Flux2Klein Model (CRT AutoDL)

Full FLUX.2 dev needs 18–24GB of VRAM even quantized, which is why Klein exists: BFL's size-distilled answer that put a genuine Flux 2 in around 13GB of VRAM - and the 9B fp8 file this node downloads is the version people actually run. CRTAutoDLFlux2KleinModel grabs flux-2-klein-9b-fp8.safetensors on first use, loads it as a MODEL, and gives you the same three speed knobs as the rest of CRT's AutoDL model loaders.

The slightly odd part of the download deserves a mention: the node pulls from PGCRYPT/OB_FK on HuggingFace (the author's own mirror of the file), and if it finds a file already named OB_FK.safetensors in your diffusion_models folder, it renames it into place rather than re-downloading. So if you've seen that oddly-named file floating around in workflows, this is why.

The three inputs

Same trio as the LTX 2.3 loader, and they do the same thing:

• sage_attention (default auto) - swaps in SageAttention for attention when the package is available. Biggest single lever for a 9B transformer, and the reason auto is the sensible default.
• enable_fp16_accumulation (default on) - fp16 accumulation in matmuls. Leave it on.
• patch_cublaslinear (default off) - enables ComfyUI's --fast cublas linear patch. Marginal; flip it on if you're benchmarking.

Wiring

MODEL out goes into a sampler - the CRT pack has Flux2-specific helpers like Flux2Klein Seamless Tile (CRT) that take it directly, and the standard KSampler path works too. Complete the stack with CRTAutoDLFlux2KleinCLIP (the Qwen3-8B encoder, loaded as the flux2 CLIP type) and CRTAutoDLFlux2KleinVAE, and you've got a local Klein setup with zero manual downloads.

Gotchas

• First queue pulls ~10GB. Console progress bar, no canvas feedback. If it looks frozen, check the terminal for CRT AutoDL before you panic.
• It's Klein, not dev. If a workflow expects the full 32B FLUX.2 dev or a different Klein variant, this node won't satisfy it - the file is specifically the 9B fp8.
• The three knobs are exactly the AutoDL standard. If you've already tuned these on the LTX loader, you know everything here.

Install once through ComfyUI Manager (CRT-Nodes) or git clone https://github.com/plugcrypt/CRT-Nodes.git into custom_nodes, pip install -r requirements.txt, restart. The model download triggers on first queue.