Prompt Phase Splitter (Arctenox's Essentials)

Here's the pitch, straight from the README: split generation into three temporal phases - composition, detail, texture - each with its own prompt, weight, and step count. That sounds like per-step prompt scheduling, which is a genuinely useful thing to have.

Here's what the source actually does: it's a prompt blender, not a scheduler. The percentage fields (composition_end_percent, detail_start_percent, and friends) are descriptive - they show up in the phase_info string, but nothing in the node consults them against sampling steps. What really happens is the phase prompts get merged into a single conditioning before it ever reaches your KSampler. If you bought this hoping for true temporal conditioning, you'll be disappointed. If you use it as "organize one prompt into three weighted chunks," it's fine.

How it works

With a CLIP model attached, the node resolves wildcards in every phase prompt using a seeded RNG (wildcard_seed −1 = randomize every run), then combines the active phase prompts with your base conditioning. The blend_mode chooses how:

• concat (default) - joins all phase prompts into one string, prefixed masterpiece, best quality, and encodes it once. Honestly? This is "build me a longer prompt" in a trench coat.
• append - adds each phase as its own conditioning entry, the way you'd append conditionings in a vanilla graph.
• add / subtract - does tensor arithmetic on the conditioning embeddings, scaling each phase by its weight. This is where the node actually earns the "splitter" name, and where the weights do real work.

The negative side passes through untouched - you get base_negative back exactly as you fed it.

The inputs that matter

base_positive and base_negative come from your CLIP Text Encode nodes, and you also feed a clip in. That clip input is the whole mechanism: if it's unconnected, the node shrugs and returns your base conditioning unchanged with a "No CLIP provided" note.

For each of composition, detail and texture there's an enable toggle, a prompt, a weight (0–3) and percent boundaries. The three enable toggles and weights are what you'll actually touch; the percent fields are mostly for your own documentation. There's also an off-by-default enable_character phase for subject emphasis, and a standalone wildcard_prompt that gets appended to composition.

The outputs

• positive - the combined CONDITIONING, wired into your KSampler.
• negative - your base negative, passed through.
• phase_info - a human-readable breakdown of which phases are active and at what weights. Worth glancing at once so you know what you're actually sending to the sampler.
• resolved_wildcards - shows what each wildcard resolved to, handy when a generation surprises you.

Install

ComfyUI Manager, search Arctenox's Essentials, install, restart. Or:

cd ComfyUI/custom_nodes/
git clone https://github.com/Arctenox/Arctenoxs-Essentials_ComfyUI

Restart, and you're done - no model downloads, just torch, numpy and optional psutil.

Gotchas

• Set expectations: default concat mode is basically prompt concatenation. Use add or subtract if you want the weighted-blend behavior the phase framing promises.
• If the output never changes, check that clip is connected. It's the difference between "works" and "silently passes through."
• The whole pack is marked DEPRECATED - Arctenox is folding it into a new combined pack. Fine for a workflow you're running today, not something to build long-term infrastructure on.