Flux Text LoRA Loader

Flux is great at a lot of things; spelling is not one of them. Rendering legible text in Flux images is the classic weak spot, and the community answer is dedicated text-rendering LoRAs trained on Flux - which don't load like ordinary LoRAs, because the good ones condition on glyphs (the actual shapes of the characters), not just prompt words. FluxTextLoraLoader is the loader that knows how to apply them: it loads the LoRA from your models/loras folder and patches Flux's transformer so glyph conditioning actually flows through.

If you've ever stared at a Flux "sign in the image" that reads as near-nonsense, this is the node that makes that sign render correctly.

How it works

Under the hood it's a two-part job. First, the LoRA file is loaded and converted (comfy.lora_convert), then installed onto the Flux model's linear layers with your strength_model. Second - and this is what a normal LoraLoader won't do - it monkeypatches the Flux transformer's double- and single-stream blocks so the glyph/text tokens get their own attention pathway. The LoRA's weights are applied only inside a gated context, which is why a naive loader produces a LoRA that "loads fine but does nothing."

It's Flux-only, and it checks: if you feed it any non-Flux model it raises an error. (The error message actually still says "PhotoDoddle" - a leftover from the author's other model, and a good reminder that this is a personal utility pack.)

Inputs and outputs

• model - your Flux model (from a checkpoint or model loader).
• lora_name - the LoRA picked from your models/loras folder, as a dropdown.
• strength_model - how hard the LoRA applies (default 1.0).

Output is the patched MODEL, wired straight into your sampler. Note there's no CLIP output here - this LoRA path patches the diffusion model, not the text encoder.

Installing it

It's part of Duanyll Nodepack:

cd ComfyUI/custom_nodes
git clone https://github.com/Duanyll/duanyll_nodepack

or via ComfyUI Manager. The LoRA itself is a download you make separately - put it in ComfyUI/models/loras like any other LoRA.

Where people get burned

• Only Flux, period. The guard raises a hard error on non-Flux models, so you can't silently misapply it. That's the good kind of failure.
• It expects a compatible text LoRA. Not every "text" LoRA is glyph-conditioned, and this loader's special handling targets the ones that are. A generic style LoRA loaded this way may behave differently than through LoraLoader.
• Strength is real but not a dial for text quality. Crank strength_model too high and you'll get the LoRA's artifacts, not better spelling. Start at 1.0 and only nudge it.
• Spelling is still stochastic. Even with a text LoRA and the right loader, Flux will occasionally misspell; this fixes the mechanism, not the universe. Keep prompts short and explicit about the text you want.

The word "text" here means text in the image, not the prompt. If you're just looking to load an ordinary character or style LoRA on Flux, the built-in LoraLoader is still the right tool - reach for this one specifically when the job is legible words.