MiniMax H3 Reference-to-Video: Image to Video Online

Every row comes from VdoBloom's live model configuration, the generation endpoint and the credit table. The detail that separates this mode from H3 Image-to-Video: the photos are references, not the first frame, so the clip can open on a composition none of them shows, and you choose the aspect ratio yourself.

The rate matches the other two H3 modes (seconds times 4 at 768P, seconds times 6 at 2K) and the number of photos does not change it; the first 5 are free and VdoBloom never sends more than you attach. The jobs below are examples, not presets; any whole second from 4 to 15 works at either resolution. Dollar figures use the Lite plan rate of $0.05 a credit. New accounts get 10 free credits, which does not reach the 16-credit minimum, so plan on a top-up before your first render.

In H3 Image-to-Video your photo becomes frame one and the clip continues from it. Here the photos never appear as a frame. MiniMax reads them for appearance, the face, hair and outfit of a person, the shape and label of a product, the look of a place, and then generates the opening frame from your prompt. That is why two people from two different photos can stand in one shot, and why the camera can start somewhere none of the photos were taken from.

The trade is that nothing is pinned. If the prompt does not say what to keep, the model decides, so the photos and the prompt have to work together.

The three H3 variants share the price list, the 4 to 15 second range and the 768P/2K tiers. They split on what controls the picture: an exact first frame, several appearance references with a free composition, or nothing but the prompt.

All three take more than one reference photo and appear in the Image to Video tab. From the credit table: H3 is the cheapest of the three per second, Wan 2.7 Reference-to-Video is 32 credits for 5 seconds at 720p and 64 for 10, and Seedance 2.5 is 126 credits for 5 seconds at 720p with its reference photos tagged @Image1 and @Image2. Wan 2.7 offers 5 or 10 second lengths at 720p or 1080p; H3 offers any second from 4 to 15, a 2K tier and up to 9 photos. 768P and 720p are not identical, so read the price columns as 720p-class.

No tab opens on this model by default, so you select it. The model, the 4 to 15 second range and the per-second rate are the same everywhere it appears.

Refer to the photos in plain language by their upload order: the woman in Image 1, the jacket in Image 2, the cafe in Image 3. Then write the scene the way you would brief a director: who does what, where it happens, how the camera moves (slow push-in, orbit, static two-shot) and what is heard, including any line of dialogue. Identity holds best when the prompt says so outright: keep each person's face, hair and outfit exactly as in their photo. Leave that out and the model is free to restyle them.

Because the first frame is generated, the prompt owns the composition. Say where each subject stands and how close the camera is, or the opening shot may crop out the person you cared about. One main action per clip reads cleanly; several unrelated actions in 5 seconds blur together. For 10 to 15 second clips describe the motion in stages. Test at 4 seconds and 768P for 16 credits with the final photos attached; once both identities hold, re-run at the length, resolution and ratio you will deliver.

Photos and a prompt are the whole input: no first or last frame to pin, and no reference video or audio on VdoBloom even though the provider lists them. If you need the clip to open on exactly your picture, use H3 Image-to-Video at the same price. Up to 9 photos are accepted, but more references do not raise the price and do not guarantee a better clip; the first 5 are free and every photo is another thing the model has to reconcile. The ceiling is 15 seconds. 2K costs 50% more per second than 768P, so drafting at 2K is the easy way to overspend. Within VdoBloom's reference models H3 is the lowest per-second rate (20 credits for 5 seconds at 768P against 32 for Wan 2.7 Reference-to-Video at 720p and 126 for Seedance 2.5 at 720p), while Wan 2.7 adds a 1080p tier and Seedance 2.5 adds first and last frame control. The 10 free signup credits do not cover the 16-credit minimum. H3 is a proprietary MiniMax model that VdoBloom runs through its provider, KIE; if a job fails on the provider side the credits come back automatically.

VdoBloom does not rank models on subjective quality. Everything above is spec and price; whether H3 keeps your two faces recognisable in one shot is something a 16-credit 4-second test answers faster than any review.