Krea2 Projector Tuner for ComfyUI ⚡
The ultimate in-memory projector tuning tool designed to patch and customize txtfusion.projector layers of Krea2 models in real-time.
Enables instant censorship/NSFW bypass and prompt-following enhancements with zero permanent changes to your model files!
💡 Ultimate Quality:
You can also pair this tuner with our companion node FLSampler (BSS) to perfectly restore any lost micro-details!
🔗 GitHub Repository (Full Guide & Docs): https://github.com/BlackSnowSkill/ComfyUI-Krea2-Projector-Tuner
🌟 Key Features (v2.0.0)
🧠 In-Memory Dynamic Patching: Patches model weights temporarily in VRAM during execution using ComfyUI's native patching pipeline. Zero disk writes, zero model file corruption risk.
📂 Dynamic JSON Presets & Easy Sharing: All configurations are stored in the local
presets/folder as simple.jsonfiles. This allows the community to easily share, import, or back up their favorite tuning setups! Just copy any community preset.jsonto yourpresets/folder and it will load instantly.🎛️ Interactive Web UI & On-Node Buttons: Save and manage your custom configurations directly from the node interface:
Save Preset (BSS): Prompt-saves your manual sliders setup as a new JSON preset on the fly.Delete Selected Preset (BSS): Instantly deletes the current preset file from your disk.Open Presets Folder (BSS): Opens the presets folder in your OS file explorer.All sliders update dynamically on your canvas when choosing a preset!
🔥 Premium Default Presets:
Strongx3_wl7(Default & Best): The absolute best bypass configuration. Applies a 3x suppression on safety layers 8–10, plus a targeted-0.7shift on layer 7 to lock down the last safety peak.Strong 2.5: Standard decensoring patch.Clean Bypass: Clean safety filter bypass.
🎚️ 12-Layer Slider Control: Custom sliders mapped directly to the 12 CLIP text encoder layers, giving you full manual control over style, details, layout, and censorship with a comfortable
0.01dragging step.
📂 Quick Installation
Method 1 (Recommended):
Open ComfyUI Manager -> click Install Custom Nodes.
Search for
Krea2 Projector Tuner.Click Install and restart ComfyUI.
Method 2 (Fallback via Git URL):
Open ComfyUI Manager -> click Install via Git URL.
Paste this URL:
https://github.com/BlackSnowSkill/ComfyUI-Krea2-Projector-TunerClick Install and restart ComfyUI.
For detailed parameters, portable environment wheels, and manual installation, please refer to the GitHub README.
☕ Support & License
Support Development: Support me and get exclusive models on Boosty!
License & Usage: © 2026 blacksnowskill (BSS). All rights reserved. This project is protected by copyright. Copying, distribution, merging, or use on other websites/repositories without explicit written permission is strictly prohibited.
Description
FAQ
Comments (26)
Is this for base Krea2, or is this something that can/should be used with the other diffusion models based on Krea2?
Likely a dumb question, apologies :)
for all krea2
how to use it in comfyUI? with what node?
loader > lora > ksampler
这些版本哪个最好？或者说各自的应用于什么场景？
"Strongx3_wl7" is the strongest for nsfw, the usage scenario depends on your personal goal. test it out
this is like Krea2FilterBypass 3Vector KREA2 by S1LV3RC01N but slightly different approach, right?
I do not know what approach the author of "Krea2FilterBypass" used, but I do not know any other ways besides what I used, so I think yes
I experimented with adjusting various knobs to see how the model responded. I eventually found that a 3.0x scale on layers 8-10 worked best, and through further iterations, I discovered that adding a small -0.7 shift to layer 7 was the missing piece to fully clear the filter.
@llikswonskcalb Thanks for the answer! To be honest, I know absolutely nothing about these mechanics, but I'm hooked on your work, so I'll definitely try your patch, and if something works out, I'll mention it when I post my new generations.
Sorry for the dumb question, what's the difference from using an abliterated version of Qwen?
