From September 13 to 17, 2026, ten new tools and LoRAs were released around the MiniMax H3 video model in ComfyUI. The week's key events include the ComfyUI 0.36.0 release with a fixed Fun ControlNet and official Fast H3 templates, the ClipProj-MiniMax-H3 v3.1 project, which reduces the text encoder's VRAM consumption from 15.7 to 4.5 GB, and the official int8 quantization of the video VAE by Comfy-Org. Everything mentioned is available for free today, and the barrier to entry for local video production is shifting from flagship cards to consumer ones.

image

What Happened

The channel "Neuronaut" released the 25th part of its digest on #minimaxH3 — a review of ten new tools and LoRAs for the MiniMax H3 video model in ComfyUI; almost all of them were released between September 13–17, 2026. The central event is the ComfyUI 0.36.0 release on September 15: PR #16240 fixes the operation of Fun ControlNet for H3 together with Comfy Compiler, the Video Concatenate node appeared (PR #16267), H3 VAE optimizations by kijai were included (PR #16187), Marigold V2 support (PR #16232), and official workflow templates. Separately, Comfy-Org published an int8 version of the video VAE under the name minimax_h3_video_vae_int8_convrot.safetensors. The ClipProj-MiniMax-H3 v3.1 project by jhong520 replaces the Qwen3-VL-32B text encoder with a small Qwen3-VL via projection matrices: the mmh3-4b-ClipProj-v3.1 file weighing 26 MB reduces VRAM consumption from 15.7 to 4.5 GB, requires a custom node version 0.1.13 or newer, and, according to the author's description, improves the recognition of non-English speech. The top ten is completed by LoRAs: FaceSwap_MiniMaxH3_REF2VA with the faceswap trigger for replacing a face in a reference video with a face from a picture (CRT-Nodes required), Retro Toon 70s for 1970s cel animation, 16bit Pixel by KennethFal for pixel art in the style of SNES cutscenes, and Minimax Ghost v0.1 on Civitai.

Context

To assess the significance of the week, it is important to understand how the local H3 pipeline is structured. The Qwen3-VL-32B text encoder was one of its heaviest components: before the appearance of projection replacements, it required about 15.7 GB of video memory and effectively excluded most consumer cards even before the start of generation. ClipProj v3.1 solves the problem not by reducing the main model, but by replacing the large encoder with a small one via trainable projection matrices — a technique that, if confirmed, can be easily transferred to other "model plus encoder" combinations. The maturity of the ecosystem is also seen in who releases the key artifacts: the int8 quantization of the video VAE was done by the Comfy-Org infrastructure maintainer themselves, not a third-party enthusiast, which increases trust in the file. An applied layer is also growing around the base model: the ExactAudioLock node turns audio quality control into a formal human approval point within automatic pipelines, float32 decoding with output to EXR meets the needs of colorists, and timeline export to DaVinci and FCP integrates generation into a familiar editing loop.

Why This Matters for the Industry

The main change of the week for the industry is the shift of local video production on H3 from flagship accelerators to consumer cards. An encoder at 4.5 GB instead of 15.7 GB, the official int8-VAE, and VAE optimizations in ComfyUI 0.36.0 collectively reshape hardware requirements: a video service prototype can be built on a single consumer card without cloud bills, and for existing teams this is a direct recalculation of the cost price and the ability to expand the audience to owners of weak GPUs. The composition of the releases is also indicative: of the ten tools of the week, almost all are adaptation and stylization, face swap, retro style, editing, and not new capabilities of the model itself. Such a profile of efforts is characteristic of an established base model around which a production ecosystem is maturing. If the projection replacement of the encoder withstands independent checks, similar techniques are likely for other video models, and the official quantization from the maintainer sets a precedent for future open releases.

Why This Matters for Users

Practically everything from the digest can be tried for free today. The first step is to update ComfyUI to 0.36.0: the portable build has already been posted on GitHub, and the release comes with ready-made official Fast H3 templates, which are convenient for starting without building a pipeline from scratch. LoRAs can be downloaded from Hugging Face or Civitai. If there is little video memory, it is worth installing ClipProj v3.1 and int8-VAE first — the encoder will take 4.5 GB instead of 15.7 GB, and decoding will become lighter and faster compared to fp16. Colorists will find float32 decoding with output to EXR useful, editors — timeline export directly to DaVinci or FCP, and for quick visual experiments — stylizations for 1970s cel animation and pixel art in the style of SNES cutscenes.

What Is Still Unknown / Limitations

Key caveats. The digest contains no cost measurements per minute of video and no benchmarks: the int8-VAE characteristic "lighter and faster than fp16 with close quality" is not accompanied by numbers, so before a production launch, your own measurements are needed. The VRAM reduction in ClipProj v3.1 and the improvement in non-English speech recognition are declared by the project author and have not yet been confirmed by independent checks — they should be rechecked on your card and your scenes. A dense week of releases proves the activity of the community and the maturity of the ecosystem, but not the growth of the capabilities of the MiniMax H3 model itself — these are different statements. Finally, the widely available face swap on open weights sharpens the questions of ethics and the origin of content, and whether H3 will maintain its position against the backdrop of future video models is unknown.

Sources

Author

Look at AI, editorial team