AI Influencer Master Class - Krea 2 Text-to-Image [VERSION 4 OUT NOW] (Patreon)

This is my straight text-to-image workflow for Krea 2 Turbo, set up for AI influencer content. You write one prompt, your character LoRA takes care of how she looks, and you get a finished image with the eyes already cleaned up. Version 4 fixes the hair texture problem a lot of you ran into, and the whole thing is laid out so you can follow it left to right.

What's new in version 4

The hair texture problem is fixed. A lot of you saw a crunchy, almost painted texture in hair. I traced it back to the regular KSampler, so v4 swaps it for the ClownsharKSampler from RES4LYF. Same speed, clean hair.

New text encoder. All my Krea 2 workflows now use the same one, qwen3-vl-4b-heretic_int8_dynamic_convrot. The download link is at the bottom.

Cleaner layout. Everything sits in labeled boxes that read left to right, in the order the image actually gets made. The How To Use note on the far left has all the download links too, so you don't have to come back here.

Optional film grain. There's a film grain node at the end. It's off by default.

The face detailer is gone. It was giving bad results with Krea 2, so only the eye pass is left.

How to use it

1. Load your character LoRA

In the Models box, find the Power Lora Loader. The last slot is your character. Right now it has my Isabella LoRA in it, so click the name and pick your own LoRA from the list.

Start with the strength at 1.0. If your character starts taking over the whole picture, for example every outfit or background keeps turning into the same thing, bring it down to 0.8 or 0.9.

The other slots above it are realism and detail LoRAs I tuned for this workflow. You can leave them exactly as they are. You'll find download links for all of them at the bottom.

2. Write your prompt

Start the Positive Prompt with your LoRA's trigger word, then the words a woman. After that, describe the photo you want: where she is, what she's wearing, how she's posing, the lighting, and what kind of shot it is (a mirror selfie, a photo a friend took, a close up).

Don't describe her face, hair, eye color or body. Your LoRA already knows what she looks like. When you describe her as well, the prompt and the LoRA start pulling in different directions and her face drifts. Describe the scene and let the LoRA handle her.

Here's an example of the shape a good prompt takes:

YourTrigger, a woman standing on a city sidewalk on a bright afternoon, shot by a friend on an iPhone from about waist height. She wears a fitted black tank top, light wash jeans and a small shoulder bag, looking down and to the side with a small smile. Natural daylight, real street behind her, true phone photo look.

There's no negative prompt to fill in. This workflow runs at cfg 1, and at cfg 1 a negative prompt does nothing, so it's switched off for you.

3. Set your image size

The width and height in the EmptyLatentImage64 node set the size of your image. The default is 904 x 1600, a tall portrait that fits Instagram. Bigger images use more VRAM and take longer. If you run out of memory, try 720 x 1280.

4. The sampler (ClownsharKSampler)

These are tuned for Krea 2 Turbo, so leave them alone unless you're testing something on purpose. For reference, they're set to eta 0.5, sampler linear/ralston_2s, scheduler simple, 8 steps, cfg 1 and bongmath on.

You'll notice the seed says 42 and fixed. That's on purpose. The Smart Seed Variance node (a couple of steps down) changes on every run instead, and that's what gives you a different image each time you hit Queue.

5. Shift

At the top of the Prompt box, the ModelSamplingAuraFlow node sets the shift to 6. Shift decides how the model splits its steps between building the overall picture (the pose, the layout, where things go) and finishing the fine detail. The sampler settings above were tuned together with 6, so keep it there.

6. Smart Seed Variance

This node is the reason your images don't all come out looking the same. On every run it nudges your prompt a little, so you get changes in pose, framing and composition even when the prompt stays exactly the same.

If your images all look too similar, raise fine_tune_variance (it starts at 50).

If your character's face starts drifting and stops looking like her, lower it.

Everything else in this node can stay as it is.

7. Eye detail

Once the image is made, the EyeDetailer finds the eyes, zooms in and redraws just that small area with more detail. Small eyes in a full body shot usually come out mushy, and this pass is what fixes that.

The denoise (set to 0.08) is how much it's allowed to change the eyes. Nudge it up toward 0.15 or 0.2 if you want sharper eyes. Past about 0.25 it starts redrawing the eyes instead of cleaning them up, and your character can stop looking like herself.

This step needs the Eyeful and SAM models from the install section below. Without them it throws an error.

8. Film grain (optional)

Right before the image saves, there's an Image Film Grain node. It's switched off (bypassed) by default, which is why it looks faded. To turn it on, click the node and press Ctrl+B. Press Ctrl+B again to switch it back off. It adds a light layer of grain so the final image looks more like it came off a real camera and less like a perfectly clean render.

Your finished images save into ComfyUI/output/ with the name T2I. You can change that name in the Save Image node.

Installing the required models

Every model here is a one time download. Put each file in the folder shown, then restart ComfyUI (or press R in ComfyUI to refresh the model lists) and it will show up in the dropdowns. Your models folder lives at ComfyUI/models/ inside your ComfyUI install.

Inside models/loras/, make two new folders named KREA2MISC and KREA2CHARS. The LoRA loader looks inside those two folders, so once your LoRAs are in there they all load automatically. If you'd rather keep your LoRAs loose in models/loras/, that works too. You'll just need to click each LoRA slot once and pick the file again.

The main models

Diffusion model: krea2_turbo_fp8_scaled.safetensors goes in models/diffusion_models/. Direct download

Text encoder: qwen3-vl-4b-heretic_int8_dynamic_convrot.safetensors goes in models/text_encoders/. Direct download. Make sure you grab this exact file. There are a few similar looking Qwen3-VL heretic files floating around, and they are not the same.

