Krea2T Enhancer Advanced

Same idea as the plain Krea2T-Enhancer, with one more knob a lot of people will actually want. If you've run into Krea 2's habit of soft-ignoring prompt details - the flattened expressions, the clauses about bodies or violence it just skips - the base Enhancer gives you a single strength dial to push the model back toward following you. Krea2T Enhancer Advanced keeps that dial and adds text_scale, a more direct lever on how loudly your text tokens speak.

Quick recap of why any of this exists: the open Krea 2 weights got an alignment pass the hosted version didn't, so the local model follows prompts worse - Krea confirmed this on the record. The community calls the missing bits "craters," and this node family is one of the fixes for them.

What text_scale actually does

The base node adjusts Krea 2's text-fusion path as a whole. The Advanced node adds a second, sharper intervention: right after the model's txtmlp layer processes your text tokens - and before they enter the shared Krea 2 stream where they steer the image - it multiplies them by text_scale. In plain terms, it turns up the volume on your prompt at the exact moment it gets injected. Bigger multiplier, harder the model leans on what you wrote.

text_scale runs from 0.25 to 4.0 with a neutral point of 1.0 (multiply by one, nothing changes). The README's suggested working range is 1.50 to 2.00, which is a real boost - this dial is meant to be pushed, unlike strength, which the author suggests keeping near 1.0. Take it too high and you'll oversaturate or fry detail; that's your cue to back off.

The other reason to pick this one: caching

The original Enhancer had a rough edge. Because it patched the model at runtime, ComfyUI sometimes couldn't reuse cached results and re-ran the graph when it shouldn't have - one tester saw a full re-render just from swapping the VAE at decode. The Advanced node was rebuilt to fix exactly that: it restores every temporary patch after each model call and keeps no step-local state in the model config, so with the same seed and the same settings ComfyUI caches and reuses normally. If you iterate a lot - tweaking downstream nodes without touching the sampler - that alone is a reason to run Advanced over the base node.

The inputs and outputs that matter

• text_scale (default 1.0, range 0.25–4.0) - the new lever. Start around 1.5, climb toward 2.0 if the model's still ignoring you, drop back if things look cooked.
• strength (default 1.0, range 0–2) - the same enhancer strength as the base node. Keep it near 0.75–1.0; use it together with text_scale rather than instead of it.
• enabled (default true) - patch on/off, for clean A/B tests.
• debug (default false) - console diagnostics; off unless you're troubleshooting.

The output is a single model (MODEL). Drop it between your Krea 2 loader and your KSampler; everything else - text encoder, VAE, latent, sampler - stays standard.

How to install it

One repo covers both nodes, so if you already installed the pack you've got this. Otherwise:

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

Then restart ComfyUI. No extra Python packages, no model downloads - a working Krea 2 setup is the only requirement. In ComfyUI Manager, search ComfyUI-Krea2T-Enhancer.

Common issues & troubleshooting

Nothing happens. The patch only applies to genuine Krea 2 models (it checks for the 12×2560 text-conditioning layout) and skips everything else. If you see no effect, confirm you're on Krea 2 Raw or Turbo and flip debug on to watch the patch report.

Results look burnt. That's almost always text_scale set too high - come down toward 1.5. And remember you now have two overlapping controls: if you crank both strength and text_scale, they stack. Move one at a time.

Mushy or cartoonish detail. More of a Krea 2 sampler quirk than a node problem - the author recommends res_2s for realism instead of the default euler.

Which one do I use, this or the base node? If you want the simplest one-knob version and don't care about caching, the base Enhancer is fine. If you want the extra text-token control, or you iterate heavily and want cached reuse to keep working, this is the one.