H3 Relay · H3 Model Loader

If you're new to H3 Relay, this is the loader to start with. It takes one MiniMax H3 diffusion checkpoint from your models/diffusion_models folder and hands you a ready-to-use H3_RELAY_MODEL bundle. That's the entire job. The Hybrid Loader exists for people who want to merge fl2va and ref2va; this one just loads whichever checkpoint you pick, and for most first projects that's exactly the right amount of loader.

Mechanically, it's a thin wrapper: the node builds a stock UNETLoader plus an internal bundle-pack node via ComfyUI's GraphBuilder, wraps the loaded model with a cache tag that fingerprints the filename, dtype, and manual revision, and returns the bundle. So behind the scenes you get exactly the loading path you'd get from native ComfyUI - no custom model format, no extra code to trust.

The three settings

• model_name - the dropdown of everything in diffusion_models. For H3 you'll want either minimax_h3_fl2va_int8_convrot.safetensors or minimax_h3_ref2va_int8_convrot.safetensors from MODELS.md.
• weight_dtype - default, or fp8 variants (fp8_e4m3fn, fp8_e4m3fn_fast, fp8_e5m2). Same meaning as the stock Load Diffusion Model node. On a 4090-class card, the int8 convrot H3 files are already the practical format; leave default unless you're chasing VRAM.
• manual_cache_revision - defaults to v1. This is not a version of the model; it's a manual cache reset. The tooltip is exact: change it when replacing model contents without changing the filename. Normal model, LoRA, dtype, strength, and attention changes are tracked automatically by the cache fingerprint. If you've never replaced a checkpoint in place, you'll never touch it.

Output is a single h3_model wire. From there the bundle usually goes through H3 Relay · Attention Backend (if you want an optimized attention kernel) and then to Generate Shot. The bundle is also what a H3 Relay · Apply Model LoRA node would consume.

Why the wrapper is worth it

You could load H3 with a stock UNETLoader and build the continuation graph yourself - the context-loop ecosystem documents that path. What the wrapper buys you is the cache contract: the bundle carries a fingerprint that changes whenever your model chain changes, so a model swap invalidates only the derived artifacts (LTX finishes, interpolations) that actually depend on it. Your accepted raw shots, keyed to the model revision that made them, stay valid. That's the resumability story of the whole pack, and the loader is where it starts.

One honest caveat: this loader gives you the diffusion model only. The H3 text encoder and video/audio VAEs are separate files you load elsewhere in the graph (the reference workflow wires them the same way the stock MiniMax H3 template does). Don't be surprised when the bundle doesn't include a CLIP - that's by design.

Install via ComfyUI Manager (search H3 Relay) or git clone https://github.com/akatz-ai/h3-relay.git into custom_nodes, then restart. ComfyUI 0.32.0+ and FFmpeg on PATH are required; model files come from MODELS.md, and the pack never redistributes weights. And the usual flag: the MiniMax H3 community license geofences out the US, EU, UK, and Korea, so verify you're allowed to run it before downloading.