Load Iris Turbo LoRA

Smallest download in the pack, biggest behaviour change. This node attaches the Iris-3B Turbo LoRA - a DMD2 distillation by pentaCoxian - to an Iris model, and it's the difference between a 54-second image and an 8-second one on a 16 GB card. You pair it with Iris Sampler (Turbo); the two are a matched set, and both of them will tell you off if you mix them up.

What a DMD2 adapter actually changes

A distillation LoRA doesn't add a style, it rewrites the denoising trajectory. The base model needs 30 to 100 steps because it's following a curved path one small increment at a time. The Turbo LoRA was trained so the model can land on a good image in four large jumps instead, with the teacher's guidance baked into the student. Distribution-matching distillation (that's the DMD2 part) is trained against the teacher's output distribution rather than individual samples, which is why four-step output doesn't come out as a smeared average of everything the teacher might have drawn.

Two mechanics worth knowing:

• The LoRA's own metadata carries a turbo_steps count and a flow shift. The Turbo sampler reads both. So the four-step schedule and the shifted times aren't hard-coded in the sampler - they travel with the adapter file. That's why an unrecognised file can't just be dropped in here; the loader checks for turbo_steps and refuses anything that doesn't have it.
• It's applied live via forward hooks, like every LoRA in this pack. The base weights are never modified. So the model stays warm and you can switch back to the normal samplers without reloading 12 GB - you just rewire the graph.

Two variants ship in the recognised list: iris-3b-turbo-r2-1000.safetensors (the recommended one) and iris-3b-turbo-r1-1500.safetensors. Both download on first use, 460 MB. You don't need to fetch them yourself; Iris-3B Text to Image (Turbo) has this node already wired up.

Inputs and output

model - from Load Iris-3B Model.

lora_name - filtered, unlike the plain LoRA loader: only files in models/loras with both "iris" and "turbo" in the name appear. Your SDXL LoRA collection won't clutter this list.

strength - default 1.0, range 0 to 2. Note there's no negative range here, unlike Load Iris LoRA. Strength 1 is what it was trained for; the pack's own guidance is to leave it there. Sliding it down doesn't blend toward the base model so much as break the four-step assumption the sampler is running on.

Output is IRIS_MODEL, and it goes into Iris Sampler (Turbo). Wiring it into the Simple or Advanced sampler produces a refusal, not a picture.

Install

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

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

Just omegaconf. Restart, then open the Turbo workflow from Templates or drag the shipped example JSON in.

Where it goes wrong

"X is not a Turbo LoRA; use Load Iris LoRA for regular ones." You pointed this at a normal Iris adapter. Wrong loader.

"Iris Sampler (Turbo) needs a model from Load Iris Turbo LoRA." You skipped this node. It's not optional - the Turbo sampler looks for the adapter on the incoming model and errors if it isn't there, because without the distillation there is no valid four-step schedule to run.

Nothing downloaded and the run failed. The LoRA is fetched on first use, which means first successful execution of this node - not when the graph loads. If you're offline, grab iris-3b-turbo-r2-1000.safetensors from the pentaCoxian repo and drop it in models/loras by hand; the loader will find it by name.

You stacked a character LoRA on top and it broke. Combining is supported - chain Load Iris LoRA with this one - but be gentler than usual. Two adapters fighting over a four-step trajectory have far less room to resolve the conflict than they would over 30 steps. Drop the style LoRA to 0.5–0.7 and see.