⭐ Star Krea2 Unbound

Krea2 is the 12B flow-matching model that took over the open ecosystem in mid-2026, and one of its quirks is that prompts sometimes come out soft - the renderer gets the mood right and the details wrong. Star Krea2 Unbound is a model patcher that cranks the model's internal text attention to make it listen harder, without you rewriting a single token. It takes a MODEL, patches it, and gives you the same MODEL back, so it drops into your Krea2 pipeline between the loader and the sampler like a small, specialized volume knob.

What it actually does

This is not a sampler setting or a CFG trick - it's a wrapper around Krea2's text-fusion stack. The node clones your model, checks it's actually a Krea2 architecture (12 text taps at dimension 2560 - the guard that stops it misbehaving on other models), then patches the txtfusion forward pass. The mechanism, from the source: it runs the text projection once to get a reference, runs it again on amplified text-tap activations (a gain profile that boosts specific tap groups, with one strong boost applied), and then blends the amplified result's delta back into the reference - capped at 75% of the reference's RMS so it can't go haywire.

In plain terms: it re-weights the model's own text conditioning so the most information-dense taps contribute harder. It's entirely a forward-time intervention - no retraining, no LoRA, no extra weights. One downside to know: because it re-runs the text taps, it adds a small amount of compute per sampling step. On a 12B model that's not free, but it's also not the bottleneck.

Inputs and outputs

• model (MODEL) - your Krea2 checkpoint, straight from a loader. This is the only input.
• model (MODEL) - the patched clone. Wire it into your KSampler's model input exactly where the original went.

The node is a no-op safety net if the input isn't Krea2 - it passes the model through untouched rather than crashing, which is the right failure mode for a helper.

Where it fits

Reach for it when Krea2 keeps ignoring a specific element of your prompt - the classic "everything's right except the thing I asked for." Use it instead of cranking CFG or adding duplicate prompt weight, which on flow-matching models can distort the output (the KB's advice about flow-matching models: they don't behave like SDXL, so don't carry old habits in). Pair it with a prompt rebalance if you're also fighting a crowded prompt.

Installing

Part of the StarNodes pack - install Starnodes via ComfyUI Manager, or:

cd ComfyUI/custom_nodes
git clone https://github.com/Starnodes2024/ComfyUI_StarNodes
cd ComfyUI_StarNodes
pip install -r requirements.txt

Restart, then search star on the canvas. No models to download - it works on whatever Krea2 weights you already have.

Common issues

If you don't own a Krea2 model this node does nothing - it's a niche tool, and that's okay. On non-Krea2 checkpoints it passes through silently, so don't expect it to fix adherence on SDXL. And if you see doubled quality cost or odd artifacts, remember it amplifies specific taps; running it alongside heavy LoRAs that also touch text conditioning can compound. When in doubt, A/B it with a wipe comparison like ⭐ Star Image Compare.