FeiHou Easy H3 Model Adapter

The Loader gives you a tidy all-in-one way to load MiniMax H3, but not everyone wants it. Maybe you already have a standard ComfyUI model chain - an Unet/DiffusionModel loader, a CLIP loader, a VAE loader - and you'd rather not maintain a second, parallel way of loading the same files. FeiHouEasyH3ModelAdapter is the bridge for that: it takes ordinary ComfyUI MODEL, CLIP and VAE objects and assembles them into the same MINIMAX_H3_BUNDLE the Loader produces, so the main Easy H3 node never knows the difference.

How it works

The adapter is a pure passthrough wrapper. It requires three standard wires - a CLIP for the text encoder, a VAE for the video VAE, and a VAE for the audio VAE - plus at least one of the two H3 transformers (fl2va_model, ref2va_model) as MODEL inputs. Give it none and it refuses: "Connect at least one transformer MODEL." It then wraps those objects into a MiniMaxH3Bundle marked as "connected" (names become placeholders since the objects came from elsewhere) and hands it to the main node.

Two things make it behave differently from the Loader. First, IS_CHANGED returns NaN, so the adapter re-executes every run - which is what you want, since it has no idea whether the models feeding it changed. Second, there's no LoRA stack input. LoRAs get applied upstream, in whatever chain you already built, before the MODEL reaches this node. If you're an "apply LoRAs with my usual nodes" person, that's a feature; if you live on the pack's stack node, use the Loader instead.

The inputs

• text_encoder (required, CLIP) - your Qwen3-VL encoder, from any CLIP loader.
• video_vae / audio_vae (required, VAE) - both H3 VAEs. Yes, both; the bundle expects them and the Output node will surface them again.
• fl2va_model / ref2va_model (optional, MODEL) - at least one required. Both for a single-adapter setup that covers every mode.

The output

h3_bundle → the main node's h3_bundle input. Same wire, same type, no conversion on the other end.

How to install it

Same pack, same steps:

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. The models themselves are the standard H3 set - the ~42.5 GB full-precision weights, or whatever quantization your loaders support - and remember the H3 Community License geofences them out of the US, EU, UK and Korea.

Common issues

The classic failure is wiring it wrong on purpose: connecting only the two VAEs and a CLIP, then wondering why the main node errors - the adapter throws if neither transformer MODEL is attached. The subtler one is forgetting that this node re-runs every time, so if the models feeding it are slow to load, the adapter will look like it's recomputing forever on every queue click. That's expected, not a bug. And if your upstream loaders don't give you an audio VAE at all, the pack's own Loader is the easier door - the adapter genuinely requires you to already have both VAEs handy.