K2 Spatial Tuning

K2 Spatial Tuning is the least exciting node in the K2 Lab, and that's fine - it's a settings bag. It outputs a single K2_TUNING object you feed into K2 Compose, and it is entirely optional: Compose uses these exact same defaults when nothing is connected. You only reach for this node when a regional render misbehaves, which is how a tuning node should work.

The knobs tune the spatial attention router, the thing that makes regional prompting possible on Krea 2 at all. Krea 2 is a single-stream model - text and image tokens share one sequence - so there's no second conditioning branch to attach a second character to. The router works at the attention-logit level instead, biasing each region's text tokens toward the image tokens inside its box. This node dials that bias.

The three knobs that matter

• inside_strength - how hard regional text binds to its box. Larger values bind harder; very large values flatten the image. If a subject's attributes are drifting into the background, nudge this up.
• outside_penalty - center-to-edge contrast, and how hard subject text is blocked outside its box. Background regions use a quarter of it.
• late_step_scale - how much spatial strength survives the last denoising steps. Relaxation starts after 55% progress, and this is what stops region edges from looking pasted on. Note it needs the K2 Regional Sampler; a plain KSampler won't drive it.

Then there's strict_isolation, the one boolean that changes the router's character. On (default), subject-owned text tokens are private to their subject and unreachable for image tokens outside the box. Off, it degrades to a soft bias and attributes bleed between subjects. Turning it off is the first thing to try when the hard partition makes two overlapping people look disembodied or stiff.

The rule that's easy to break

spatial_enabled off disables the router completely - which sounds harmless until you remember the router is what keeps regional LoRAs local. If a regional LoRA is active and you switch the router off, its text delta becomes shared scene conditioning and leaks into every region. The node won't let you do it, and for good reason.

Output and wiring

One output, tuning, into K2 Compose. That's the whole wiring story.

Measured advice, if you're going further

The K2 Lab docs are unusually honest about the multi-region ceiling: two to three subjects sit reliably, and at four the model starts staggering people into depth instead of sizing them evenly. The measured help is a landscape aspect, inside_strength around 2.6 with outside_penalty around 2.0, and "shown full length" in every region prompt. The included four-region example workflow ships with exactly those values.

Installing it

Same as the pack: git clone https://github.com/ping1979ping/comfyui-FVMtools into custom_nodes, restart ComfyUI. Zero extra dependencies - this node is pure configuration on top of the K2 runtime.