KREA 2 IMAGE SUITE [version 3] IS HERE (text2img + img2img) - Enjoy, Master Class Subs (Patreon)

v2 has been doing its job, and this is where it goes next. v3 rebuilds the LoRA layer and brings in a few new nodes, all tuned around one goal: making your girl look like a real person taking real photos on a real phone, not a model in a studio render.

Both v3 workflows are live now on the AI Influencer Masterclass tier. Grab them here:

Krea 2 text2img v3 https://www.patreon.com/LoRAtech/posts/ai-influencer-2-166643382

Krea 2 img2img v3 https://www.patreon.com/LoRAtech/posts/ai-influencer-2-166643953

The new Influencer Style LoRA stack

v2 ran two realism LoRAs at 0.6. v3 runs four at 0.4 each. Bloomgirls and Famegrid stay, and I added Canon UltraReal and NiceGirls UltraReal to the mix. Four lighter layers blend into each other instead of stamping over each other, and dropping every one to 0.4 leaves real room for your character LoRA to actually come through.

What you get out of it is the everyday look. Slightly uneven skin, real pore texture, natural color instead of that overcooked saturation, and lighting that reads like a phone camera handled it rather than a lighting rig. This is the stack I run on my own accounts, tuned over a lot of bad batches so you don't have to repeat them.

The new nodes

1. Power Lora Loader

The old stack gave you four fixed slots and no clean way to switch one off. The Power Lora Loader gives you unlimited slots with a toggle on every single one, so you can pull a LoRA in and out mid session and feel exactly what it's contributing without ever touching your wiring.

2. Smart Seed Variance

This is the one that fixes the clone problem. It injects controlled, seed keyed noise into your conditioning before it reaches the sampler, so pose, framing and composition genuinely move between generations instead of giving you four slightly different crops of the same shot. Real people don't post the same photo four times. Now your batches don't either. It ships on the Krea2 SingleStream preset. If things still feel samey, raise the variance strength here before you touch anything else.

3. The detailer section, live out of the box

SAM, the bbox detectors and the eye detailer are wired and running by default now, so the detail pass is on from the first queue instead of something you have to go switch on. You still need the SAM and bbox models installed for it to fire.

4. Image Film Grain

Swapped in and turned on at the end of the chain at a light setting. Just enough noise to kill the clean digital surface that gives AI images away at a glance.

New LoRAs to download

Both are the Krea 2 versions from Danrisi. Drop them in your loras folder.

Canon UltraReal (canon_krea2.safetensors) https://civitai.com/models/2783143/canon-ultrareal Trained on Canon 1Ds photography for its color rendering and bokeh. Trigger word is c2n0n.

NiceGirls UltraReal (nicegirls_krea2.safetensors) https://civitai.com/models/1862761/nicegirls-ultrareal General realism and skin quality lift across faces and bodies.

You also need the Smart Seed Variance node pack installed or that node shows up red on load: https://github.com/RamonGuthrie/ComfyUI-RBG-SmartSeedVariance

Read this before you run it

The workflows ship with no character LoRA loaded. The four slots you see filled are the Influencer Style stack, realism only. You have to add your own character LoRA in the Power Lora Loader yourself, in a new slot, or you will get a generic woman who looks nothing like your character.

Start her at 1.0 and pull down if she starts overpowering the realism layers. Then update the positive prompt so it opens with your trigger word before anything else.

Coming next

v3 of the NSFW workflow suite is in testing right now and lands here soon. Same rebuilt LoRA layer, same variance handling, tuned for that side of the pipeline. Stay tuned.

Thanks for the v2 feedback in the Discord, most of this came directly out of it.