MiniMax H3-World Model Composer / 模型组合 (T8)

If you've loaded H3-World workflows, this node is doing more than the name suggests. On the surface it's a LoRA loader for the H3-World control LoRA - swap it in wherever you'd normally put a LoraLoaderModelOnly. But it also installs two things a stock LoRA loader can't: the per-action text refiner segmentation and the directed FlexAttention runtime that make the per-latent action control work at all. In other words, this node is the reason the whole H3-World trick functions.

What you're actually loading

The workflow needs the model at ComfyUI/models/loras/minimax/H3-World/step-10000.safetensors. Easiest path is the pre-organized pack from the author's HF repo:

hf download t8star/Minimax-H3-World-Comfy --include "loras/**" --local-dir ComfyUI/models

That drops the LoRA in the right relative folder. It's already ComfyUI-loadable 104-pair A/B weights, byte-identical to the original DANNY621/H3-World release at a pinned revision - nothing converted, merged or quantized. If you grab the original from elsewhere, get it into that exact folder or just pick it from the node's lora_name dropdown instead.

Inputs

• model - a native ComfyUI MiniMax H3 FL2VA model, straight from your diffusion model loader. Don't pass an already-OpenVDN'd or already-World'd model; this node wants the clean base.
• lora_name - the dropdown, defaulting to the H3-World LoRA.
• strength_model - default 1.0. This is the trained-strength control LoRA; resist cranking it to 1.5 to "make it more obvious." The node's report will tell you if something's off.
• compile_flex_attention - default true, and the tooltip's "Recommended" is right. The directed attention that keeps action sentence N glued to latent N is a custom FlexAttention pass, and the first run compiles one shape-specific kernel. Keep it on; the one-time compile hiccup is worth the speed on every frame after.

Why the plain LoRA loader isn't enough

The H3-World LoRA shapes the model, but the per-latent action binding lives in how text is segmented and attended to. Each action sentence has to be refined and steered toward exactly one latent region - otherwise "camera pans left" at latent 5 bleeds into the whole clip and you're back to regular H3 prompting. This node composes the LoRA and installs that segmentation + directed-attention runtime on top of the model, then hands the prepared model out.

Wiring

You'll use it exactly once per chain, right after loading your H3 base model:

Model loader → MiniMaxH3WorldModelComposerT8Advanced → sampler (with conditioning from MiniMaxH3WorldI2VAConditioningT8Advanced)

Follow the examples/workflows/26-h3-world example and don't improvise: the README is explicit that you should not layer OpenVDN, SLA, VSA, Sol-Attn, BlockCache or any other model/attention takeover node onto the H3-World chain. Those all fight over the same branches this runtime just claimed. This is the "more speed nodes = faster" trap wearing its worst disguise - it usually produces a conflict error or quietly wrong output, not a speedup.

Install and expectations

Same shared install as every node in the pack: ComfyUI Manager (search "MiniMax H3 Audio T8") or clone into custom_nodes, restart. No new pip packages. You need a recent ComfyUI with native H3 support - if the node shows red, update ComfyUI, the frontend and Manager first, then do a full exit and relaunch.

Then set expectations: H3-World is one fixed contract (832×480, 124 frames, first-frame I2VA). It passed an anonymous blind test on a fixed parking-garage sample - stable motion, matched quality - but that's one contract, not a promise for arbitrary characters, actions, or VRAM sizes. MiniMax H3 itself is a 33B model that's genuinely heavy, so on a 16GB card run this chain alone and nothing else at the same time.