Skip to main content

Same 10 prompts, two models — Wan 3.0 on the API vs MiniMax H3 on a local GPU

· 13 min read
AICU API Team
api.aicu.ai

Frame strips for the "authority" prompt. Top: Wan 3.0 pushes in and tilts up to the glowing tip. Bottom: MiniMax H3 keeps the trophy and its text almost unchanged while a flare grows at the tip

In Key visuals with Wan 3.0 (1), we turned one trophy photo into 10 teaser videos using prompts written the way Alibaba Cloud's official guide recommends. This time we passed the same photo and the same 10 prompts to MiniMax H3, running locally in ComfyUI on a single 20 GB GPU, and put the results side by side.

Short version: H3 is the one that keeps the logo and text intact, and it moves cloth and light boldly. Wan is the one that follows camera directions, and it is about 3 times faster.

The two also differ in what you need to run them. Wan on the AICU API needs only a key: no GPU, three clips at once, and you can still end a clip on your untouched logo by passing it as last_frame. H3 needs a GPU with about 20 GB of memory and about 8 minutes per clip, and its license excludes the EU, the UK, South Korea and the United States.

Setup​

Wan 3.0MiniMax H3
Where it runsAICU API (wan3.0-video)ComfyUI on our own GPU (NVIDIA RTX 4000 Ada, 20 GB)
InputThe trophy photo (1671×941) as first_frameThe same photo cropped to 1088×608, as first_frame only (no last frame)
Resolution / length720p, 5 seconds (10 seconds for #10)1088×608, 124 frames = 5.2 seconds (243 frames = 10.1 seconds for #10), 24 fps
SettingsDefaults20 steps, res_multistep sampler, seed 20261113
SoundGeneratedGenerated
Time per clip111–204 seconds, up to 3 in parallel481–501 seconds (5 s), 1,342 seconds (10 s), one at a time
Cost1,400 AP per second at 720p: 7,000 AP for a 5-second clip, 77,000 AP for all 10. No GPU neededNo AP, but you need a GPU with about 20 GB of memory (we used our own)
Failures00

H3's length must be 5 + 17n frames, so 5 seconds becomes 124 frames (5.17 s) and 10 seconds becomes 243 frames (10.1 s). Peak VRAM was 17–19 GB for 5 seconds. The 10-second clip finished without running out of memory.

The prompts are exactly the ones in the Wan article. They follow the official guide's "image to video = motion + camera movement" formula, so they are not tuned for H3.

The ComfyUI workflow​

H3 runs on ComfyUI's built-in nodes (ComfyUI 0.30.0 or later). The workflow is short: load the four model files, then MiniMaxH3ImageToVideo → sampler → decode video and audio → save. Download the API-format JSON here and change the prompt and image name: h3-i2v-api.json

NodeSetting
UNETLoaderminimax_h3_fl2va_pruned_int8_convrot.safetensors
CLIPLoaderqwen3vl_32b_minimax_h3_nvfp4_awq.safetensors, type minimax
VAELoader × 2minimax_h3_video_vae_fp16.safetensors, minimax_h3_audio_vae_fp32.safetensors
MiniMaxH3ImageToVideoprompt, width 1088, height 608, length 124, first_frame
CreateVideo24 fps, with the decoded audio

The model files are on Hugging Face at Comfy-Org/MiniMax-H3. H3 is released under the MiniMax H3 Community License Agreement, which excludes some territories (the EU, the UK, South Korea and the United States). Read the license before you use it.

The 10 results​

In every comparison, the top row is Wan 3.0 and the bottom row is MiniMax H3, one frame per second from left to right. The video below each comparison is H3. The Wan videos are in the original article.

01 The request as written​

A teaser video that evokes shimmering radiance, wind, influence, intelligence, history,
authority, creativity, and contribution to the community.

01: Wan (top) cuts between wide and close shots; H3 (bottom) stays on one shot

Wan cut the shots on its own (wide → close → wide). H3 stayed on a single shot: the cloth ripples and a flare lights up at the tip. With only abstract words, H3 falls back to "the photo, gently moving."

02 The basic formula (motion + camera)​

Sunlight shimmers across the facets of the crystal trophy, and the white cloth ripples gently
in the breeze. The camera slowly pushes in.

02: Wan pushes in until the text fills the frame; H3 barely pushes in

Wan pushes in clearly. H3 pushes in only a little. Its light on the cloth and the facets moves more than Wan's.

03 Shimmering radiance​

Golden caustic light dances and ripples across the white cloth around the trophy, and a soft
lens flare pulses at its tip. Backlight, rim light, warm tones. Fixed camera.

03: both keep the camera still; H3's flare at the tip pulses and grows

With a fixed camera, the two look much alike. H3's flare at the tip pulses and grows as the prompt says, and it ends with a soft lens flare across the frame. Both are good candidates for loop footage.

04 Wind​

A gust of wind sweeps across the scene: the white cloth billows in slow waves, the leaves behind
sway, and dappled light flickers over the crystal. The camera moves slowly to the left.

04: H3's cloth billows much more than Wan's, and the camera drifts left

H3's best result. The cloth billows in big waves and the camera drifts to the left. Wan moves the cloth too, but H3's motion reads more clearly as "a gust of wind." The trophy and its text stay intact the whole time.

05 Authority (low angle)​

The trophy stands still and solemn as the light grows stronger along its edges. Low-angle shot,
the camera slowly pushes in and tilts up toward the glowing tip. Side light, high contrast.

