Load Iris LoRA

This is the adapter slot for Iris-3B, and the first thing to know is that it's not a universal LoRA loader. Iris is a pixel-space DiT with its own layer naming, so an SDXL or Flux LoRA will not do anything useful here - and the pack is honest about that, but only after you've picked one.

The dropdown is a trap, briefly

The lora_name list is everything in models/loras, Iris or not. That's how it has to be, since ComfyUI's LoRA folder is shared. What saves you is that the loader validates on the way through: it opens the file, looks for the rank and alpha metadata an Iris LoRA carries, and if they're missing it stops with "is not an Iris LoRA (no rank/alpha metadata)". If the metadata's there but the layers don't exist on the Iris model, you get "targets <layer>, which this Iris model lacks". So you can't silently attach the wrong thing - you get a clear error instead of a picture that quietly didn't change. Annoying the first time, much better than the alternative.

Where do Iris LoRAs come from? The author also maintains an Iris-3B Trainer, and that's what you'd use for your own. Community Iris LoRAs are thin on the ground so far.

How it applies, and why strength 0 matters

The interesting engineering choice: LoRAs are attached live, while sampling, through forward hooks on the target linear layers. The base weights are never modified and no second copy of the model is made. Two consequences fall out of that, and both are genuinely useful:

• Strength 0 gives you exactly the base image. Not "roughly" - the hook is skipped entirely when the effective scale is zero. So you can keep a LoRA wired in your graph, set it to 0, and A/B against a clean baseline by nudging one number. No bypassing, no reloading.
• Switching between Turbo and normal sampling costs nothing. The Turbo LoRA is just another entry in the same live stack, so you're not reloading 12 GB when you move between the fast and slow paths.

The effective scale is strength * alpha / rank - the standard LoRA math, read from the file's own metadata rather than guessed. So strength is a multiplier on what the trainer baked in, not an absolute weight.

Inputs and output

model - from Load Iris-3B Model, or from another loader in the chain.

lora_name - one entry from your LoRA folder. See above.

strength - default 1.0, and the range here is unusually wide: -2 to 3. Negative strengths are legal and can be interesting for pushing a concept out rather than in. In practice, 0.6 to 1.2 is where most adapters live; past 1.5 you're usually just cooking the image.

The single output is IRIS_MODEL, which goes to any of the three samplers. Chain several of these nodes to stack LoRAs - each one adds to the same live stack rather than overwriting. Stacking is where the classic LoRA problems show up: two character LoRAs at strength 1 will bleed into each other, and the fix is lower strengths on both rather than more steps.

Install

ComfyUI-Manager, search Iris-3B. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/bani4kaskashka/Iris-3B-Comfy-Nodes
pip install -r requirements.txt

Drop your Iris LoRAs in ComfyUI/models/loras, restart, and they'll appear in the list. There's a Text to Image (Advanced) example workflow to start from; if you want the Turbo path, put this node after Load Iris Turbo LoRA - the two stack.

Where it goes wrong

"is not an Iris LoRA (no rank/alpha metadata)". You picked an SDXL or Flux LoRA. This loader is not for those. There's no conversion path; train or download an Iris-format adapter.

"targets X, which this Iris model lacks". The file claims to be for Iris but was trained against a different Iris build - most likely a mismatched base version. Try a different adapter.

"X is not a Turbo LoRA; use Load Iris LoRA for regular ones" - the reverse error, from Load Iris Turbo LoRA. You put a normal LoRA in the Turbo slot, or vice versa. The two loaders are type-strict about which metadata they accept.

The image barely moved. Check the effective strength, not just your number: a rank-4 adapter with a low alpha is a small change at strength 1. Try 1.3 before you conclude the LoRA is broken - but if the base image also looks different from the no-LoRA run, you've got the wrong file rather than a weak one.