Load Ideogram4 V9 TE Module-LoRA

Here's the piece that makes Ideogram4 V9 triggers unlike anything you've used. Gen2_LoadIdeogram4V9TriggerTEAdapter loads exactly 36 independent rank-4 mlp.down_proj module-LoRA pairs - tiny LoRAs that hook into the Qwen3-VL text encoder's 36 layers, not into the diffusion model. The idea is that a trigger token isn't just a vector to inject; it's also a behavior in the encoder. When the token fires, all 36 hooks fire with it, and the encoder produces conditioning that actually matches whatever was trained.

Why a text-encoder LoRA at all

Your text encoder is the model that turns your prompt into the vectors the diffusion model attends to. On modern stacks that encoder is a real LLM - Qwen3-VL here - and the community discovered you can shape how it encodes a token by LoRA-ing the encoder, not the denoiser. That's what this artifact is: a set of module-LoRAs scoped to the encoder's mlp.down_proj weights, one per layer, all rank 4. The pack is strict about the shape - it will refuse anything that isn't exactly 36 hookable pairs. It also checks a compatibility fingerprint, and the composer node refuses to pair an embedding and a TE adapter that were trained against different fingerprints. Annoying if you mix artifacts from different sources, but it's the difference between a trigger that works and one that silently half-works.

This is also the node that trips people up because of its name. It is not a general-purpose LoRA loader. The pack has a separate node - Gen2_Ideogram4AITKLoRALoader - for ordinary joint Ideogram4 LoRAs. This one only understands V9 trigger TE artifacts, and if you feed it a normal LoRA it will fail cleanly.

The one input

• artifact_name - dropdown from ComfyUI/models/gen2/. Same folder as the embedding loader, so keep both halves of a trigger together and clearly named (they're usually distributed as a pair).

Output: te_adapter, which - like the embedding - doesn't go to a text encoder directly. It feeds Gen2_ComposeIdeogram4V9TriggerActivator, which binds it to the matching embedding and produces the activator the text-encode node actually consumes.

Install

ComfyUI Manager: search ComfyUI-gen2. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/petmycat/ComfyUI-gen2.git
cd ComfyUI-gen2
pip install -r requirements.txt

Restart, place the artifact in ComfyUI/models/gen2/, and you're set. Requires a native Ideogram4 + Qwen3-VL backend - the pack fails closed if it can't confirm it, and it explicitly won't run on Flux/Klein's Qwen3-8B even though the layer counts look similar. And remember the Ideogram 4 weights are non-commercial with a license that even reaches your generated images, so know what you're allowed to ship.