Krea 2 Turbo Model (CRT AutoDL)

This is the checkpoint most people actually mean when they say "Krea 2" day to day. Run the node, it fetches Krea 2 Turbo, and you get a MODEL back - no manual download, no hunting for the right HuggingFace repo.

Why Turbo, specifically

Krea 2 shipped as two checkpoints in June 2026, and Turbo is the one built for generation rather than training. It's distilled from Raw down to 8 steps via a technique called Trajectory Distribution Matching, runs with guidance disabled at official settings (mu 1.15), and renders up to 2K. If you just want to type a prompt and get a good-looking Krea 2 image without thinking about training curricula, this is the checkpoint.

Two things worth knowing before you use it. First, the community's actual preferred configuration is neither Turbo alone nor Raw alone - it's Raw with the separately-published Turbo LoRA applied at around 0.6 weight and roughly 12 steps on euler_cfg_pp with the beta scheduler, reported as clearly better than running Turbo by itself. This node gets you straight Turbo; if you want that hybrid setup you'll need the Raw model loader plus the Turbo LoRA as a separate download. Second, like Raw, Turbo carries the same alignment-training gap from the hosted Krea API - expressions and prompt clauses about bodies or violence get silently dropped locally in a way they aren't on Krea's own website, which Krea's team has confirmed is intentional alignment work applied only to the open release.

The inputs and outputs that matter

Same three optimization toggles as the rest of the AutoDL model loaders:

• patch_cublaslinear (default false) - cuBLAS-routed Linear layers for faster matmuls.
• sage_attention (default auto) - disabled, auto, or a named SageAttention kernel variant. auto is the safe default unless you know your GPU generation and want to force a specific kernel.
• enable_fp16_accumulation (default true) - fp16 matmul accumulation for a speed gain on supporting GPUs.

Output: a single MODEL.

Installing it

Bundled with CRT-Nodes, no separate install:

ComfyUI Manager - search CRT-Nodes, install, restart.

Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/PGCRT/CRT-Nodes.git
pip install -r requirements.txt

Restart ComfyUI. The Turbo weights download the first time you run the node - a community fp8 conversion of Krea 2 brings the model down to around 12GB, though whether this AutoDL loader fetches full-precision or a quantized build isn't documented in the node's schema, so treat the first-run download size as something to watch in your console rather than assume.

Common issues

• Output looks a bit generic or less punchy than screenshots you've seen online - that's very likely the Raw+Turbo-LoRA hybrid setup people are actually running, not plain Turbo. If you want that look, you need Krea 2 Raw Model (CRT AutoDL) plus the separately-published Turbo LoRA, not this node alone.
• CFG slider seems to do nothing - Turbo's official setting runs with guidance disabled; that's expected behavior for a distilled checkpoint, not a broken widget.
• Expressions look flat, or prompts about bodies/violence get ignored - the documented alignment gap between the open weights and Krea's hosted API, not something specific to this loader. It affects both Raw and Turbo equally.
• First run is slow - that's the model download; give it time before assuming the node has failed.