Qwen allows you to write, expand, and refine sensitive prompts using a local LLM without it refusing or censoring the text.
Krea2 Projector Tuner bypasses censorship in the image generation phase (diffusion model). Even if you feed a perfectly detailed prompt written by an abliterated Qwen, the Krea2 image model will still censor the final image (causing severe anatomy degradation, weird distortions, or abstract noise) because of its built-in text-to-image projector alignment.
Simply put: Abliterated Qwen allows you to write the prompt without restrictions, while this Tuner allows the Krea2 model to actually draw it.
@llikswonskcalb Thank you ❤️
in Anatomy Opt, how much strength did you put on each knob?
could you tell me the same for the others? i like to expriment with it, thank you for the amazing tuner !
wrote all the values in "About this version"
@llikswonskcalb thank you so much ! do you suggest using the default text encoder? or the heretic one?
@necroryona I highly recommend using the default text encoder with this tuner!
Here is why:
No extra downloads needed: The tuner node handles the censorship bypass dynamically in-memory. You don’t need to download large custom "heretic" CLIP/T5 files.
Better prompt adherence & quality: "Heretic" encoders can sometimes degrade text comprehension, spelling, and style quality on non-NSFW concepts. By using the default text encoder, you keep 100% of the default model's spelling and coherence, while the tuner handles the decensoring on the fly.
Redundancy: While the tuner is fully compatible with heretic encoders if you choose to use them, the default encoder + tuner is already more than powerful enough on its own.
@llikswonskcalb thank you so much, will share my testings once i'm done, so far it's working really good but i just wish there was a description of what each Knob does exactly, Krea2 is the most powerful I2V so far in my eyes, it's a shame they had to censor it in this complicated way.
but i do understand since it's too damn deep.
How should the 11 knobs be set?Thank you!
You don't need to manually adjust the knobs to get a great result. For plug-and-play usage, just select one of the built-in presets from the dropdown. The default Strongx3_wl7 for example.
However, if you want to run your own experiments or fine-tune specific parameters for a special case, you can set the preset to "custom" and adjust the sliders manually.
This is heavy! I don't mean the file size of course, but the feeling I got, when I realized that this patch is almost as effective as the cumshot lora I've just trained. I'm puzzled... Using this patch alone, can do most of the NSFW scenes if you prompted correctly. wow and thanks for sharing!
Would love to see the node updated so that when selecting a preset it shows what the preset is doing to the knobs rather than always showing 0, this would help people learn and understand what is actually happening and tweak values themselves. Cheers for the awesome node!
done
@llikswonskcalb Ran into a couple issues on V2 but vibe coded a fix, had a summary done up for you in case this helps:
Found Issue: Moving the mouse over the node triggers spurious LiteGraph widget callbacks, causing the active preset to instantly snap back to "custom".
Fix: Tracked loaded preset values in a local array (currentPresetValues). In the knob callback, only switch the dropdown to "custom" if the incoming slider value genuinely differs from the tracked preset state (Math.abs(diff) > 1e-6).
Found Issue: Sliders do not visually update when a preset is selected until the user interacts with the node to force a redraw.
Fix: Appended node.setDirtyCanvas(true, true); immediately after the asynchronous preset fetch() applies values to the widgets, forcing an instant UI refresh.
Found Issue: Manually selecting "custom" from the dropdown leaves the sliders at their previous preset values instead of zeroing them out.
Fix: Added logic to presetWidget.callback to reset all 12 knobs to 0.0 if "custom" is explicitly selected. Used a boolean flag (isProgrammaticCustom) to bypass this reset when the script automatically switches to "custom" because a user began dragging a preset slider.
@Adeptus08 Thank you so much for the feedback! I've successfully implemented all your fixes (spurious callbacks filtering, instant setDirtyCanvas redraw, and manual custom reset with programmatic checks) in the new update. The repository has been updated. Thanks for making the tuner better!
Team work baaaby. i was just impressed with this convo tbh lol