05: Wan pushes in and tilts up to the tip; H3 barely moves and lets the flare grow

The biggest difference. Wan pushes in and tilts up, ending on the glowing tip. H3 barely moves the camera; only the flare at the tip grows stronger. If you need a specific camera move, Wan is the one that delivers it.

06 History (time-lapse)​

Time-lapse photography: the shadow of the trophy sweeps slowly across the white cloth as the
light shifts from bright morning white to deep sunset gold. Fixed camera.

06: both shift from white light to sunset gold

Both shift from white light to sunset gold. Wan draws a band-like shadow that doesn't match the trophy's shape. H3 turns gold toward the end and the shading on the cloth changes, but you don't see a shadow clearly sweeping across it.

07 Intelligence (circuits of light)​

Inside the crystal, fine lines of light trace and branch like a network, rising slowly from the
base to the tip, while refracted light glints across its facets. Orbiting camera movement, a slow
arc of less than 30 degrees.

07: Wan fills the trophy with branching lines of light; H3 shows them only faintly at the end

Wan fills the trophy with lines of light from the base up, just as written. H3 shows the lines only faintly in the last second or two. H3 tends not to add things that aren't in the photo, and in exchange the text stays readable.

08 Community (spreading light)​

Small sparks of golden light spread outward from the base of the trophy across the cloth like
ripples, multiplying as they travel toward the edges of the frame. The camera slowly pulls out
to reveal the whole scene.

08: Wan draws concentric rings and pulls far back; H3 spreads radial streaks of light across the cloth

The two read the prompt differently. Wan draws concentric rings and pulls the camera far back, so the text becomes unreadable. H3 spreads radial streaks of light across the cloth toward the edges of the frame and keeps the trophy large. H3's version is closer to "spreading outward."

09 The sound formula (sound effects + music)​

The white cloth ripples in the wind and light shimmers across the crystal trophy. Sound: wind
rustling through the leaves, a clear crystal chime as a flare passes the tip, and a slow, rising
orchestral score with strings and soft brass. No dialogue.

09: both have calm motion; H3's cloth moves a little more

Like Wan, H3 generates sound for every clip, even when the prompt says nothing about sound (01–08). Listen to the sound in 09 and judge it yourself. As with Wan, if you plan to add narration or music, assume you will replace the generated audio.

10 Multi-shot (10 seconds)​

A teaser for an international award, in three shots. Shot 1 [0-3 s]: close-up, light glides
along the crystal edge of the trophy, fixed camera. Shot 2 [3-6 s]: medium shot, the white cloth
ripples in the wind and golden caustics dance around the base, the camera moves left. Shot 3
[6-10 s]: the camera pulls out and rises to reveal the trophy standing in sunlight against the
green trees, and a lens flare blooms at its tip. Sound: soft wind, a crystal chime, an orchestral
swell. No dialogue.

10: Wan starts from the photo's wide composition; H3 jumps to a close-up of the crystal edge right after the first frame, then follows the three shots

Both follow the three-shot structure, but they treat the first frame differently. Wan starts from the photo's wide composition, so Shot 1 is not a close-up. H3 shows the photo for just one frame, then cuts straight to a close-up of the crystal edge as Shot 1 asks. Shots 2 and 3 follow the prompt. If you want a multi-shot clip that opens on a close-up, H3 is closer to the script; if you want it to open on your image, Wan is.

This clip took 1,342 seconds (about 22 minutes) to generate.

What we learned​

Wan 3.0 (API)MiniMax H3 (local)
Keeping the logo and text intactGood, but lost when pushing in or pulling backBest. Readable throughout almost every clip
Following camera directions (push in, tilt, pull out)Follows themMostly ignores them; small moves at most
Physical motion (cloth, wind, light)GoodBold (04, 08)
Adding things that aren't in the photo (07)Draws themDraws them faintly
Multi-shot openingStarts from your first frameShows your first frame for 1 frame, then follows the script
Speed2–3 minutes, parallelAbout 8 minutes for 5 s, 22 minutes for 10 s, one at a time
What you needAn API keyA GPU with about 20 GB of memory, ComfyUI, and a location the license covers

How to choose:

  • Logo shots, product shots and material where the text must stay readable, without a rush: H3.
  • Teasers that rely on camera moves, effects that add something new, or anything you need today: Wan on the AICU API.
  • Camera moves and a readable logo in the same clip: Wan with the original image as last_frame. In Key visuals with Wan 3.0 (2), the last frame of a transition to the logo matched the original at PSNR 37.8 dB, and we still swapped in the original still at the cut.
  • Both make sound by default. Plan to replace it for screenings.

Try it on the AICU API​

Pass your photo as first_frame and, if the clip must land on your logo, the logo as last_frame:

curl -X POST https://api.aicu.ai/v1/video/generate \
-H "Authorization: Bearer $AICU_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "wan3.0-video",
"first_frame": "<base64 of your photo>",
"last_frame": "<base64 of your logo>",
"prompt": "the camera slowly pushes in as light sweeps across the trophy",
"resolution": "720p",
"seconds": 5
}'

The call returns at once with a status_url to poll. The AP is held when the job starts and returned in full if it fails. Up to three jobs run at once per key. Get a key on the dashboard, and see the Video API guide for every parameter.

⭐︎ Prices are as of publication. You can check current rates on Pricing and with GET /v1/models.


MiniMax H3 is licensed under the MiniMax H3 Community License Agreement, Copyright © 2026 MiniMax. All Rights Reserved.