沐阳 H3 加载器

MiniMax H3 doesn't need five separate loader nodes to be usable - it needs five model slots wired up correctly, and that's exactly the job 沐阳 H3 加载器 (H3Loader) exists for. It's the entry point of basically every workflow in this pack, including the bundled Director and LongVideo examples. You pick your models once, out pops a single MYANG_H3 bundle, and every other Myang node reads from it. If you've ever fumbled your way through three different checkpoints and two VAEs in one image workflow, this is the video-omni version of that, done in one place.

The clever bit is that it's lazy. H3 actually runs two diffusion models: ref2va (reference-to-video, for motion transfer and video continuation) and fl2va (text-to-video and first/last-frame jobs). The loader registers both but only loads the one the current job actually calls - so a pure motion-transfer run doesn't eat VRAM loading the fl2va checkpoint and vice versa. Under the hood it's just delegating to ComfyUI's own CLIPLoader (with the minimax_h3 type) and VAELoader, which is why this pack ships with zero Python dependencies beyond the stock ComfyUI environment.

The inputs that matter

• ref2va_model - the reference-to-video checkpoint. Motion transfer and video continuation use this one.
• fl2va_model - the text-to-video / first-last-frame checkpoint.
• text_encoder - the Qwen encoder, loaded as a minimax_h3 CLIP.
• video_vae and audio_vae - H3 splits picture and sound into separate VAEs, so you genuinely need both. Missing the audio VAE is the classic first-run failure.
• weight_dtype - default, or fp8_e4m3fn / fp8_e4m3fn_fast / fp8_e5m2. The full weights are big (~42.5GB reported for H3), so if the dropdown on your card is VRAM-anxious, fp8 is the lever you pull first.

Output is one h3 (MYANG_H3) socket. That feeds 沐阳 H3 取模型 (to pull a MODEL out for patch chains), 沐阳 H3 条件 (to build conditioning and the empty latent), or straight into 沐阳 H3 · 长视频 / 导演台.

Install and models

Install via ComfyUI Manager (search "ComfyUI-MiniMaxH3-Myang"), or manually:

cd ComfyUI\custom_nodes
git clone https://github.com/civilcoco/ComfyUI-MiniMaxH3-Myang.git

Restart ComfyUI, hard-refresh the browser if the old UI lingers. The nodes appear under the 沐阳 H3 category. The pack does not distribute weights - you supply the H3 diffusion model, the Qwen text encoder, and both VAEs yourself, all the usual places in ComfyUI/models.

Where people get burned

Forgetting the audio VAE, and picking the wrong model for the task - put your motion-transfer checkpoint in ref2va, your text-to-video in fl2va, and make sure any 沐阳 H3 取模型 node's kind matches the task you actually run. And a license heads-up that isn't this pack's fault: H3's Community License geofences the local weights out of the US, EU, UK and South Korea, so check LEGAL.md before you download or publish anything.