The BEST local AI video generator just got BETTER!

MHMiniMax H3MiniMaxUnverified

Open-source video generator released the previous week; the video covers faster and lower-VRAM ways to run it in ComfyUI, prompting guides, and using it for audio and image generation.

COComfyUI-Spectrum-MiniMax-H3Available in EU

Spectrum speed-up node, updated to a newer version with lower quality loss. Works with res multistep but not with the ERSD sampler.

Tiny VAE (under 10 MB) used with the KJ nodes model preview override node to preview video generation live and cancel early if it looks bad.

MIMiniMax-H3-Realism-People-LoRAfalAvailable in EU

131 MB LoRA that makes videos look more realistic and amateur. Uses the trigger word 'realism'; demoed in a text-to-video cafe example.

MHMiniMax H3 Turbo LoRAsLightX2VAvailable in EU

Turbo LoRAs from Larry VRH, LightX2V, Joy Fox and Kijai cut steps from 20 to about 4, up to 5x faster. Demoed Kijai's resized 315 MB version at 0.75 strength with the ERSD sampler.

Compressed W4A8 models of about 12 GB, plus a smaller 3 GB int8 video VAE. With Spectrum and sage attention they could run on about 8 GB VRAM; demoed with an orbit shot of a car.

S2Seedance 2.5ByteDanceUnverified

Sponsor segment on Higgsfield: up to 30 seconds per pass with multiple shots and audio, up to 50 references, and finer editing control.

COComfyUI-MiniMaxH3-Easynkxx188Available in EU

A simple single ComfyUI workflow covering text, image, first/last frame and reference to video, with a clean resolution selector and @-references in prompts.

MHMiniMax H3 audio generationMiniMaxUnverified

Bypassing the video steps and saving only the audio lets H3 generate music and sound effects from a prompt. Demoed with orchestral battle music and a screaming voice over rock music.

Community ComfyUI workflow using H3 with a frame count of one to produce images. Demoed with an anime fox-eared shrine maiden; the creator says it isn't a frontier image model and prefers other models.

MHMiniMax H3 GGUFsUnslothAvailable in EU

GGUF quantizations from Abiray (Q3 15.6 GB) and Unsloth (Q2 6.7 GB). The creator says ComfyUI now recommends quantized native versions over GGUFs for better quality.