EditUtils: Flux2Klein Model Config lrzjason

Flux2KleinModelConfig_EditUtils is the "I'm editing with Flux 2 Klein, treat it accordingly" handshake node for this pack. One optional instruction input, one model_config DICT output. Everything the encode pipeline needs to know about the model arrives in that one wire, and it's a blessedly small set of facts.

What's actually in the dict, since the node keeps it invisible: model_name: "flux2klein" - the routing key that sends EditTextEncode_EditUtils down the Flux2Klein code path; vae_unit: 16 - the VAE spatial unit for the Klein family, which is not the same 8 that Qwen, Boogu, and Krea 2 use, and getting this wrong is exactly the kind of thing that produces misaligned padding; and llama_template - which stays empty unless you type something into instruction.

That instruction field deserves a real warning, because this is where Flux2Klein differs from every other config node in the pack. For Qwen and Krea 2, the instruction builds a custom system prompt and that's the point. For Flux 2 Klein it's the opposite: the Klein tokenizer works best with its own built-in prompt handling, so the recommended state is instruction left blank, which yields an empty llama_template. The source code literally comments that setting it is "not recommended but allowed for customization." The honest reading: it's a back door for experimentation, and the pack author doesn't think you'll beat the defaults. Start blank.

You can think of this node as the per-model key in a switch statement. Swap Flux2KleinModelConfig_EditUtils for BooguModelConfig_EditUtils or Krea2ModelConfig_EditUtils on the same model_config wire and the whole encode chain re-targets itself - same EditTextEncode_EditUtils, same configs chain, different model. That's the unified-interface design this pack is built around, and it's why these one-field config nodes are worth having rather than just typing flux2klein into a text box.

The context you're slotting this into, if you're newer to Klein: it's Black Forest Labs' small distilled line (4B and 9B, the 4B Apache 2.0), released January 2026, and it does text-to-image, single-reference and multi-reference editing in one checkpoint. This pack's job is the edit plumbing around it - the weights, the Qwen3 text encoder, and the VAE are all yours to download. Install: ComfyUI Manager (search "EditUtils") or git clone https://github.com/lrzjason/ComfyUI-EditUtils into custom_nodes, restart. Zero dependencies beyond stock ComfyUI.