Built around uninterrupted visual iteration.

Krea 2 Operator is the public version of the workspace I built through daily Krea 2 use. It was never originally meant to be released. It grew around the small things that interrupt a session: overwriting a prompt I was not finished with, losing track of a useful seed, digging through LoRAs, moving too far away from the image, and breaking the train of thought while generating.

The layout follows attention, not symmetry.

The active prompt is the center of the workspace. The KSampler sits beside it because seed mode, steps, sampler, and scheduler are controls I change constantly. Model and resolution controls sit above it. The image preview stays large because the point is to generate, inspect, adjust, and run again without leaving the graph.

No upscaler maze. No extra processing added just to make the workflow look advanced.

WHAT IS INCLUDED

• Large in-graph image review area

• Prompt parking for preserving a complete prompt while testing another

• Blank scratch Notes that can be duplicated whenever more space is needed

• ENDURING RECIPE space for exact phrases that consistently work

• STYLE LORA TRIGGERS space for trigger words and style instructions

• Krea 2 facial Expression Bank

• Face Micro-Controls for gaze, eyebrows, mouth, lips, jaw, and facial tension

• Separate Bypass, Primary, and Style LoRA racks

• Increment-based seed workflow for easier backtracking

• Date-organized output folders

THE LORA RACKS

PRIMARY

The Primary rack is intentionally allowed to extend straight to Hell.

Permanent residents naturally stay near the top. Familiar combinations become a kind of breaker box in the middle. New LoRAs enter farther down as experiments. When weak or overly specific LoRAs are removed, the ones that survive gradually move upward.

This is not automatic sorting. It is simply how the rack organizes itself through real use.

BYPASS

The Bypass rack lives above the main workspace. It stays out of attention until Krea starts fighting the prompt, then it is close enough to reach without pulling you far away from the image.

STYLE

The Style rack sits beneath the image review area. Style changes are useful, but they are secondary to the main prompt and subject iteration loop, so they do not crowd the active workspace.

PUBLIC STARTING POINT

This release is saved with the original Krea 2 Turbo MXFP8 model and a familiar base-Turbo setup:

• 8 steps

• Euler

• Simple scheduler

• CFG 1

• Denoise 1

• 1:1 aspect ratio

• 2 megapixels

• Batch size 1

• Seed mode: increment

Two megapixels is the normal working resolution. Three megapixels can be used when you want a larger final render.

Increment keeps nearby seed numbers navigable. Switch to fixed when comparing prompt or LoRA changes. Use randomize when seed history no longer matters.

REQUIRED SETUP

The workflow uses:

• ComfyUI with subgraph support

• rgthree-comfy Power LoRA Loader

• Krea 2 Turbo MXFP8 diffusion model (I use many fine-tunes featured on Civitai)

• qwen3vl_4b_bf16.safetensors (You all know the drill)

• qwen_image_vae.safetensors (wan_2.1_vae.safetensors works as well)

The workflow currently references:

krea2Turbo_mxfp8.safetensors

Depending on where/when you downloaded it, your local copy may be named differently.

I've tested many models featured here on Civitai in this workflow.

Model weights and custom-node code are not included in the workflow JSON.

OPTIONAL SAVED LORAS

All saved LoRA entries begin OFF. They are references, not requirements.

• Fedor Krea2 Filter Bypass

https://civarchive.com/models/2746817/krea2-filter-bypass-fedor?modelVersionId=3089754

• Krea2 Filter Bypass

The workflow references both Bypass and Bypass 3 from this resource page.

https://civarchive.com/models/2728234/krea2filterbypass?modelVersionId=3067151

• Krea2 TextFusion Refusal Reduction

https://civarchive.com/models/2775340/krea2-textfusion-refusal-reduction-lora?modelVersionId=3125118

• Wetness Slider Krea2 / ZIT

Saved in the Primary rack at strength -2, disabled by default.

https://civarchive.com/models/2738333/wetness-slider-krea2-zit?modelVersionId=3079282

Install only what interests you. Reselect the files from your own folders before enabling them, or remove the entries you do not use.

OPTIONAL CLIP ROUTE

By default, the prompt encoder receives CLIP directly from KREA 2 CORE.

The CLIP path through the LoRA racks is wired and available, but its final output is deliberately unused. Connect the STYLE LORAS CLIP output to the prompt encoder only when you specifically want compatible text-encoder LoRA weights applied.

The Note beside that route can be deleted after reading.

MAKE IT YOURS

The Expression Bank and Face Micro-Controls contain ordinary descriptive prompt phrases. They are not special trigger tokens and do not require an expression LoRA.

The Notes are manual copy-and-paste workspaces. They do not automatically modify the active prompt.

Move anything that does not fit the way you work. Keep what does. Let the rack grow.

If Operator earns a place in your setup, post what you make. Buzz tips are appreciated, but useful feedback about friction is valuable too.

That is how this workflow got built.

I’ll be releasing my “BREACHED STACK” version with the fine-tune I actually use every day, plus every LoRA that has earned a permanent place in my working stack.