Krea 2 ONE STUDIO (TJ)

Krea 2 ONE STUDIO (TJ) is the one-node UI wrapper for Krea 2, the 12B DiT that most people consider the best open base model released in the first half of 2026 - the one that took an entire ecosystem off Ideogram 4 in about a week. If you've been hearing that Krea 2 is astonishing and also that it's fiddly, this node is both of those things at once: the base text-to-image and img2img experience is genuinely good, and the marquee extra features (ControlNet, identity editing) are marked experimental and will absolutely bite you.

Let's be clear about what Krea 2 is, because it shapes everything. It's Krea's first fully in-house model - single-stream DiT, a Qwen3-VL text encoder, the Qwen-Image VAE, and a license permissive enough to matter. The community's favorite way to run it isn't even either shipped checkpoint; it's the Raw base with the Turbo LoRA at 0.6 weight, about 12 steps. This node just wants the Comfy-Org Krea2 files in the standard folders and takes care of the graph for you.

How it works

Same architecture as the rest of the TJ NODE STUDIO ONE family: the node itself is a thin Python shell that returns the last generated image as an IMAGE, and all the real work is the JavaScript loading a prebuilt workflow per mode and running it invisibly. So the rules that apply everywhere in this pack apply here: generate with the node's ▶ Generate button, not ComfyUI's RUN - RUN only re-emits the last image through the output slot. And if a mode fails, it's usually a missing dependency or model, not a setting.

The inputs and outputs that matter

One optional input, prompt_override - a multiline STRING prepended before the prompt you type in the panel. That's your hook for driving it from a Primitive (String) node or TJ_NODE's Prompt Studio. One output, image (IMAGE), and it's an output node. Everything else - steps, CFG, sampler, the up-to-three LoRA chain - is panel and Settings territory.

Installing it

ComfyUI Manager (search ComfyUI-TJ_NODE_STUDIO_ONE) or:

cd ComfyUI/custom_nodes
git clone https://github.com/designloves2/ComfyUI-TJ_NODE_STUDIO_ONE.git

Restart ComfyUI, then grab the dependencies with the bundled install_requirements.sh / install_requirements.bat. Base usage needs just three model files from Comfy-Org's Krea-2 repo: the diffusion model to models/diffusion_models/, the Qwen3-VL text encoder to models/text_encoders/, and the Qwen-Image VAE to models/vae/. People routinely skip the encoder and the VAE and then report the model as broken - it needs all three.

Where people get stuck

Everything labeled 🧪 is where the pain lives. The IDENTITY tab and the ControlNet depth/canny modes are experimental, depend on extra custom nodes (comfyui-krea2edit, comfyui-krea2-controlnet, ComfyUI-NK2E, comfyui_controlnet_aux), and need separate LoRA files registered in Settings. The README is refreshingly honest that these "may produce errors or unexpected results" and that the depth/canny control is LoRA-based, not pixel-accurate. Treat them as a bonus, not a feature. The one genuinely nice detail: the depth preprocessor auto-downloads DepthAnything v2 (Large) on first use - though the Giant variant is unsupported because its repo is gated.

Also remember Krea 2's own quirks, which no wrapper can fix. At CFG 1 you need a negative prompt slot even if you leave it empty - the README's Settings has a shared negative prompt for exactly this reason. And if details look mushy at the default sampler, the community consensus is to switch (e.g. to euler with a beta scheduler) rather than fight it.

Is this node the way to try Krea 2? If you want Krea 2 as a simple T2I/I2I/upscale tool and you're happy letting a hidden graph do the wiring, yes - it's the least intimidating on-ramp to the year's best base model. If you came for the ControlNet and identity features specifically, lower your expectations; that's the unfinished quarter of an otherwise solid package.