Iris Sampler (Turbo)

Four model passes instead of thirty. On the author's reference card that's about 8 seconds per 1024x1024 image against 54 seconds for the fast preset on the normal sampler. That's the pitch, and it's real.

The catch is the usual distillation catch: guidance is baked into the LoRA during training, so this sampler uses only the positive prompt and computes no unconditional pass at all. You don't get a CFG field here. You don't get steps-as-you-choose either, not really - the LoRA was distilled for 4, and the field lets you pick 1 to 16 but 4 is what it learned.

How the four steps work

This isn't DPM-Solver++ with a short step count, it's a different loop. The sampler takes four evenly spaced flow times from pure noise, shifted by a value the LoRA itself carries in its metadata, and at each one does a single model pass. From the velocity prediction it computes the clean image at that step, then - if there's another step coming - re-noises that clean image with fresh noise to the next time level. Predict, project, renoise, repeat. That's the standard few-step distilled loop, and it's why extra steps past 4 don't help: there's no solver accumulating history, just four jumps.

If you know DMD2, this is DMD2. Distribution-matching distillation trains a student to land where the teacher's distribution lands rather than where one particular teacher sample landed, which is what keeps 4-step output from looking like a blurry average. The Turbo LoRA here is pentaCoxian's, and the pack downloads it rather than bundling it.

Inputs and the output

model - and this is the important one - must come from Load Iris Turbo LoRA, not straight from the model loader. If there's no Turbo LoRA on it, the node errors with "Iris Sampler (Turbo) needs a model from Load Iris Turbo LoRA". Conversely, the simple and advanced samplers refuse to accept a model that has one. The two sampler families are mutually exclusive by design, which is fine because the base weights are never modified, so flipping between them is a graph edit and a reload, not a re-download.

conditioning - from Iris Text Encode. Only your positive text is used. Nothing you write in the negative box will reach the model here, so don't bother.

resolution - eleven fixed aspect ratios, all about 1 MP, and no match input image option. That's not an oversight: it was trained around a megapixel and the sizes are the same ones the normal samplers use. 21:9 through 9:21 is plenty for composition work.

steps - default 4, range 1 to 16.

seed and batch_size - as everywhere else, batch 1 to 8.

One output: IMAGE. Straight to Save Image, your preview node, or Iris Upscale.

Install

Search Iris-3B in ComfyUI-Manager, or:

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

The LoRA itself is 460 MB and downloads on first use - iris-3b-turbo-r2-1000.safetensors is the recommended one, with iris-3b-turbo-r1-1500 also recognised. Easiest path is the Iris-3B Text to Image (Turbo) example workflow, which already has the LoRA loader wired between the model loader and this node.

Where it goes wrong

"Iris Sampler (Turbo) needs a model from Load Iris Turbo LoRA." You wired the model loader straight in. Add Load Iris Turbo LoRA in between.

"this model has a Turbo LoRA attached: connect it to Iris Sampler (Turbo) instead." The mirror image. You're on Simple or Advanced with a Turbo model.

Strength that isn't 1. The Turbo loader allows 0 to 2, but 1 is what it was trained for. Dropping it toward 0 doesn't blend gracefully toward the base model the way a normal LoRA does - it just breaks the four-step trajectory, because the schedule and the adapter are a matched pair.

Faces and text got worse. This is the documented tradeoff, not something you've misconfigured. The Turbo LoRA's card notes less variety and sometimes weaker text than the full sampler. Use it for iteration, seed farming and batch exploration; go back to Iris Sampler (Simple) at 30–50 steps when you've got the shot and you want the good version.

You want a LoRA and Turbo. Do it - the loaders chain, and Turbo stacks with ordinary Iris LoRAs. Put Load Iris LoRA after Load Iris Turbo LoRA, or the other way round; both orders accumulate.