Iris Text Encode

This is the CLIP Text Encode of the Iris world, and it behaves a little differently in one way that trips people up: the negative box wants to stay empty.

Why there's a null prompt at all

Iris-3B is a flow-matching DiT, so it still does classifier-free guidance - the sampler predicts each step twice, once conditioned on your prompt and once without it, and CFG sets how hard the difference gets amplified. That means something has to play the role of "without your prompt". Iris was trained with a specific null prompt for that slot, and this node lets you supply your own text there if you really want to.

The author's tooltip is blunt about the recommended setting: empty gives you the null prompt the model was trained with. Type SD-era negative-prompt incantations in there and you're not steering away from bad hands, you're just conditioning the model on a second prompt it was never given in that role. On a model whose defaults are tuned to the model card, that's usually a downgrade.

If you want more or less prompt adherence, change prompt_adherence (or cfg) on the sampler. That's the dial built for this. Don't reach for the negative box first.

Inputs and output

Three fields, and only one is really a field.

text_encoder takes the IRIS_TEXT_ENCODER from Load Iris Text Encoder (Qwen3-VL-4B). It's a hard requirement - there's no way to hand this node a CLIP, and no way to type text without one, because Qwen3-VL-4B is what produces the 12 stacked hidden layers Iris conditions on.

positive is your prompt, multiline. The default is a red fox in snow at golden hour, which is a decent hint at what the model likes: it's a natural-language model, not a tag soup model. Write sentences. "a red fox sleeping in fresh snow, golden hour" beats "fox, snow, sunset, masterpiece, 8k, best quality" - those last three do nothing here, the same way they do nothing on any 2026-era encoder. Describe the photo, not your wishes for it.

negative, multiline, empty by default. See above. If you must use one, keep it short and structural - something like "blurry" - and be ready to discover that it made things slightly worse.

The output is IRIS_CONDITIONING, a single wire carrying both the positive encoding and the null encoding that were computed together. It plugs into any of the three Iris samplers and into nothing else. Note that it does not have separate positive and negative jacks like you're used to - that's why the pair of text boxes lives in one node.

Install

Part of the same pack. 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

Restart after. No model of its own - the encoding work happens in the text encoder loader's model, which this node briefly moves onto the GPU, runs, and moves back.

Where it goes wrong

"You must connect a text encoder." You didn't wire the loader in, or you wired a ComfyUI CLIP output. Different type, no conversion exists.

The image barely changed when you rewrote the prompt. Check prompt_adherence on the sampler before you blame the encoder - the loose (cfg 2) preset genuinely is loose. And check you didn't put the good prompt in the negative box; the two boxes are easy to swap when you're moving fast.

Nothing renders and the log mentions the encoder. Give it a moment and watch RAM. The text encoder is the largest single thing the pack loads into system memory, and if you're running Iris alongside a big video model you can genuinely run the machine out of memory at encode time.