Conditioning Pipeline (Combine)

Prompts are global. Ask a diffusion model for "a red-haired woman on the left, a blonde man on the right" and you're rolling dice: sometimes it lands, more often their features swap, merge, or one of them just quietly ceases to exist. That's attribute bleeding, the oldest problem in multi-subject generation, and the standing fix is regional conditioning - split the canvas into areas and give each area its own prompt. Conditioning Pipeline (Combine) is the merge point of that fix. It takes the area pipeline you built plus your global prompt and hands the sampler finished positive and negative conditionings. Nothing upstream matters if you skip this node.

How it works

The node doesn't do any sampling itself. It reads the CONDITIONING_PIPELINE list built by its sibling Conditioning Pipeline (Set Area), converts each area's normalized coordinates into a mask on a 64×64 grid, then expands itself into a chain of ComfyUI's built-in conditioning-properties nodes behind the scenes. First entry sets the first area; each following entry combines onto the running result; then your global prompt gets combined in as a full-image entry; and finally a default-combine pass acts as a safety net for any pixels no area covered. You never see that sub-graph - it's constructed at execution time - but knowing it exists explains the failure modes below.

The inputs that matter

Four of them, and only one is a dial:

• global_positive / global_negative - the ordinary conditioning from your CLIP Text Encode. Keep the positive minimal and general ("a cozy cafe, daylight") and let the areas carry the subject detail.
• pipeline - the chained pipeline_out from your last Set Area node.
• global_strength - default 0.3, range 0–10. This is the one people overdrive. The author's tested rule of thumb: keep it below 0.5. Push it higher and the global prompt starts washing out your per-area control - you get a nicely lit image where both characters look like neither of them.

Outputs

• positive_out / negative_out (CONDITIONING) - wire these into your KSampler, or into ControlNet first if you're guiding composition with poses.
• areas_out (CONDITIONING_AREAS) - a structured list of every area's x, y, width, height and strength. Its intended consumer is the author's companion ComfyUI-OpenPose-Studio pack, which mirrors these areas straight into the pose editor so your OpenPose poses land exactly where you conditioned. Any node that accepts area metadata can use it too.

Install

This ships inside andreszs/ComfyUI-Lora-Pipeline. Easiest via ComfyUI Manager - search "LoRA Pipeline" - or:

cd ComfyUI/custom_nodes
git clone https://github.com/andreszs/ComfyUI-Lora-Pipeline

Restart ComfyUI. You'll find the nodes under LoRA Pipeline → Conditioning. Python 3.10+ is required; these three conditioning nodes have no extra dependencies (the optional matplotlib install in the README is only for the scheduled LoRA loader's curve preview). Heads up: the pack is brand new - v1.1.2 shipped mid-2026, near-zero community footprint - so you're early, and there's not much of a support trail if you get stuck. The README is unusually thorough, start there.

Common issues

• You feed only one stream. The README calls this out explicitly - both global_positive and global_negative are required, so give the negative node something to work with.
• Empty pipeline. If the pipeline is invalid or has no entries, the node silently falls back to passing the globals through and areas_out is an empty list. If your areas "disappear", trace the chain back - you probably forgot to connect the last Set Area into pipeline.
• Washed-out areas. Almost always global_strength cranked too high, or every strength maxed at once. Lower area strength improves overall quality but weakens per-area control; balance all three knobs instead of going to eleven.
• Forgetting the outputs exist. Build all the areas, then never wire positive_out/negative_out into the sampler path. Classic.

One more honest note: regional conditioning only holds subjects apart if the model listens to masks, and composition still wanders. The author says flat-out that ControlNet + OpenPose is strongly recommended for consistent multi-subject layouts - treat this node as the conditioning half of a two-part setup, not the whole answer.