ComfyUI-Krea2T-Enhancer Introduction
ComfyUI-Krea2T-Enhancer is a collection of ComfyUI nodes for working with Krea2 diffusion models. Its main enhancer nodes aim to improve how closely a generated image follows your prompt. The collection also gives you more control over important phrases, provides a reference sampling schedule for Krea2 Turbo, and offers ways to apply LoRAs while limiting their effects on particular parts of the model.
You can use the tools separately. If you simply want to test whether the main enhancer helps a prompt, place it between your Krea2 model loader and sampler. If a particular detail keeps getting overlooked, try the attention-weighted phrase node. If a character LoRA improves likeness but disrupts other requested details, try the image-only character LoRA loader.
These controls change how the model processes information; they do not guarantee that a particular object or detail will appear. Compare results using the same seed and workflow settings so you can judge what each change actually does.
How ComfyUI-Krea2T-Enhancer Works
A Krea2 workflow turns your written prompt into text information that the diffusion model uses while building an image. Think of the prompt as a set of creative directions and the model as an artist repeatedly checking those directions while refining a draft.
The main enhancer nodes adjust part of Krea2’s internal text-fusion path during sampling. This is where prompt information is prepared for use alongside image information. They return an adjusted MODEL connection, so your usual text encoder, VAE, latent, and sampler setup can remain in place.
The phrase-weighting node works differently. It lets you mark a phrase that deserves more or less attention:
A scene containing a (primary subject:2.0) beside a (secondary object:0.6)
Here, the primary subject receives greater attention priority, while the secondary object receives less. The numbers affect the model’s attention to those words, not the literal size, number, or visibility of objects in the finished image. The node removes the weight notation before encoding the prompt, preserving the words and their order.
LoRAs introduce another kind of adjustment. A character LoRA can improve a character’s appearance but may also affect how the model handles the rest of your prompt. The image-only character loader applies compatible character-LoRA updates to image information without directly applying those updates to text information. Text and image still interact normally, so this reduces one source of interference rather than creating complete isolation.
ComfyUI-Krea2T-Enhancer Features
Main Prompt Enhancer
ComfyUI-Krea2T-Enhancer returns a patched MODEL intended to improve prompt adherence. Connect it between your Krea2 diffusion model loader and sampler:
Load Diffusion Model -> ComfyUI-Krea2T-Enhancer -> KSampler
| Control | What it does |
|---|---|
enabled | Turns the enhancement on or off for a comparison. On by default. |
strength | Sets enhancement strength from 0.0 (neutral) to 2.0 (full range). Default: 1.0. |
debug | Prints brief diagnostic information in the ComfyUI console. Off by default. |
If your result is already close to the prompt, start at the default strength and change only this setting while comparing images.
Advanced Prompt Enhancer
Krea2T Enhancer Advanced uses the same enhancer path and adds text_scale, which changes the strength of fused text information before it enters Krea2’s shared text-and-image processing layers.
Load Diffusion Model -> Krea2T Enhancer Advanced -> KSampler
It has the same enabled, strength, and debug controls as the main node. text_scale defaults to 1.0, its neutral value; the author suggests 1.50 to 2.00 as a starting range for experimentation. Try changing text_scale independently of strength so you can tell which adjustment affects your result.
Attention-Weighted Phrases
Krea2 Text Encode — Attention-Weighted Phrases lets you prioritize specific words or complete phrases. It returns both a MODEL and positive CONDITIONING output, and both must be connected to the workflow:
Final Krea2 model, including LoRAs -> Phrase node -> MODEL to sampler
Krea2 CLIP text encoder -----------> Phrase node -> CONDITIONING to positive input
Enter your prompt in the node’s text field. Write each weighted section as (phrase:weight):
| Weight | Meaning |
|---|---|
1.0 | Neutral: keeps the phrase’s original attention priority. |
Above 1.0 | Increases attention priority. |
Between 0.0 and 1.0 | Reduces attention priority. |
0.0 | Applies the strongest suppression available through this control. |
For example, (red embroidered jacket:1.5) gives the complete clothing detail more priority. Weighting the whole phrase can be more useful than weighting only (jacket:1.5) when the color and embroidery are what matter. You can weight several separate phrases, but weighted sections cannot overlap or contain one another.
Start with 1.0 or an unannotated prompt, then try a moderate value such as 1.5 or 2.0. Hold the seed, resolution, sampler, sigma schedule, and LoRA strengths steady while testing. Change one phrase at a time. Very large weights can compete with composition and other details, and the same weight can look different across seeds.
The optional STRING output reports the cleaned prompt and the text-token pieces selected for each phrase. Connect it to a text preview node if you need to check exactly what was weighted. The phrase node is validated for text-only Krea2 conditioning in the expected 12 x 2560 layout; it rejects incompatible encoders or token layouts rather than guessing.
Krea2 Turbo Reference Sigmas
Krea2 Turbo Reference Sigmas (From Latent) creates a sampling schedule based on the official Krea 2 Turbo scheduler settings described in the original README. Connect the loaded Turbo model and the same latent you will send to the sampler. The node outputs both SIGMAS for sampling and the validated, unchanged LATENT.
| Control | What it does |
|---|---|
steps | Number of Euler denoising steps. Default: 8, matching the reference Turbo setup. |
denoise | Uses the complete schedule at 1.0; a lower value keeps the final requested steps from a longer schedule. |
The node checks that the latent’s image dimensions are divisible by 16. It uses Turbo’s fixed timestep shift rather than the resolution-dependent rule associated with the RAW checkpoint.
