What MiniMax H3 Is Actually Good At

The curtain gave the models away. MiniMax H3 barely disturbed the finished advertisement around it. WAN 2.2 made the curtain billow, but also reconstructed more of the scene. LTX 2.3 moved fastest and changed the composition most. None was simply better. Each protected a different thing.

Across six finished layouts, H3 was our preferred result when preserving the source mattered most. It also generated stereo audio with the video. The tradeoff was clear: H3 was the slowest route, and its restraint could look almost static when the brief called for bold movement.

We ran a series of fixed-seed comparisons on an RTX 3090 with 24 GB of VRAM. Each comparison kept its source, prompt, requested duration, output shape, and seed fixed within that test. We saved the ComfyUI workflow, output, media metadata, and timing so that a surprising result could be checked instead of remembered.

• Six finished vertical advertisements animated with H3, WAN 2.2, and LTX 2.3.
• A 10, 12, 16, and 20-step H3 sweep on the same source and seed.
• Four fixed scenes comparing native 12, native 20, and conservative Spectrum 20.
• A generated-speech control comparing audio activity with visible mouth motion.
• Three reference-conditioning routes using the same two characters.
• A three-scene long-form chain comparing native quality with a six-step Turbo LoRA.

Automated metrics helped locate drift and timing changes, but visual and auditory review made the final decisions. Pixel motion can increase because a model is doing something useful, or because it is destroying the design. A single number cannot tell those cases apart.

This is also a comparison of products and workflows, not model weights in isolation. WAN 2.2 used interpolation, the LTX result used our LTX 2.3 generation workflow, and MiniMax H3 used its matching generation graph. A shared seed improves repeatability, but different models do not turn that seed into the same initial latent or sampling path.

Six finished vertical campaign layouts were resized to the same portrait dimensions. Each model received the same five-second target and a scene-specific prompt that locked typography, logo, crop, and graphic panels while requesting motion only in the photographed region. No overlay was reapplied afterward. This deliberately made text preservation part of the model test.

In our review, H3 preserved the layout best in all six cases. It also changed the fewest pixels, which helps explain both its stability and its restrained feeling. WAN moved more. LTX moved most and finished much sooner, but rebuilt more of the composition.

The times are averages across the six layouts. Source match and motion combine manual review with pixel-change checks, but the labels describe only this test set. Restrained motion can protect a layout or leave a shot feeling lifeless, which is why the videos matter more than a single score.

Native 20 steps were our reference trajectory. Native 12 was faster but visibly weaker in the harder people and couple scenes. Conservative Spectrum kept a 20-step schedule while forecasting six of the expensive transformer evaluations. Across four fixed cases it averaged about eleven minutes, roughly four minutes faster than native 20 but one and a half minutes slower than native 12.

Decoded-video SSIM placed this Spectrum configuration closer to native 20 than native 12 in all four cases. SSIM measures structural pixel similarity, not narrative correctness, so we also watched and listened to every result. A more aggressive preset moved further from the native trajectory and reduced both motion and audio level.

Tested latency-quality compromise
schedule: 20 steps
native evaluations: 14
forecast evaluations: 6
sampler: RES multistep
reference: native 20

H3 predicts separate visual and audio latents inside one joint generation process. That lets visible action, ambience, effects, music, and speech influence the same sequence. It does not guarantee accurate words or perfect lip sync.

The two clips had similar visible speech timing, while the tested Spectrum run finished about two and a half minutes sooner. That is useful, but it is not proof of accurate lip sync. Watch whether mouth movement follows the cadence, then listen for the actual words and any abrupt changes in the voice.

Listening still has to catch wrong words, stutter, extra voices, or an unnatural onset. A later continuation test produced exactly that kind of onset artifact even though the media file and overall waveform looked valid.

In the first two-character test, we passed the same references through three routes with the same fixed settings and native 20-step sampling. The official Ref2VA route was soft and relatively static in this example. An experimental FL2VA route using the same reference node looked sharper.

The late-block hybrid kept a little more of the mage's hair curl and the gold material on his robe, but it was slower and the visual gain was small. That was one seed, not enough to crown a new default.

Two follow-up subjects made the conclusion less tidy. On a character with a product, the hybrid had a small material advantage but no decisive identity gain. On a branded bicycle, both experimental routes approximated the lettering and neither was safe for client-facing brand detail. The official Ref2VA checkpoint remains the sensible general reference route until a broader comparison proves otherwise.

In one three-scene test, native 20-step H3 produced the better reference-heavy chain but took about 51 minutes. The six-step LightX Turbo route was roughly 2.6 times faster and failed semantically: one character changed into another, duplicates appeared, and an audio onset jumped at a join. Upscaling could not repair those failures.

A third-party context-loop workflow then carried a short overlap of video and audio into each new scene and removed that overlap from the final assembly. It reduced the visual seam and eliminated the malformed vocal spike seen in a naive continuation. It still could not force new dialogue to start at an exact timestamp.

One successful chain does not establish a universal recipe, but it supports a practical architecture: generate bounded shots, carry compact context, keep persistent identity references, checkpoint every accepted segment, and master long-form audio separately when exact chronology matters.

H3 adds a useful high-fidelity option, but the model is too slow and too restrained to be the automatic answer to every brief. Start with the cost of failure. If a changed logo or face ruins the shot, test H3. If a still result ruins it, compare WAN. If iteration speed matters most, LTX may be the better creative tool.

• Keep exact typography deterministic when the design permits it.
• Test H3 first for completed designs and reference-critical shots.
• Compare WAN when stronger movement matters and LTX when speed matters.
• Treat fast Turbo reference chains as previews until identity and chronology pass review.
• Judge audio onset, dialogue accuracy, identity, and geometry separately.

The cleanest interface is not a wall of model switches. Ask whether the user wants to protect the source, maximize movement, or iterate quickly, then map that intent to a tested route. The model name is an implementation detail; the creative tradeoff is the useful choice.

Compare the two strongest routes directly in MiniMax H3 vs WAN 2.2, then test a source frame in Movey's image-to-video workflow.

Primary model, implementation, and acceleration sources used to interpret the fixed-seed comparisons.

• Official MiniMax H3 model card documents the open FL2VA and Ref2VA checkpoints, joint audio-video architecture, and supported generation modes.
• Official ComfyUI MiniMax H3 guide documents the local model modes, native stereo audio workflow, resolution guidance, and required ComfyUI version.
• Comfy-Org MiniMax H3 model repository provides the packaged local model assets used by the ComfyUI workflows.
• ComfyUI core documents native audiovisual generation support, model loading, offloading, quantization, and workflow execution.
• Spectrum for MiniMax H3 implements the forecasted transformer evaluations tested here and documents the method's approximation and audio-related limitations.
• Spectrum research implementation describes adaptive spectral feature forecasting for diffusion sampling acceleration.
• MiniMax H3 Contex Loop provides the bounded video and audio context mechanism used for the continuation follow-up.