MiniMax Ref Hybrid Loader

Here's the situation the H3 local scene keeps running into: one checkpoint, the fl2va lineage, generates noticeably nicer output. Another, ref2va, is the one with the proper reference-conditioning pathway - the one that actually follows your reference image, video, and audio. You want fl2va's quality and ref2va's references, and MiniMaxRefHybridLoader is the blunt instrument for getting both: it loads one checkpoint as the base and surgically overlays selected tensor groups from a second one on top.

The mechanism is honest about what it's doing. MiniMax H3's DiT is 50 blocks (indexed 0..49), and the loader only bothers touching the adaln_proj tensors - the adaptive-layer-norm projection weights that, in a reference-conditioning model, are exactly where the "condition this block on what it sees" signal lives. So the default preset, ref2va_adaln_over_fl2va, runs a fl2va base and pulls each block's adaln_proj from ref2va: quality from one, reference-following from the other. That's the whole pitch, and for a reference-to-video workflow it's a genuinely useful trick rather than a gimmick.

The three required inputs are checkpoints, checkpoints, and intent. base_model is where every tensor starts; overlay_model is where the preset-matched tensors come from (you can set it equal to base, which is just "load base only"). overlay_preset is the real control:

• none - pure base load, equivalent to the stock Load Diffusion Model node
• ref2va_adaln_over_fl2va (default) - per-block adaln_proj from the overlay
• ref2va_all_adaln_over_fl2va - also takes final_layer.adaln_proj
• ref2va_full_over_fl2va / fl2va_full_over_ref2va - everything from the overlay, which the tooltip calls out as a sanity check rather than a use case
• block_range_adaln - only blocks in [block_range_start, block_range_end], for when you want to experiment with where the reference conditioning lives
• custom - you hand-write the tensor selection

Both checkpoints must live in ComfyUI/models/diffusion_models. The custom tier is where this gets deep: custom_overlays and custom_base take comma-separated keys, prefixes, or globs - blocks.49. for a prefix, blocks.[0-4].*.attn.qkv_proj.weight as a glob - to force tensors over from the overlay or drag them back to the base. final_adaln_from_overlay is a boolean additive toggle for pulling final_layer.adaln_proj regardless of preset. Unless you have a specific hypothesis about a block range, the default preset plus maybe that final-adaln toggle is where you'll live.

weight_dtype mirrors the stock loader (default, fp8_e4m3fn, fp8_e4m3fn_fast, fp8_e5m2) - leave it unless you know why you're touching it. Output is a single model, wired into your sampler chain. It's categorized under model/loaders and is the one node in this pack that's independently useful outside the director workflow.

Be honest about what this is: a checkpoint-merging experiment, not a scientific guarantee. There's a companion analysis file in the repo describing what differs between fl2va and ref2va, but the "best" blend is empirical - log what you load (the node prints the combo to the console), render the same shot under a couple of presets, and keep the one that wins. Installation is the pack's shared story: ComfyUI Manager or clone https://github.com/eaglering/MiniMaxRefDirector-ComfyUI into custom_nodes, pip install -r requirements.txt, restart. And a reminder that H3's local weights sit under a community licence that excludes the US, EU, UK, and Korea - worth knowing before you download a second ~42 GB checkpoint to merge.