I Just Remade All Of My PRIVATE Workflows With Krea2. Here's One Of em For Free 🤫 (Patreon)

I rebuilt every single one of my core workflows around Krea2 Turbo this week. This model is genuinely so god damn good if you know how to use it. I am blown away. Text to image, image to image (which is free for patrons, attached to this post), inpainting, NSFW, all of it. And the reason isn't just quality. It's that this one model now does two jobs that used to require two separate pipelines.

Node, Setting, and Synopsis Breakdown

Krea2 Turbo on its own is already a low step, high quality generator. I'm running it through `ClownsharKSampler_Beta` at 6 to 8 steps, cfg locked to 1.0, denoise between 0.48 and 0.55 depending on whether I'm doing a fresh generation or working off an existing image. Sampler and scheduler pairs I've landed on are linear/euler with simple scheduler for faster passes, and exponential/dpmpp_sde_2s with beta scheduler when I want more detail resolved. Bongmath on for both. That's already competitive with anything else I've run locally at this step count.

What changed everything is the ComfyUI-Krea2T-Enhancer node. This turns the same krea2_turbo_fp8_scaled checkpoint into an edit model without swapping anything else in the pipeline. Normally, editing an existing image means either a full character remaster through i2i at high denoise, or bringing in a completely separate edit model like Flux Kontext or Klein. Now it's the same UNET, same sampler node, same everything. You just drop the enhancer into the chain and it reads your prompt as an edit instruction instead of a generation instruction.

This Might Kill Leading Edit Model Popularity

Practically, this means my image to image workflow and my edit workflow are now the same workflow. If I want a full remaster, I run it like normal i2i. If I want to change one small thing, a shirt color, an expression, a background object, I don't touch the graph at all. I just change the prompt text and drop denoise down toward that 0.48 range. No node swapping, no reloading a different checkpoint, no rerouting conditioning.

This kills the biggest friction point in most editing pipelines, which is maintaining two separate graphs for "generate" and "edit" and constantly checking which one you're in. One wrong prompt style in the wrong graph and you either get a full remaster when you wanted a small tweak, or nothing changes when you wanted a full remaster. Collapsing that into one workflow means fewer mistakes and way less setup time per client request, since I'm not rebuilding graphs mid-session anymore.

The NSFW side deserves its own callout. I got an NSFW result by accident during a random batch test, no NSFW lora loaded, nothing intentional, and the anatomical quality was clean. That's not normal for a base model. So I loaded an NSFW lora on top of it and at that point it's not close. Best local image model I've run for NSFW quality, full stop, and that's before you factor in that the same checkpoint is also handling my NSFW content and edits.

Will Krea2 Edit Be Better?

Worth being straight about this. There's no official Krea2 Edit model out right now. Krea's own technical report lists robust editing as something they're still working toward, not something shipped. What I'm running is a community node, ComfyUI-Krea2T-Enhancer, doing the edit trick on the same Turbo checkpoint, and there's a separate community LoRA floating around too, Krea2 Identity Edit, built specifically for likeness-preserving edits with its own steps and CFG recommendations. Neither of those is Krea themselves.

So if Krea ships an official edit model built and trained for this from scratch, it'll probably beat the community workaround on likeness preservation and edit precision, since it won't be riding on top of a text to image checkpoint that wasn't built for it. But that's speculation until it's actually out. Right now the enhancer node is what's live and it's already good enough that I switched my whole pipeline over to it.

Real Example

This is already live in my inpainting and NSFW pipelines. Same krea2_turbo_fp8_scaled UNET, same ClownsharKSampler_Beta node, and I'm switching between full remask generations and small prompt edits without touching the graph structure at all. For client work specifically, this means I can do a full character generation and then immediately hop into small revision requests, "change the pose slightly," "different lighting," in the exact same session without reloading anything.

Model: krea2_turbo_fp8_scaled.safetensors (UNETLoader)Sampler node: ClownsharKSampler_BetaSampler/scheduler: linear/euler, simpleSteps: 6CFG: 1.0Denoise: 0.48Eta: 0.5Bongmath: on

The quality is literally nuts...

Takeaway

If you're still running separate checkpoints for generation and editing, stop. Grab ComfyUI-Krea2T-Enhancer, drop it into your existing Krea2 Turbo graph, and you've got both in one pipeline. Steps in the 6 to 8 range, cfg at 1.0, denoise around 0.48 for edits and up toward 0.55 for heavier i2i. This is the first model I've run where I didn't feel like I was compromising on either speed or quality to get both jobs done.