MiniMax-H3
A multimodal video generation model released by MiniMax capable of simultaneously generating video and synchronized stereo audio tracks within a single request. It supports multiple task modes including text-to-video with audio, first-and-last-frame conditioned generation, and image/video/audio reference-conditioned generation.
Timeline (last 90 days)
Using the MiniMax-H3 image-to-video model on the RTX 3090, three attempts received blind scores of 2/10, 4/10, and 9/10 respectively.
The third attempt preserved the geometry with no banding, but when shown to the person responsible for approving the demo, the feedback was "fake as hell, my Minecraft rain physics," pointing out that the heavy explicit rain streaks rendered as particle sprites.
The key fix was setting the CLIPLoader type to `minimax` rather than `hidream` and adopting the official sampler chain: RandomNoise and BasicGuider (no negative prompt, no CFG), KSamplerSelect with res_multistep, BasicScheduler simple 8 steps, loading the turbo 8-step LoRA via LoraLoaderModelOnly, with the latent extracted from slot 1 of the MiniMaxH3ImageToVideo node.
MiniMax-H3 achieved a success rate of 29.90% on WorldGuide-Bench under the condition of receiving a reference action plan.
MiniMax-H3 received an additional action plan input in the comparison while WorldGuide did not, indicating that merely providing action planning information is not enough to guarantee task completion, and whether the executor can faithfully realize the plan is where the bottleneck lies.
MiniMax-H3 achieved a success rate of 32.73% on VideoCraft-Bench when given only the goal condition.
MiniMax-H3 achieved a TextVScore of 68.32 on the T2V track of TextVBench, representing the strongest performance.
MiniMax-H3 can be run in ComfyUI via SGLang Diffusion.
First-frame rendering for the MiniMax-H3 model can complete up to approximately one minute earlier, and peak VRAM usage on A100, B200, and RTX PRO 6000 has been reduced by 25–29 GiB.
Kijai updated the VAE model for MiniMax-H3 (MH3), adopting an int8 quantized version that significantly reduces VRAM usage with almost no loss in image quality.
22 more timeline events
All Facts (20)
Using the MiniMax-H3 image-to-video model on the RTX 3090, three attempts received blind scores of 2/10, 4/10, and 9/10 respectively.
50%UnverifiedThe third attempt preserved the geometry with no banding, but when shown to the person responsible for approving the demo, the feedback was "fake as hell, my Minecraft rain physics," pointing out that the heavy explicit rain streaks rendered as particle sprites.
50%UnverifiedThe key fix was setting the CLIPLoader type to `minimax` rather than `hidream` and adopting the official sampler chain: RandomNoise and BasicGuider (no negative prompt, no CFG), KSamplerSelect with res_multistep, BasicScheduler simple 8 steps, loading the turbo 8-step LoRA via LoraLoaderModelOnly, with the latent extracted from slot 1 of the MiniMaxH3ImageToVideo node.
50%UnverifiedMiniMax-H3 achieved a success rate of 32.73% on VideoCraft-Bench when given only the goal condition.
50%UnverifiedMiniMax-H3 received an additional action plan input in the comparison while WorldGuide did not, indicating that merely providing action planning information is not enough to guarantee task completion, and whether the executor can faithfully realize the plan is where the bottleneck lies.
50%UnverifiedMiniMax-H3 achieved a success rate of 29.90% on WorldGuide-Bench under the condition of receiving a reference action plan.
50%UnverifiedMiniMax-H3 achieved a TextVScore of 68.32 on the T2V track of TextVBench, representing the strongest performance.
50%UnverifiedMiniMax-H3 can be run in ComfyUI via SGLang Diffusion.
50%UnverifiedFirst-frame rendering for the MiniMax-H3 model can complete up to approximately one minute earlier, and peak VRAM usage on A100, B200, and RTX PRO 6000 has been reduced by 25–29 GiB.
50%UnverifiedThe updated VAE is hosted in the Comfy-Org repository on Hugging Face (huggingface.co/Comfy-Org/MiniMax-H3), and users can download it to replace the original VAE file.
50%UnverifiedThe updated VAE model filename is minimax_h3_video_vae_int8_convrot.safetensors.
50%UnverifiedBy cropping a damaged-wall shot as a reference image and instructing the model to pay attention to it in retention_analysis, environmental damage consistency can be maintained across different shots.
50%UnverifiedIn MiniMax-H3 video generation, photorealism is strongly correlated with the megapixel (MP) scale at generation time—higher resolution yields more realistic images.
50%UnverifiedOn a single 5070 Ti GPU, generating a 1 MP, 10-second clip using the fast model requires only 8 steps and takes under 6 minutes.
50%UnverifiedMiniMax-H3 handles ordinary motion well, but fast motion is a clear limitation.
50%UnverifiedProducing decent combat footage requires one or two dedicated LoRAs; prompts alone are hard-pressed to elicit high-quality combat shots.
50%UnverifiedMiniMax-H3's handling of faces degrades noticeably in wide shots; FaceRefine can improve this but works best with a single subject and performs worse in multi-person scenes.
50%UnverifiedThe method for controlling expressions is to prepare a reference image with the target expression, add retention_analysis, and explicitly invoke it in the shot description, which substantially improves expression performance.
50%UnverifiedWhen using retention_analysis to maintain environmental continuity, the only flaw is that damage is occasionally copied incorrectly onto a different wall.
50%Unverifiedretention_analysis is a mechanism in the MiniMax-H3 workflow used to instruct the model to preserve the state of specific visual elements from a reference image when generating new shots.
50%