MiniMax H3 RAVEN Guarded Loader / 受保护加载器 (Advanced EXP/T8)

RAVEN Streaming is a separate, MIT-licensed ComfyUI plugin that runs MiniMax H3 as a causal streaming sampler - it's how people squeeze long clips out of the 33B joint AV transformer without blowing up VRAM in one go. This node is the guard that stands in front of it. It preflights the RAVEN plugin, the model, the LoRA, and your memory envelope, and only then delegates to the external loader. If the check fails, it blocks before the weights are loaded. The name is the whole pitch: guarded.

The reason it exists is that H3 is heavy and RAVEN's own reviewed envelope is not casual hardware. The default block_outside_reviewed_envelope mode wants roughly 24 GiB GPU plus 192 GiB host memory to pass. If your box isn't in that neighborhood, the node says no. That's a feature - the README is explicit that the author's own 16GB/128GB machine is below the reviewed resource range and would be refused rather than pretending it can run.

Inputs

• unet_name - must be the full, non-pruned, non-quantized BF16 H3 diffusion model. INT8/FP8/NVFP4/ConvRot checkpoints are explicitly incompatible with RAVEN v0.1. This is where people get caught: you grab a quantized build that works fine for stock sampling, and the guard (correctly) rejects it.
• lora_name - the mandatory RAVEN Streaming LoRA, applied at strength 1.0. Not optional.
• weight_dtype - default / bf16 / fp32.
• enforcement - block_outside_reviewed_envelope (default), block_mechanical_conflicts, or report_only. report_only will load the model even when the envelope fails, and the tooltip flags it as explicitly high risk. Don't reach for it because the default annoyed you.

Outputs: model (for the streaming graph) and report_json with whatever passed or failed.

Installing

Two installs happen here, and the README is clear that this pack does not bundle the external runtime:

cd ComfyUI/custom_nodes
git clone https://github.com/T8mars/comfyui-minimax-h3-audio-T8   # the T8 pack
git clone <RAVEN streaming plugin repo>                             # the external RAVEN plugin (v0.1.0)

T8's node does not copy or execute upstream RAVEN source; it checks that RAVEN is present and then delegates. You also need the full BF16 H3 base plus the RAVEN LoRA on disk. Everything else in this pack installs with no pip deps.

The honest take

This node is for a narrow, expensive corner: long-form streaming H3 on beefy hardware, where a mistake costs you a 24 GB load followed by an OOM. If you're on 16GB trying to run RAVEN, the guard will (correctly) tell you this route isn't yours yet - use stock dual-clock sampling with short segments instead. If you do have the machine, this is exactly the kind of gate you want between you and a model load you can't afford to get wrong. Set enforcement to default, read report_json, and treat report_only as a last resort with a second person watching.