Realtime LoRA Trainer (FLUX Klein - Musubi Tuner)

What it is

FLUX Klein is Black Forest Labs' small, fast, distilled Flux 2 line, and it's been the default local image editor since it shipped in January 2026 - four checkpoints at once, split across size (4B/9B) and distillation (base/distilled). This node trains on the base variants: Klein Base 4B or Klein Base 9B, the undistilled ones BFL says explicitly are not meant for generation, only for fine-tuning. Whatever LoRA comes out gets applied at inference against the distilled 4B or 9B model, same pattern as Z-Image Base training on this pack.

Why bother: Klein tolerates LoRA stacking noticeably better than Z-Image (community reports of three LoRAs at full strength on Klein 9B without the quality collapse Z-Image shows past two), and its 4B variant is Apache 2.0 - the only Flux 2 weight you can build a commercial LoRA on without a BFL licensing conversation. Training landed in every major trainer within three days of Klein's release, which tells you the ecosystem cared.

How it works

Musubi Tuner trains against the Klein base checkpoint, using the Qwen3 text encoder Klein was built around (Qwen3-4B for the 4B model, Qwen3-8B for the 9B) and the FLUX.2 VAE. The node's blocks_to_swap control here is deliberately separate from resolution and precision settings, so you can dial VRAM usage independently of image quality - useful given how much heavier the 9B variant is than the 4B.

Inputs and outputs that matter

• model_variant - Klein Base 4B or Klein Base 9B. This decision drives everything downstream: the 4B needs a Qwen3-4B text encoder and roughly 13GB VRAM footprint, the 9B needs Qwen3-8B and closer to 29GB.
• dit_model, vae_model, text_encoder dropdowns, filled once you've downloaded the base (undistilled) checkpoint - not the distilled generation model - plus the matching Qwen3 encoder and the shared FLUX.2 VAE (ae.safetensors, the same file across all FLUX.2 variants).
• inputcount plus image_N / caption_N pairs, or images_path for a folder.
• training_steps (400), learning_rate (0.0001), lora_rank (32) - all deliberately lower/higher than this pack's older-architecture trainers, reflecting how sensitive current models are to a hot LR.
• blocks_to_swap (0–16) as an independent VRAM lever, plus vram_mode with eight presets from Max (1256px) down to Min (512px), several offering fp8 and fp8-offload variants.

Output is a single lora_path string, ready for ApplyTrainedLora.

Installing it

cd ComfyUI/custom_nodes
git clone https://github.com/ShootTheSound/comfyUI-Realtime-Lora

Restart ComfyUI, install Musubi Tuner separately, and set musubi_path. From Hugging Face's black-forest-labs org, download flux-2-klein-base-4b.safetensors or flux-2-klein-base-9b.safetensors - the base, undistilled versions - plus ae.safetensors and the matching Qwen3-4B or Qwen3-8B text encoder. Dedicated example workflows for both sizes ship in the pack's workflows/Training/ folder.

Common issues

Downloaded the distilled model by mistake. BFL ships four Klein checkpoints and it's easy to grab the wrong one; only the base (undistilled) variants train correctly here.

9B running out of VRAM. The 9B is genuinely large - lean on blocks_to_swap first, then drop into an fp8 vram_mode preset, before concluding your card can't do it.

Applying a 4B LoRA to the 9B model or vice versa. The two sizes are not interchangeable; a LoRA trained on Klein Base 4B only applies to Klein 4B at inference, and the same goes for 9B. Publish and load for the size you actually trained on.