Hailuo Minimax Image-to-Video: Subject Reference Fidelity, Motion Quality, and Limits

Hailuo Minimax excels at fast action and complex motion in image-to-video generation, but higher motion magnitudes increase the risk of subject drift and limb mutation.

Hailuo Minimax Image-to-Video: Subject Reference Fidelity, Motion Quality, and Limits

MiniMax (Hailuo AI) released the video-01 variant in mid 2025 and the video-02 variant later that same year, according to the MiniMax API documentation dated September 2025. The web platform at hailuoai.com provides the primary access point, with mobile apps for iOS and Android and a REST API. The API endpoint is POST /v1/video_generation, accepting an image URL via the file_url parameter. Free tier users get one concurrent generation; paid tier limits vary by subscription plan and MiniMax publishes those on the platform pricing page. Generated output ownership stays with the user for both free and paid accounts, and commercial use is permitted under the terms of service, which were last checked in October 2025. The model embeds C2PA metadata to comply with the EU AI Act transparency obligations.

Image Input Specifications and Output Capabilities

File Requirements And Format Support

The Hailuo Minimax image to video pipeline accepts JPG, PNG, and WebP files up to 30 MB. The source image aspect ratio is preserved by default. You can select among output aspect ratios of 16:9, 9:16, 1:1, 4:3, 3:4, and 21:9.

Resolution, Duration And Control Limits

Video 02 generates native 1080p resolution at 24 fps for a maximum clip duration of six seconds. Video 01 outputs 720p native resolution for the same six second clip length. There is no audio generation; the output is silent. The model supports start frame control by using the uploaded image as the first frame, and it also accepts a separate end frame image for specifying the final frame. There is no explicit motion strength slider, no camera control parameters for pan, zoom, tilt, or dolly, and no region mask input for selective animation.

MiniMax Video Generation Subject Reference During Fast Action

The Artificial Analysis Video Arena evaluation in 2025 ranked Hailuo below Kling 1.6 and Runway Gen 3 Alpha on motion quality and prompt adherence. That ranking reflects the specific failure pattern of subject drift during fast action. If your source image shows a cyclist leaning into a turn, Hailuo will generate the lean and the turn. The cyclist's face and clothing may drift by the end of the clip. The model is better at maintaining the scene structure than the subject's details. Backgrounds tend to stay recognisable because they produce less pixel movement. Foreground subjects in rapid motion are the highest risk input.

Hailuo Minimax Image to Video Specifications at 2025 September Release
SpecificationVideo 01Video 02
Native Output Resolution720p (1280x720 for 16:9)1080p (1920x1080 for 16:9)
Clip Duration6 seconds6 seconds
Frame Rate24 fps24 fps
Input FormatsJPG, PNG, WebPJPG, PNG, WebP
Input Size Limit30 MB30 MB
Audio GenerationNoneNone
Start Frame ControlImage used as first frameImage used as first frame
End Frame ControlSeparate end frame image supportedSeparate end frame image supported
Motion Strength SliderNot documentedNot documented
Camera ControlNot documentedNot documented
Region Mask ControlNot documentedNot documented
WatermarkingC2PA metadata embeddedC2PA metadata embedded

Hailuo Fast Action Complex Motion: Where the Model Excels and Where It Breaks

Flicker is another risk with fast action clips. Rapid frame to frame changes in brightness or texture appear when the latent space decoding is inconsistent across frames. Hailuo minimises flicker compared to earlier open weights models, but it is still present in clips with abrupt lighting changes. If your source image has strong contrast or specular highlights, expect flicker in the generated clip. The model's 24 fps output means each frame is a separate denoising step. High contrast regions are more likely to produce divergent latent codes.

Hailuo Minimax Failure Modes and What Causes Them

Aspect ratio distortion appears when the source image aspect ratio does not match the selected output aspect ratio. The preprocessing pipeline resizes the image naively, stretching or compressing subjects. Always crop your source image to match the target aspect ratio before uploading. Hailuo preserves the source aspect ratio by default. If you want a 16:9 output, upload a 16:9 source image.

Hailuo Minimax Image to Video Web Platform Access and API Usage

Content Moderation Rules

Content moderation is strict. MiniMax's terms of service dated October 2025 prohibit sexually explicit content, graphic violence, illegal activities, impersonation, and hate speech. The safety filters can produce false positives on benign source images. If your upload is rejected, try a different crop or lighting condition. The classifiers are broad and may flag skin tones, certain object shapes, or patterns that resemble prohibited content. No model card has been published by MiniMax as of September 2026, so the specific training data and classifier thresholds are unknown.

Consent For Animating Real People

If you plan to animate a real person, you need their explicit consent. MiniMax's terms of service prohibit impersonation, and using someone's likeness without permission violates that clause. Get written consent before uploading a portrait. The platform does not verify consent at upload time, but disputes can result in account suspension and legal liability under the terms.

The Trade Off Between Motion and Fidelity in Hailuo Minimax

The single most important thing to understand about Hailuo Minimax image to video is that you cannot have both maximum motion and maximum subject fidelity. The model's strength on fast action directly increases the risk of source image fidelity loss at higher motion magnitudes. This is not a bug; it is the structural consequence of predicting pixel displacement from a single reference frame. Every image to video model faces this trade off, but Hailuo leans further toward motion than competitors like Runway Gen 3 Alpha or Kling 1.6. If your use case demands a recognisable subject above all else, keep motion minimal. Use slow camera movements, small subject displacements, and provide an end frame image. If your use case demands dramatic motion, accept that the subject may drift, mutate, or shift identity by the end of the six second clip. Plan your edit around that reality. Cut the clip before the drift becomes distracting, or use the first two to three seconds of output where subject fidelity is highest.

Common Questions

How long does a Hailuo Minimax image to video generation take?

The inference time per second of video is not published by MiniMax. Through the web platform, generation takes 30 to 90 seconds for a six second clip during off peak hours. The API documented inference time excludes server queue wait time, which can multiply total turnaround by two to ten times during peak usage. Paid tiers get priority queue placement, but no specific inference speed difference is documented.

Can I generate video longer than six seconds with Hailuo Minimax?

No. The model's native clip duration is six seconds for both video 01 and video 02. You can concatenate multiple clips using FFmpeg, but seam artefacts will appear at the join points because the model does not support true start and end frame conditioning across concatenated clips. The end frame control helps with the transition between a source and target frame, but it does not produce a smooth join between independently generated clips.

Does Hailuo Minimax work for face consistent generation?

Not reliably. Identity shift is a documented failure mode. The model cannot maintain a face's identity across six seconds of motion, especially if the face moves or changes angle. For face consistent generation, use the end frame control with a portrait of the same person as the end frame. This anchors the final frame, but the intermediate frames may still drift. For professional face consistent work, consider models with dedicated identity preservation features that Hailuo does not have.

What happens if my source image is rejected by content moderation?

The platform will return an error and not generate a clip. The safety classifiers are broad and not documented. If a benign image is rejected, try reducing contrast, removing skin tone dominant regions, or cropping out patterned areas that might resemble prohibited content. The false positive rate is not published. There is no appeal mechanism documented in the terms of service.

Can I use Hailuo Minimax output commercially?

Yes. The terms of service permit commercial use for both free and paid users. You retain ownership of the generated outputs. However, the outputs carry C2PA metadata that labels them as AI generated. Platforms like YouTube, TikTok, and Instagram may require you to disclose synthetic content. YouTube uses an altered content checkbox in the upload flow. TikTok applies auto labelling to content with C2PA metadata. Instagram requires a made with AI tag for photorealistic video generated with AI. Failure to apply the label may result in demonetisation or removal.