FeiHou Easy H3 Loader

FeiHouEasyH3Loader is the pack's main model loader, and it answers a question the main node deliberately doesn't want to answer: which transformer are we running? MiniMax H3 splits its generation path into two transformers - FL2VA for text-to-image/video and first/last-frame work, REF2VA for reference-conditioned video - plus a Qwen3-VL text encoder and two separate VAEs, one for video latents and one for audio. This node loads whichever of those you've selected, applies your LoRAs internally, and hands the whole pile to the main node as a single MINIMAX_H3_BUNDLE. You get one clean wire instead of a tangle.

How it works

Pick a model in each dropdown and hit run. The loader loads the text encoder, video VAE and audio VAE through ComfyUI's standard loaders, then bundles them with the chosen transformer names, the normalized LoRA stack, and any second-sampling model config into a MiniMaxH3Bundle. The bundle type is how this pack keeps everything on one wire: the main node unpacks it and routes FL2VA or REF2VA depending on your mode. Validation is blunt - if both transformer selectors are "none", it raises "Select at least one MiniMax H3 transformer" before you waste a generation cycle.

The dropdowns scan your real model folders: models/diffusion_models, unet and unet_gguf for the transformers (so GGUF quantizations show up), text_encoders, clip and clip_gguf for the encoder, and the VAE folders for both VAEs. New files appear after a refresh or restart - that stale-dropdown bug was fixed in v1.3.3.

The inputs that matter

• fl2va_model and ref2va_model - the two H3 transformers. You need at least one. Standard pattern: pick FL2VA for image/keyframe work, REF2VA for reference video, or load both so a single loader covers every mode.
• text_encoder - defaults to qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors, a 32B Qwen3-VL variant. This is a big file; it's the thing eating your RAM if you leave it unoffloaded.
• video_vae (minimax_h3_video_vae_fp16.safetensors) and audio_vae (minimax_h3_audio_vae_fp32.safetensors) - yes, H3 really does carry two VAEs. They get packed into the bundle and surfaced again by the Output node.
• lora_stack (optional) - the FEIHOU_MERGE_LORA_STACK wire from the pack's LoRA stack node. The loader applies every enabled LoRA to the transformer inside the loader; the LoRA node never sits in the main node's model chain. That's the whole "bypass, model only" design.
• custom_second_sampling_models plus second_fl2va_model / second_ref2va_model and second_sampling_use_lora - the two-pass setup. Enable it and you can run a second, different transformer for a second sampling stage; the LoRA toggle decides whether that second pass inherits the stack.

The output

One h3_bundle → the main node's h3_bundle input. That's the only output. Everything the main node needs - models, names, VAEs, LoRA stacks, second-sampling state - rides along inside it.

How to install it

Install the pack, not the node:

cd ComfyUI/custom_nodes
git clone https://github.com/FX-FeiHou/ComfyUI-FeiHou-Easy-H3

restart, or ComfyUI Manager → search "ComfyUI-FeiHou-Easy-H3". No pip dependencies. You need a ComfyUI with official MiniMax H3 support and the models in the folders above. One honest warning: H3's full-precision weights are ~42.5 GB and the license geofences them out of the US, EU, UK and Korea - check the H3 Community License before downloading.

Common issues

Most loader pain is model-folder pain. A model doesn't show in a dropdown → it's in the wrong folder or ComfyUI hasn't rescanned. Selecting "none" for both transformers → instant error, so pick at least one. And if you enabled custom_second_sampling_models but left both second-pass selectors at none, the main node will fail when it tries to build the second pass - either pick a model or turn the toggle off.