CLIP Text Encode SDXL v3 (Global + Enhanced)

If you've seen BlenderNeko's BNK_CutoffRegionsToConditioning, you know the two-text-per-region idea: one text field is what you see (the target), another is what the model actually encodes (the region text). This node is that model, rebuilt ergonomically and fully self-contained - no ComfyUI_Cutoff plugin needed.

Each section in this v3 encoder has two fields. section_N_global_text is a chunk of your natural prompt - the plain text you'd paste anywhere. Comma-joined, all the global texts form your user-visible prompt, and they double as Cutoff's target text: space-split into words, each word gets masked in other sections' encoding passes. section_N_enhanced_text is the expanded version that the model actually encodes at this section's position - the region text. If enhanced_text is empty or identical to global_text, the section is a passthrough chunk: contributes tokens to the base prompt, creates no region.

Why two fields? Because you can write a short readable global prompt ("a weathered lighthouse") and an expanded region text ("a weathered lighthouse with peeling white paint and a rusted lantern room") that gives the model more to work with while the isolation targets stay clean.

Per-stream control

The node runs Cutoff independently per CLIP stream (L and G), and each section has clip_l_strength and clip_g_strength widgets (default 1, range -10..10). Set one to 0 and that section is excluded from that stream entirely - a cleaner way to do "this goes only to G" than the old (tag:0) tricks. The underlying cutoff knobs (strict_mask, start_from_masked) are hardcoded to 1.0/1.0. The author tried exposing them in the earlier node, users got confused, so they're not exposed here - and 1.0/1.0 is what worked in practice anyway.

Inputs and outputs

clip, section_count (0..32, default 2), the per-section *_global_text / *_enhanced_text / *_weight / *_clip_l_strength / *_clip_g_strength widgets, plus the shared settings: join_separator (default ","), zoom/offset_x/offset_y for the SDXL size/crop metadata, upscaled_conditioning_multiplier for two-stage workflows, and the A1111 embedding toggles. Optional latent input sets target W/H.

Outputs are the same triple as the rest of the enhanced family: conditioning, upscaled_conditioning, and reference_full_prompt (STRING - dump what CLIP actually encoded).

Install

ComfyUI Manager → search "Unified Conditioning", or:

cd ComfyUI/custom_nodes
git clone https://github.com/RandyHaylor/unified-conditioning-combine-with-timestep-range.git unified-conditioning-merge

Restart ComfyUI; the node is under unified-conditioning-merge. No models, no pip deps.

Troubleshooting

• enhanced_text must contain all the words in global_text - the node validates this and surfaces warnings via a realtime widget. A mismatch is the most common setup mistake.
• If section_count is smaller than the number of sections you've filled, the extra widgets are hidden but their values persist in the workflow JSON - re-expand and they come back.
• Multi-section output is combine-style (parallel branches). Keep that in mind if you chain it into a flat merge downstream - the pack blocks that configuration for good reason.