FeiHou Easy H3 Remix加载器

FeiHouEasyH3RemixLoader is the loader you pick when your H3 weights don't split neatly into the FL2VA/REF2VA boxes. The regular Loader treats FL2VA and REF2VA as separate model selectors because that's how the two generation paths usually work - FL2VA for text/image-driven generation, REF2VA for reference-conditioned video. A "Remix" checkpoint is a single transformer that serves both paths at once, so this loader takes exactly one remix_model and points both FL2VA and REF2VA at it. One download, one selector, covers every mode.

It's also the two-pass specialist. Added in v1.3.10, it exists for workflows where you sample twice - a first pass with one LoRA stack and a second pass with a different one - and want the two passes kept cleanly apart.

How it works

Pick the Remix model, the text encoder and both VAEs, and the loader builds a MiniMaxH3Bundle with fl2va_model_name and ref2va_model_name both set to that one file, tagged with a remix_loader flag so the main node knows there's no second file to hunt for. The interesting controls are the LoRA stacks and the second pass:

• first_pass_lora_stack and second_pass_lora_stack (optional) - independent FEIHOU_MERGE_LORA_STACK inputs from the pack's LoRA stack node. Feed different stacks and the two passes stay truly separate; leave the second one empty and the second pass just runs the base model.
• second_sampling_model - optional second-pass model selector, default "none". When set, second-sampling is enabled in the bundle and the main node's second_sampling_model output becomes active. Pick this on the Remix Loader and connect that output on the main node, or the main node errors out - its exact words: "connected a second-sampling model output, but the Remix Loader has no second model selected."

Under the hood the bundle carries second_sampling_use_lora set to true only when you actually supplied a second stack, so an unlinked second pass silently falls back to the base model instead of guessing.

The inputs that matter

• remix_model (required) - the Remix transformer. Refuses to run empty.
• text_encoder - the Qwen3-VL encoder (default qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors), same as the regular Loader.
• video_vae / audio_vae - the two H3 VAEs, minimax_h3_video_vae_fp16 and minimax_h3_audio_vae_fp32.
• second_sampling_model - the second-pass transformer, if you're doing two-stage work.

The output

One h3_bundle, feeding the main node exactly like the regular Loader's. On the main node, remember to connect its second_sampling_model output if you enabled a second pass - that connection is what tells the main node to actually build it.

How to install it

Pack install:

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. Models live in models/diffusion_models (GGUF under unet_gguf), and the H3 weights carry the community license that excludes the US, EU, UK and Korea - worth checking before you pull ~42.5 GB.

Common issues

Most Remix problems are second-pass wiring problems: enable a second model in the loader, forget the second_sampling_model connection on the main node, and get the error above. And if you expected your second pass to inherit the first stack's LoRAs, it won't by default - the two stacks are independent by design, and an empty second stack means base weights. That's the feature, not a bug.