H3
An AI video and image generation tool that supports AI-driven creation and editing of images and videos. It provides features such as Camera Control for creative direction and is targeted at AI content creators.
Core Facts
Prompt Studio can generate structured MiniMax H3 video scripts.
With the accelerator enabled, MiniMax H3 takes approximately 2 minutes to generate 1 second of video, approximately 14 minutes for a 480P ten-second video, and approximately 20–30 minutes for a 768P ten-second video.
H3 supports multiple video generation modes including text-to-video, image-to-video, and 3×3 grid video generation.
Timeline (last 90 days)
Among T0-tier models, Seedance 2.5 has the highest visual quality, with H3, the Gemini series, and other leading models following closely behind.
In the 2026 AI video generation model leaderboard, closed-source models overall still slightly lead open-source ones, while H3 from MiniMax stands out as the top model in the open-source space.
Cloud platform credit conversion pricing: Seedance 2.5 averages approximately 2 yuan per second, Seedance 2.0 approximately 1 yuan, and H3 and the Gemini series approximately 0.6 to 0.8 yuan.
The refinement solution improved degradation control, but contrast still worsens over time and dissolve artifacts still appear in the background, meaning the problem has not been fully resolved.
The author cited a long single-take video of SeeDance 2.5 as a reference and expressed belief that the same could be achieved with H3.
The test configuration uses 1 character reference image paired with ref2va to generate 10 segments of 8 seconds each (totaling approximately 80 seconds).
The test uses 10 base sampling steps plus 4 refinement steps, with the author deliberately lowering the step count to expose degradation issues under extreme conditions.
The image frames used in the test came from videos generated by H3, with a resolution of 0.3MP (approximately 300,000 pixels).
H3's official announcement currently only consists of a "Cooking Updates" teaser, with no specific release date, technical implementation details, or actual effect demonstrations disclosed
H3 released a preview of the Camera Control v3 update in the community
31 more timeline events
All Facts (20)
Prompt Studio can generate structured MiniMax H3 video scripts.
80%VerifiedWith the accelerator enabled, MiniMax H3 takes approximately 2 minutes to generate 1 second of video, approximately 14 minutes for a 480P ten-second video, and approximately 20–30 minutes for a 768P ten-second video.
80%VerifiedH3 supports multiple video generation modes including text-to-video, image-to-video, and 3×3 grid video generation.
70%UnverifiedThe BF16 full model of MiniMax H3 is over 50 GB in size and requires close to 60 GB of VRAM to load completely.
60%UnverifiedCloud platform credit conversion pricing: Seedance 2.5 averages approximately 2 yuan per second, Seedance 2.0 approximately 1 yuan, and H3 and the Gemini series approximately 0.6 to 0.8 yuan.
50%UnverifiedIn the 2026 AI video generation model leaderboard, closed-source models overall still slightly lead open-source ones, while H3 from MiniMax stands out as the top model in the open-source space.
50%UnverifiedAmong T0-tier models, Seedance 2.5 has the highest visual quality, with H3, the Gemini series, and other leading models following closely behind.
50%UnverifiedThe author cited a long single-take video of SeeDance 2.5 as a reference and expressed belief that the same could be achieved with H3.
50%UnverifiedThe refinement solution improved degradation control, but contrast still worsens over time and dissolve artifacts still appear in the background, meaning the problem has not been fully resolved.
50%UnverifiedH3 is an AI video and image generation tool
50%UnverifiedMiniMax H3 is integrated into ComfyUI workflows via Custom Nodes.
50%UnverifiedOn Windows, virtual memory can be configured to run MiniMax H3 when VRAM is insufficient.
50%UnverifiedThe output results of different quantized versions of MiniMax H3 differ little on routine tasks and are difficult to distinguish with the naked eye.
50%UnverifiedThe tail-frame generation module of MiniMax H3 is 7.6 GB in file size; it is not the complete model but is a required component.
50%UnverifiedSeparating references—using one reference image for the face and another for the clothing—can achieve decoupling of features across multiple images.
50%UnverifiedWhen two characters with distinctive physical features are present in the same scene, the model tends to incorrectly mix one character's hair color and clothing onto the other.
50%UnverifiedUsing strong negative phrasing (such as "nothing of that picture's is used") and strongly contrasting colors can prevent attribute cross-contamination between reference images.
50%UnverifiedAI video models have a tendency to accidentally generate inappropriate content when producing sexy outfits such as sequined or form-fitting clothing, requiring explicit constraints (such as "always dressed") to avoid this.
50%UnverifiedThe prompt distinguishes between layered sound design for overall_soundscape (overall soundscape) and non_diegetic_music (non-diegetic music).
50%UnverifiedPrompts can plan action timelines down to the millisecond, for example by specifying different actions at 00:02.500 and 00:05.500.
50%