VAE: qwen_image_vae.safetensors goes in models/vae/. Direct download

The LoRAs (all go in models/loras/KREA2MISC/)

Civitai needs you to be logged in to download, so grab these in your browser. The number after each name is the strength I have it set to in the workflow.

krea2-bloomgirls-realism-step00004000.safetensors at 0.3. Realism (BloomGirls UltraRealism). Download on Civitai

Famegrid-Natural-V1-Krea-2.safetensors at 0.4. Natural smartphone look (Famegrid Krea 2 Natural V1). Download on Civitai

nicegirls_krea2.safetensors at 0.4. Realism (NiceGirls UltraReal, Krea2 version). Download on Civitai

Detailer-KREA2.safetensors at 2.0. Overall detail slider. Download on Civitai

fedor_bypass.safetensors at 1.0. Krea 2 filter bypass. The file is tiny (about 1 KB), that's normal. Download on Civitai

skindetails_krea2_loraholic.safetensors at 0.7. Skin detail slider. Download on Civitai

Your character LoRA at 1.0. Whatever you trained, in models/loras/KREA2CHARS/.

The eye detailer models

Eyeful_v2-Paired.pt goes in models/ultralytics/bbox/. Direct download

sam_vit_b_01ec64.pth goes in models/sams/. Direct download (Meta)

Custom nodes

Open ComfyUI Manager, click Install Custom Nodes, search each name and install it. Or open the workflow and use Install Missing Custom Nodes, which finds them for you.

RES4LYF (the ClownsharKSampler). GitHub

rgthree-comfy (the Power Lora Loader). GitHub

ComfyUI Impact Pack (the eye detailer). GitHub

ComfyUI Impact Subpack (the eye finder). GitHub

RBG Smart Seed Variance. GitHub

WAS Node Suite (the film grain). GitHub

Install commands (Windows)

Open your ComfyUI folder in File Explorer (the folder that has the models folder inside it). Click the address bar at the top, type cmd and press Enter. That opens a Command Prompt already sitting in the right place. Paste these in, one line at a time or all at once:

mkdir models\ultralytics\bbox mkdir models\loras\KREA2MISC mkdir models\loras\KREA2CHARS curl.exe -L -o models\diffusion_models\krea2_turbo_fp8_scaled.safetensors https://huggingface.co/Comfy-Org/Krea-2/resolve/main/diffusion_models/krea2_turbo_fp8_scaled.safetensors curl.exe -L -o models\text_encoders\qwen3-vl-4b-heretic_int8_dynamic_convrot.safetensors https://huggingface.co/ethanfel/Qwen3-VL-4B-Heretic-ComfyUI-ConvRot-INT8/resolve/main/qwen3-vl-4b-heretic_int8_dynamic_convrot.safetensors curl.exe -L -o models\vae\qwen_image_vae.safetensors https://huggingface.co/Comfy-Org/Krea-2/resolve/main/vae/qwen_image_vae.safetensors curl.exe -L -o models\ultralytics\bbox\Eyeful_v2-Paired.pt https://huggingface.co/Tenofas/ComfyUI/resolve/main/ultralytics/bbox/Eyeful_v2-Paired.pt curl.exe -L -o models\sams\sam_vit_b_01ec64.pth https://dl.fbaipublicfiles.com/segment_anything/sam_vit_b_01ec64.pth

If a mkdir line says the folder already exists, that's fine, ignore it. Type curl.exe exactly like that, with the .exe. In PowerShell, plain curl runs a different command and these lines fail. The LoRAs aren't in here because Civitai needs a login, so download those from the links above.

Install commands (RunPod / JupyterLab)

On RunPod your ComfyUI lives at /workspace/ComfyUI, so you can paste these into the JupyterLab terminal from anywhere:

mkdir -p /workspace/ComfyUI/models/ultralytics/bbox /workspace/ComfyUI/models/sams /workspace/ComfyUI/models/loras/KREA2MISC /workspace/ComfyUI/models/loras/KREA2CHARS curl -L -o /workspace/ComfyUI/models/diffusion_models/krea2_turbo_fp8_scaled.safetensors https://huggingface.co/Comfy-Org/Krea-2/resolve/main/diffusion_models/krea2_turbo_fp8_scaled.safetensors curl -L -o /workspace/ComfyUI/models/text_encoders/qwen3-vl-4b-heretic_int8_dynamic_convrot.safetensors https://huggingface.co/ethanfel/Qwen3-VL-4B-Heretic-ComfyUI-ConvRot-INT8/resolve/main/qwen3-vl-4b-heretic_int8_dynamic_convrot.safetensors curl -L -o /workspace/ComfyUI/models/vae/qwen_image_vae.safetensors https://huggingface.co/Comfy-Org/Krea-2/resolve/main/vae/qwen_image_vae.safetensors curl -L -o /workspace/ComfyUI/models/ultralytics/bbox/Eyeful_v2-Paired.pt https://huggingface.co/Tenofas/ComfyUI/resolve/main/ultralytics/bbox/Eyeful_v2-Paired.pt curl -L -o /workspace/ComfyUI/models/sams/sam_vit_b_01ec64.pth https://dl.fbaipublicfiles.com/segment_anything/sam_vit_b_01ec64.pth

For the LoRAs on RunPod, download them on your computer and drag them into the KREA2MISC folder in JupyterLab's file browser.

That's everything. No checkpoint or upscale model needed for this one.

Workflow by LoRAtech