ComfyUI-Krea2T-Enhancer

If you've run Krea 2 locally and felt like it's only half-listening - you ask for a specific expression, or something with a bit of edge to it, and the model quietly renders something blander instead - you're not imagining it. The open Krea 2 weights went through an alignment pass the hosted API version never got, and Krea's own head of research confirmed it: the local model follows prompts noticeably worse than the one on their website. The community's word for what's left behind is "craters" - whole areas (bodies, violence, strong emotion) the model skips rather than botches. ComfyUI-Krea2T-Enhancer is one attempt to fill them back in.

It's a single MODEL-in, MODEL-out node you drop between your diffusion model loader and your sampler. No prompt rewriting, no extra conditioning nodes, nothing to download. It reaches into Krea 2's text-fusion stage - the point where your encoded text tokens get merged into the image stream - and nudges that signal up so the model weights your prompt details the way it should.

How it works

Krea 2 reads your prompt through a Qwen3-VL text encoder and fuses those tokens into the diffusion transformer during sampling. The alignment training baked into the open weights damps some of that fusion, which is why clauses get dropped. This node patches that path at runtime, applies a controlled adjustment to the internal conditioning, then cleans up after itself when the run finishes. It only touches the MODEL that comes out the other side - your text encoder, VAE, latent and sampler stay exactly as they were. If what you feed it isn't a Krea 2 (the patch checks for Krea 2's specific 12×2560 text-conditioning layout), it just skips rather than erroring.

Worth knowing where this sits: it's roughly the third community fix for the same problem. There's the conditioning-rebalance node and an uncensor LoRA already, and both cost you image quality and character knowledge. This one is pitched as a continuation of that line of work, and early testers reported it hurts the model less than the rebalance node does. The author is upfront that it's experimental - a good-enough first pass, not a finished solution.

The inputs and outputs that matter

There are only three knobs, and one of them is a toggle:

• strength (default 1.0, range 0–2) - the whole node in one dial. It blends from neutral (0.0, does nothing) up to full (2.0). The author's own advice is to keep it between 0.75 and 1.00; the default of 1.0 is a fine starting point and you rarely want to shove it hard.
• enabled (default true) - turns the patch on and off so you can A/B against a clean run without unplugging anything.
• debug (default false) - prints diagnostics to the ComfyUI console; leave it off unless something looks wrong.

The only output is model (MODEL) - wire it straight into your KSampler where your Krea 2 model used to go.

How to install it

No dependencies, no model files. The manual route:

cd ComfyUI/custom_nodes
git clone https://github.com/capitan01R/ComfyUI-Krea2T-Enhancer

Then restart ComfyUI. The README is explicit that nothing beyond a working Krea 2 setup is required, so if Krea 2 already runs for you, this will too. Prefer ComfyUI Manager? Search the pack title, ComfyUI-Krea2T-Enhancer, install, restart.

Common issues & troubleshooting

The node forces re-renders. A tester hit a case where swapping the VAE at decode re-ran the whole graph - something a normal node wouldn't do. That's the runtime patch leaking into ComfyUI's caching. If it bites you, the sibling Krea2T Enhancer Advanced was rebuilt specifically to restore its patches cleanly so cached results reuse normally; switch to that one.

Output looks mushy or too cartoonish. First results tend to skew illustration/anime. The author's fix is a sampler swap: use res_2s for realism (euler is just what the default workflow happened to use). That's a Krea 2 sampler thing, not a node thing.

Nothing changes. If the patch silently no-ops, you're probably not on a real Krea 2 checkpoint - the layout check has to pass first. Confirm you're loading Krea 2 Raw or Turbo, then flip debug on to see whether the patch actually applied.

You expected an off switch for the filter. This isn't a checkbox that disables Krea 2's safety training. It boosts prompt adherence, which recovers some of what the alignment pass buried, but it's a strength dial, not a kill switch.