ComfyUI-Krea2T-Enhancer (encoder added)

I added phrase-level attention control to the latest update.

The idea is simple: sometimes you don’t need to push the entire prompt harder. You need Krea2 to pay more attention to the few words that actually reinforce the idea you’re trying to get through.

So now you can do:

a (specific important phrase:1.8) with the rest of the prompt written normally

and selectively increase or decrease how much attention those words receive.

This is done without scaling, duplicating, deleting, or otherwise changing Krea2’s original 12×2560 Qwen conditioning. The prompt gets encoded normally, the node finds the exact Qwen token rows belonging to the weighted phrase, and the weight is applied to image→text attention inside the shared DiT blocks.

1.0 = untouched
>1.0 = more attention priority
<1.0 = less attention priority
0.0 = suppression

It also has an inspection output showing exactly which Qwen token rows/pieces were matched, so there’s no guessing about what the weight actually landed on.

Basically: instead of turning the entire prompt up, you can now point at the parts that matter and tell Krea2 pay more attention to this.

Included in the latest Krea2T Enhancer update.

It is best when used with the refusal reduction LoRA as they complement each-other nicely

Sample Workflow

And yes, I know (word:1.5) looks like we somehow got teleported back to the SDXL days haha. The logic underneath it definitely did not though.

Additional update

Added Krea2 Turbo Reference Sigmas (From Latent) to the same repo.

This is a separate scheduler node that creates a Turbo sigma schedule based on the official Krea2 Turbo scheduler settings. Connect the loaded model and the same Empty Latent Image used for sampling, then send its SIGMAS and LATENT outputs to the sampler.

It defaults to 8 steps, checks that the image dimensions are valid for the model, and keeps the Turbo schedule aligned without manually guessing sigma values.

IMPORTANT: The Empty Latent width and height must both be divisible by 16. If the node gives you a dimension or alignment error, fix the Empty Latent dimensions first. That error is intentional and does not mean the node is broken. Please verify both values are divisible by 16 before reporting an issue.

More technical details are available in the GitHub README.

Additional update

Added Krea2T Enhancer Advanced to the same repo now, original node is still there and untouched.

This is basically the same enhancer but with one extra knob called text_scale, this knob controls the fused text tokens right after txtmlp before they enter the shared Krea2 stream. so if strength alone is not landing the prompt fully, this gives another cleaner push without adding more workflow mess.

Use either the original node or the advanced node, do not stack both unless you are intentionally testing weird behavior.

Start with:

- strength: 1.00

- text_scale: 1.00

Then raise text_scale, I have been pushing it around 1.50 - 2.00.

Again this is still experimental, but this update gives a more direct text scale control while keeping the setup as one MODEL-in / MODEL-out node.

Updated the node with bug fixes now you can push the strength to 2.00 start with 1.00 and increase to get better adherence for certain cases. sweet spot 1.00 - 1.05

With Lora it varies as sometimes even 0.90-0.95 is more than enough. so when using a lora start experimenting from 0.05 until you find a sweet spot :)

Small experimental ComfyUI node for Krea2 Turbo prompt adherence. patches the Krea2 model path at runtime and applies a controlled adjustment inside the text-fusion stage, instead of rewriting the prompt or adding extra conditioning nodes. It current scope is simple:

- one MODEL-in / MODEL-out node

- strength control

- debug toggle for runtime diagnostics

- no extra dependencies

Repo: https://github.com/capitan01R/ComfyUI-Krea2T-Enhancer

This is still experimental, so treat it as an early testing release. The goal is better prompt follow-through on Krea2 workflows while keeping the setup minimal.

Your discretion is advised "proof of concept only" Gore content advisory, Do NOT click the link if you are not comfortable with this type of content
https://imgur.com/a/5ZwrCIb

For people who's interested in the technical side you can find a bit more of explanation here