Ideogram4 AI Toolkit LoRA Loader

If you've trained an Ideogram4 LoRA with Ostris's AI Toolkit - and if you trained one at all, you probably did, since it's the trainer that ships support for each new architecture within days - you've hit the moment where the LoRA loads fine into ComfyUI but the text-encoder half silently doesn't apply. That's what this node fixes. Gen2_Ideogram4AITKLoRALoader is a joint MODEL + CLIP LoRA loader that adds adapter-base aliases for AI-Toolkit's lora_te.language_model.* Qwen text-encoder keys, then hands off to stock ComfyUI loading. Same load, but the CLIP half finally lands.

What's actually broken without it

Ideogram4's text encoder is a Qwen3-VL model, and AI-Toolkit saves its text-encoder LoRA weights under keys that stock ComfyUI doesn't recognize - lora_te.language_model.* paths that don't map to the encoder structure ComfyUI expects. The result: your LoRA's diffusion half applies, your LoRA's prompt-encoding half doesn't, and the trained style or character comes out half-baked with no obvious error. This loader detects those AI-Toolkit TE tensors, builds the aliases that remap them onto the connected CLIP, and only then calls the stock loader. The LoRA tensors themselves pass through untouched - it's a naming-compatibility shim, not a re-quantizer or converter.

It's a joint loader, so unlike a model-only loader it needs both:

• model - the Ideogram4 diffusion model to patch.
• clip - the Ideogram4 Qwen3-VL text encoder to patch.
• lora_name - dropdown from ComfyUI/models/loras/. Works for both AI-Toolkit and standard ComfyUI-format LoRAs.
• strength_model, strength_clip - separate weights for the two halves, both default 1.0. Setting both to 0 is a valid bypass that returns the inputs untouched.

Outputs: model and clip, both patched, ready to feed your sampler and text encoder.

Two honest warnings. First, this is not the loader for V9 trigger TE artifacts - those are a different, strictly-shaped format that must use the V9 nodes. Second, the loader fails loudly if it can't match all the AI-Toolkit Qwen TE keys to your CLIP (e.g. you connected the wrong encoder, or ComfyUI restructured things). Loud is good here: a half-applied LoRA is a silent disaster, an error you can fix.

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, put your Ideogram4 LoRA in ComfyUI/models/loras/. No extra Python deps for this node. As with anything on the local Ideogram4 weights: they're non-commercial under a license that even reaches generated images, so know your terms before you ship anything built with them.