H3RelayInternalHybridLoader

If you've read the H3 Relay Hybrid Model Loader article, you already know the trick: load a MiniMax H3 checkpoint, merge selected tensor groups from an overlay checkpoint onto it, and end up with a model that has fl2va's quality and ref2va's reference-conditioning pathway. This internal node is the same loader - literally the same class from the vendored MiniMax H3 Hybrid Loader - registered under a namespaced id, with one crucial difference: it returns a plain MODEL, not an H3_RELAY_MODEL bundle.

That single difference is the entire reason this node exists. The public Hybrid Loader wraps its output in H3 Relay's cache-aware bundle so the pack can fingerprint model changes and invalidate derived artifacts. This one skips the wrapper and hands you a stock MODEL socket, which makes it useful in exactly one scenario: you're building a hand-rolled H3 graph with native ComfyUI nodes - stock samplers, conditioning, the works - and you want the hybrid merge without committing to H3 Relay's bundle-and-cache contract.

The inputs

Everything matches the public hybrid loader's schema minus the manual_cache_revision field (there's no cache to bump, since the output is raw):

• base_model - primary checkpoint; every tensor starts here.
• overlay_model - secondary checkpoint; tensors matched by overlay_preset come from here. It may equal base_model (effectively a base-only load).
• overlay_preset - none, ref2va_adaln_over_fl2va (default), ref2va_all_adaln_over_fl2va, the two full-overlay sanity presets, block_range_adaln, or custom. Same semantics as the public node: none is a pure base load equivalent to a stock UNETLoader.
• block_range_start / block_range_end - used only with block_range_adaln; the H3 DiT has 50 blocks indexed 0..49. block_range_end < block_range_start takes nothing from the overlay.
• final_adaln_from_overlay - pulls final_layer.adaln_proj from the overlay on top of whatever the preset does.
• custom_overlays / custom_base - comma-separated keys, prefixes, or fnmatch globs to add or force back.
• weight_dtype - default or fp8 variants, same meaning as stock Load Diffusion Model.

Output is a single MODEL socket. Wire it wherever a native ComfyUI model would go.

When to use it (and when not to)

Use it if you're building the stock-style MiniMax H3 continuation graph by hand and want the hybrid merge as the model source. Don't use it inside a reference H3 Relay workflow - the staged pipeline and the cache contract both assume the bundle form, and a raw MODEL won't carry the H3_RELAY_MODEL type the relay nodes expect. If you're not hand-building, the public H3 Relay · H3 Hybrid Model Loader is the right node, full stop.

This is the classic internal-node trap: it looks like a drop-in alternative and is, but only if you know what you're doing with the surrounding graph. If you found it in the expanded graph of a reference workflow, it's because the staged H3 generation path uses it internally to build the H3-only stage - leave it be.

Install the pack via ComfyUI Manager (search H3 Relay) or git clone https://github.com/akatz-ai/h3-relay.git into custom_nodes. ComfyUI 0.32.0+, FFmpeg, and both H3 checkpoints from MODELS.md are required; the MiniMax H3 community license geofence (US/EU/UK/Korea) applies regardless of which loader you pick.