Minimax H3 hybrid loader

MiniMax H3, the open-weight video model MiniMax dropped at the end of July 2026, ships as two checkpoints that force an awkward choice. FL2VA gives you the strong image generation. Ref2VA gives you reference-driven shot organization. And both are huge - roughly 19.5 GB for a pruned int8-convrot checkpoint, 31.7 GB for the full one. You are not loading two of them on a 32 GB machine. MinimaxH3_HybridLoader is the community answer: load the FL2VA, and fold a small, validated slice of the Ref2VA on top so you stop having to pick.

This is a fresh pack for a fresh model - H3 only went open-weight in July 2026, and this loader is one of the first tools built around it. It's worth knowing about even if you've never touched a hybrid loader, because the reasoning behind it is the interesting part.

What it actually does

The pack's whitepaper makes a claim that flips the usual "hybrid" logic on its head: FL2VA already contains a complete reference-conditioning path. ComfyUI's own MiniMaxH3ReferenceToVideo runs the identical code path for both checkpoint variants, and the reference-token components (condition_proj, token_refiner, the AdaLN table) are tensor-for-tensor nearly identical between the two. So the overlay isn't "buying reference capability" - it's modulation fine-tuning. This matters because an earlier community hybrid loader (scottmudge's) overlaid Ref2VA's AdaLN weights over blocks 25–49 based on weight-space analysis that was never validated against actual video, and that full range accumulates structural artifacts.

Hence the three presets, which are the whole node:

• none - pure FL2VA load. The Ref2VA checkpoint is never even opened, so memory and speed match a plain single-model load.
• block_range_adaln (default) - overlays adaln_proj for blocks 45–49 only. This is the production trade-off the author converged on after video-validating every range.
• all_adaln - overlays every block's AdaLN plus the final AdaLN and the format-appropriate time embedding. Experimental, reference-first.

In every hybrid preset the output heads stay FL2VA. The loader also auto-detects the pruned (adaln_t_table) vs full (time_embedder.*) checkpoint format, rejects mixed pairs, validates key sets, metadata, and the shape/dtype of every tensor it overlays, and keeps quantization sibling keys (.comfy_quant, weight_scale, …) on the same side as their owning weight. That last one is the kind of detail that silently breaks a model if you get it wrong.

The inputs and outputs

Only three inputs, and you'll set all of them:

• base_model - an FL2VA checkpoint. The dropdown is filtered to files with "fl2va" in the name.
• overlay_model - an optional Ref2VA checkpoint (filtered on "ref2va"). Ignored in none mode, but required for the other two presets or the node errors out.
• overlay_preset - the three-way choice above.

The single output is a MODEL, which plugs into the same sampler chain as any video model you've already got wired up. Nothing exotic downstream.

Installing it

ComfyUI Manager can find it if you search the pack title, "ComfyUI MiniMax H3 Hybrid". Otherwise:

cd ComfyUI/custom_nodes
git clone https://github.com/ANe5s/ComfyUI-MiniMax-H3-Hybrid

Restart ComfyUI. That's it - the plugin has zero Python dependencies beyond what ComfyUI bundles, which is refreshing. What you do need is the H3 checkpoints themselves, sitting in ComfyUI/models/diffusion_models. The node picks sensible defaults: whichever "fl2va" file has "pruned" in the name becomes the base default, same for "ref2va".

Where people get burned

• Your file doesn't show up in the dropdown. The node filters on the filename containing "fl2va" or "ref2va". Rename the file to match, or it's invisible.
• Mixed formats. Base and overlay must both be pruned or both full. The loader refuses to guess, and rightly so - the two layouts compute modulation differently, and a mixed pair is a semantically wrong model.
• The RAM guard. With DynamicVRAM/AIMDO off, a hybrid load needs free system RAM ≥ the base checkpoint size plus a GiB. That's the error message's hint to use the pruned pair or enable DynamicVRAM, not a bug.
• Expecting a magic reference upgrade. The honest baseline is none plus ComfyUI's reference-conditioning node - the whitepaper's own video tests show that's the most stable lossless setup, and the hybrid presets are refinements, not a rewrite of H3.

MiniMax H3 is new and it's heavy; expect slow generation even on a 3060, which the community confirms runs it but without any speed. For the choice between FL2VA's sharpness and Ref2VA's shot composition, though, this loader is the cleanest resolution on the board - one node, three presets, and a genuinely readable whitepaper explaining why each exists.