Krea 2 Raw Model (CRT AutoDL)

This node fetches one specific checkpoint - Krea 2 Raw - and hands you back a MODEL, no file browsing required. Krea 2 shipped in June 2026 as two checkpoints, Raw and Turbo, and picking the right one matters more here than with most model pairs: they're built for genuinely different jobs.

Raw vs. Turbo, and why Raw is the one you train on

Krea 2 Raw is the undistilled base - no aesthetic finetune applied at all, which Comfy Org's own team called something they hadn't seen a modern base model ship as before. Krea's own reasoning: the open-source community lacks quality base models to train on, so they shipped one with nothing baked in, specifically so LoRA training and further finetuning have something clean to work against. Raw's official settings are 52 steps at CFG 3.5 - noticeably heavier than Turbo's 8-step, CFG-disabled defaults - and the community's converged practice is to train LoRAs on Raw and then apply them to Turbo at inference, which is also Krea's own stated guidance.

Worth knowing before you invest time in this: the open weights went through a licensor-acknowledged alignment pass the hosted Krea API version never got. That means prompt clauses about bodies, violence, and strong emotional expression get silently dropped rather than degraded - Krea's own head of research confirmed this directly, responding to a user who found the local model much worse at prompt-following than the API. Two community bypasses circulate (a conditioning-rebalance node and an uncensor LoRA); the uncensor LoRA is generally preferred because it costs less quality. None of that affects Raw's specific role as a training base, but it's the thing to know before you're confused by why local output doesn't match what you've seen from Krea's website.

The inputs and outputs that matter

Three performance toggles, same pattern as the other CRT AutoDL model loaders:

• patch_cublaslinear (default false) - routes Linear-layer matmuls through cuBLAS for speed, the loader-level version of ComfyUI's --fast cublas_ops flag.
• sage_attention (default auto) - disabled, auto, or a specific named SageAttention kernel (sageattn_qk_int8_pv_fp16_cuda, sageattn3, and others). Leave it on auto unless you know exactly which kernel your GPU supports.
• enable_fp16_accumulation (default true) - trades a sliver of precision for speed on GPUs that support fp16 matmul accumulation.

Output: a single MODEL.

Installing it

Ships with CRT-Nodes - no separate download for the node code itself:

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 actual Krea 2 Raw weights (a 12B model - roughly 24.76GB in bf16) fetch on first use of the node, not at pack-install time; this is the biggest of the AutoDL family's downloads, so expect a real wait on first run.

Common issues

• Using Raw for a quick generation and getting a slow, mediocre result - that's expected; Raw has no aesthetic finetune and needs its full 52-step, CFG 3.5 settings to look its best. If you just want fast, good-looking images, use Krea 2 Turbo Model (CRT AutoDL) instead - Raw's job is training, not daily generation.
• Local output feels flatter or less expressive than what you've seen from Krea's website - that's the alignment-training gap, not a bug in this node. It's a documented difference between the open weights and the hosted API version.
• Out of VRAM on first load - Krea 2 is a 12B model; a full-precision load needs real headroom. If you're VRAM-constrained, this auto-downloaded checkpoint won't automatically be quantized for you - you'd need a manually-sourced fp8 or GGUF build instead.
• Download stalls on first run - give it time; this is the largest checkpoint in the AutoDL family. Check your console for active HTTP transfer before assuming it's hung.