Character LoRA — Image Only
Krea2T Character LoRA — Image Only is intended for cases where a character LoRA helps likeness but makes other prompt details less reliable. Use it in place of the regular loader for that character:
Krea2 model -> other LoRA loaders (optional) -> Character LoRA — Image Only -> sampler
Choose your character LoRA or LoKr in lora_name. strength_model controls its influence: 1.0 is the default and 0 turns its effect off. Negative values are supported. Turn enabled off to bypass the loader. Load each character LoRA only once.
The loader supports compatible standard LoRA and LoKr exports. It skips unsupported weights and formats; a file with no compatible weights passes through without changing the model. Its optional report tells you what loaded or was skipped, including saved adapter scaling. A LoRA’s saved scaling can make its effect weaker than you might expect from strength_model alone.
If you also use attention-weighted phrases, place the phrase node after this loader and send the phrase node’s MODEL and CONDITIONING outputs onward.
Selective LoRA Utilities
Krea2 Selective LoRA Loader Model Only (Projection Split) offers a different approach for LoRAs that do not work well with the image-only character loader. It uses normal model-only LoRA loading while letting you switch three Krea2 text-conditioning branches on or off independently: load_projection, load_textfusion, and load_txtmlp. Other LoRA weights remain eligible. Its report shows how many weights were found and retained in each branch. This is not the same as image-token-only isolation.
Krea2 · Projector + External MLP · LoRA Isolation is a focused testing utility. It applies only selected saved updates for the projector and two external text-MLP layers from a reference adapter.
ComfyUI-Krea2T-Enhancer Models
The extension’s nodes are designed for Krea2 models; the provided information does not describe any model checkpoint bundled with the extension. Choose the node that fits the Krea2 setup you are already using:
- Krea2 models with the expected text-fusion layout: Use the main or advanced enhancer. If a loaded model does not match the expected layout, the enhancer patch is skipped.
- Krea2 with text-only
12 x 2560conditioning: Use the attention-weighted phrase node when you need control over specific prompt sections. - Krea2 Turbo: Use the Turbo reference sigma node when you want the supplied Turbo sampling schedule. It is specifically intended for a loaded Krea2 Turbo model.
- Krea2 with a compatible character LoRA or LoKr: Try the image-only loader when likeness is useful but the adapter interferes with other directions.
The original README also mentions a RAW checkpoint to distinguish its resolution-dependent shift rule from Turbo’s fixed shift. It does not provide a broader checkpoint catalog or model-by-model output comparison.
Troubleshooting ComfyUI-Krea2T-Enhancer
The enhancer seems to have no effect.
- Confirm that its
enabledcontrol is on andstrengthis above0.0. - Check that its
MODELoutput, rather than the unpatched model connection, reaches the sampler. - Confirm that you loaded a Krea2 model with the expected text-fusion layout. The patch is skipped when the layout does not match.
- For more information, turn on
debugand check the ComfyUI console. Compare outputs with the same seed and other settings.
A weighted phrase does not change the result as expected.
- Check the syntax:
(complete phrase:1.5). - Confirm that the phrase node receives the final model chain and the Krea2 CLIP text encoder.
- Connect both its
MODELoutput to the sampler and itsCONDITIONINGoutput to the positive-conditioning path. - Start from a neutral weight of
1.0, then change one phrase at a time with the seed and sampling settings fixed. - If needed, inspect the
STRINGreport in a text preview node to see which token pieces received the weight. Remember that attention priority is not a guaranteed visible change.
The phrase node rejects the workflow.
- Confirm that the text encoder is loaded with the Krea2 CLIP type.
- Use text-only Krea2 conditioning with the expected
12 x 2560layout. - Remove visual or custom embedding tokens that change the expected token mapping.
- Confirm that the connected model exposes the expected Krea2 text-fusion and shared-block architecture.
A character LoRA is not loading, or its effect seems unexpectedly weak or strong.
- Check the loader’s
reportfor loaded and skipped weights. Zero loaded weights means the file supplied no compatible updates. - Confirm that you selected the intended LoRA or LoKr and loaded that character adapter only once.
- Check
enabled,strength_model, and the report’s saved scaling information. Saved alpha/rank scaling is applied before your strength setting. - If an older loader produced excessive noise with an alpha-scaled LoRA, update both loader files together and restart ComfyUI. Re-test from your intended strength instead of a value chosen to compensate for the older behavior.
The Turbo sigma node rejects the latent.
- Check that you connected the same latent to the sigma node and sampler.
- Make sure its image dimensions are divisible by 16.
- Confirm that the connected model is Krea2 Turbo, then start with the reference
stepsvalue of8anddenoisevalue of1.0.
Frequently asked questions
- Does phrase weight
2.0make an object twice as visible? No. It doubles the selected phrase’s attention odds before attention is balanced across the available information; the visible effect depends on the image and workflow. - Does the image-only loader completely separate character and prompt influences? No. It prevents the compatible character updates from being applied directly to text tokens, but Krea2 still lets text and image information interact.
- Can other LoRAs stay in the workflow? Yes. The image-only character loader preserves LoRAs already applied to the model. Connect the phrase node after the completed LoRA chain if you use it.
- Do the main enhancer nodes require a different prompt encoder? No. They return a patched
MODELand work with the usual Krea2 text encoder and conditioning setup. - What does the selective loader do differently? It includes or excludes chosen text-conditioning branches during model-only LoRA loading; it does not apply a LoRA only to image tokens.
Learn More about ComfyUI-Krea2T-Enhancer
The ComfyUI-Krea2T-Enhancer repository contains the extension’s original README and project files. For artistic comparisons, the most useful starting point is a working Krea2 workflow: keep its seed and sampling settings fixed, test one node or control at a time, and use the nodes’ reports or debug output when a result differs from what you expected.
