toobusy Flux2 Klein

Flux2 Klein 9B became the default local image editor of 2026 - the distilled model that does text-to-image, single-reference editing, and multi-reference editing in one checkpoint, at ~4 steps. The catch is that a proper multi-reference Klein workflow in ComfyUI is a wall of nodes: the model loader, the KV cache, the Qwen text encoder, the VAE, three reference-conditioning steps, a guider, noise, scheduler, empty latent, sampler, and decode. toobusy Flux2 Klein folds that entire graph - the README counts thirteen replaced nodes - into one node that takes reference images and a prompt and returns an image.

How it works

Internally it calls the exact native chain: UNETLoader + FluxKVCache (the kv-fp8 caching that stops reference images from being recomputed every pass), CLIPLoader + VAELoader, CLIPTextEncode, three Klein reference-conditioning nodes, then BasicGuider + RandomNoise + KSamplerSelect + Flux2Scheduler + EmptyFlux2LatentImage + SamplerCustomAdvanced + VAEDecode. Your reference_1_image through reference_5_image fill the reference slots; the defaults it ships with - flux-2-klein-9b-kv-fp8.safetensors, the Qwen3 8B text encoder, and flux2-vae.safetensors - are exactly the files the official example uses.

Two behaviors are worth knowing before you touch anything. First, size_mode: from reference (the default) matches reference #1's size and ignores the ratio/megapixel fields - that's the Klein-native behavior. Pick ratio + megapixels to force a 9:16, or manual for explicit width/height. Second, the bundle integration: connect a toobusy_bundle and its empty reference slots get filled from bundle roles, its resolved_prompt replaces the positive text, and its LoRA cards apply - that's how this node plugs into the Reference Board → Prompt Director flow. bundle_reference_order set to auto follows whatever swap flags the Director set.

The inputs that matter

• model_name / clip_name / vae_name - auto-detected with fuzzy scanning; override if it grabs the wrong file.
• positive - the edit instruction. Klein is a Photoshop-command model, not a "dream me a scene" one; "direct, technical, post-processing prompts" beat poetic ones for edits.
• reference_1_image ... reference_5_image - applied in order; reference #1 also drives the default output size.
• size_mode + ratio_preset/megapixels or width/height - see above; the 9:16-with-full-body gotcha is below.
• steps (4), sampler_name (euler), seed - the distilled defaults are genuinely what people run: 4 steps, euler, CFG-1-ish via the native scheduler.
• lora_slots (0–5) and the lora_N_enable/name/strength triplets - Klein tolerates LoRA stacking well; this is the built-in stack.
• use_bundle_prompt / use_bundle_loras (both on) - how much the bundle overrides manual entries.

Outputs and where they go

• image - the result, into a Save Image / preview.
• latent - the sampled latent, for a second pass without re-running the graph.
• model, model_clean, clip, vae, positive - passthroughs for building an external refine loop. Use model when continuing at the same resolution; model_clean is safer when you're changing resolution or swapping LoRAs.

Install and gotchas

Standard pack install (Manager "toobusy", or clone + restart). The three model files are yours to fetch - the pack ships none of them.

Two things bite people. First, full-body vertical shots: Klein loves to crop. The README's fix is to spell out composition in the prompt (full-length, head to toe, entire body visible, both feet on the ground) and pin 9:16 via ratio + megapixels before generating. Second, identity drift on edits is Klein's known weakness - it looks great but doesn't hold a face like Qwen-Image-Edit does; a consistency LoRA is the usual mitigation. And remember the 9B carries the FLUX non-commercial license; if you need commercial, that's the 4B's job.