node to reach for when you want a shot that resolves somewhere specific rather than drifting.

It's the image-animation sibling of the H3 Text-to-Video node in the same pack, same hosted API, same one-pass picture-plus-stereo-audio at 24fps. Feed it a portrait and a prompt describing the motion, and you get a clip that moves the way you asked - with sound, if your prompt asks for it.

How it works

The node gives you two ways to hand it images, because API-video nodes live in two worlds. The preferred path is a ComfyUI IMAGE tensor:

• first_frame (IMAGE) - sent to WaveSpeed as a base64 data URI. This takes precedence the moment it's connected.
• first_frame_url (STRING) - the fallback when no tensor is connected. Handy for chaining off another node that already emits a URL.

Same split for the ending frame: last_frame (IMAGE) or last_frame_url (STRING). Leave both empty and it's a plain one-frame animation.

Two details in the tooltips are worth reading twice. First, the output canvas follows the first image's aspect ratio - there's no aspect_ratio dropdown here because your image is the canvas. Second, duration snaps to the model's frame grid, so a 5s request lands near 5.2s. Don't fight it.

Inputs that matter

• prompt - required. Motion and camera movement. Same Audio: line convention as the H3 T2V node if you want a steered soundtrack.
• first_frame / last_frame - the two IMAGE sockets above. Connect the first one and you're animating; connect both and you're interpolating.
• duration - 3–15s, default 5.
• resolution - 480p (~$0.04/s) or 768p (~$0.10/s), straight from the tooltip.
• seed - a real API seed, sent to the model. Fixed value = reproducible video and a cached result you don't get re-billed for; -1 randomizes each queue.

Output is video_url (STRING), the pack-wide convention for video nodes - pipe it into Preview Anything to see or download the clip.

Installing and configuring

Part of the ERPK Collection pack. Fastest route is ComfyUI Manager: search erpk, install ERPK Custom Nodes, restart. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/eRepublik-Labs/comfyui-nodes-erpk.git erpk
cd erpk && pip install -r requirements.txt

It runs against WaveSpeed's API, so you need a WaveSpeed key from wavespeed.ai in Settings > ERPK > API Keys. Nothing is downloaded, nothing runs on your GPU - the server does the generating, the node just polls (10s interval, 15-minute timeout) and hands back a URL.

Common issues

• My first_frame is being ignored - check whether you also left text in first_frame_url. When a tensor is connected, it wins outright; the URL input is only a fallback for when the socket is empty.
• The output is a different aspect ratio than I expected - the canvas follows the first frame's ratio, so a square source gives a square clip. If you want widescreen, start from a 16:9 image.
• Interpolation looks weak - treat the two frames as the same scene, not two unrelated photos. H3 interpolates between them, and like any interpolation it's happiest when the endpoints share composition and lighting.