Shima Master Prompt

Shima Master Prompt is the pack's conditioning hub: one node that takes your prompt and produces ready-to-sample conditioning, and it adapts which text encoders it uses based on the model you're actually running. Set model_type to SD1.5 and it uses the single CLIP. Pick SDXL and it layers CLIP-L and CLIP-G with independent weights. Pick SD3 or Flux and it adds a T5 field for natural-language prompting. The same node carries you across three generations of architectures without you rebuilding the graph.

This matters because "CLIP Text Encode" is fine until it isn't. SDXL genuinely benefits from splitting style (CLIP-L) from subject (CLIP-G); Flux and SD3 want a T5 encoder that understands full sentences. Doing that with stock nodes means a different wiring diagram per model. Master Prompt collapses it into one node with the L/G/T5 fields and per-encoder weight sliders, so you get the correct encoder wiring without re-plumbing.

The inputs that matter

• positive / negative - the main prompt boxes, always present. Everything else is supplementary.
• model_type - 14 choices (sdxl, sd1.5, sd2.x, sd3, flux, pony, illustrious, auraflow, hunyuan, lumina2, chroma, hidream, two z-image variants). The boss input that decides which encoders are live.
• positive_l / negative_l (style/detail), positive_g / negative_g (subject), positive_t5 / negative_t5 (natural language) - the per-encoder fields, with clip_l_weight, clip_g_weight, t5_weight sliders to balance them. Leave the encoder fields empty and the main prompt is used for that encoder.
• flux_guidance - guidance scale for Flux/Chroma, auto-applied when model_type is flux or chroma.
• modelcitizen.bndl - plug the Model Citizen bundle here and it pulls the CLIP from it; clip_override / model_override / vae_override give manual overrides.

Outputs are the meat: positive and negative conditioning straight into your sampler, plus CLIP_L_ONLY, CLIP_G_ONLY, T5_ONLY conditioning (handy for debugging or chaining), the resolved vae, the final pos_string/neg_string, and a masterprompt.bndl for the panel system.

Install & what to expect

cd ComfyUI/custom_nodes
git clone https://github.com/KDB-USJP/shima_wf

then restart, or ComfyUI Manager → search "Shima". No models of its own - it encodes with whatever CLIP your Model Citizen (or override) provides.

The honest caveat: the pack's 14 "model types" don't all have fully distinct encoder handling, and some (like the z-image entries) map onto the same flux/16-channel family. If you're on a mainstream model, this behaves well; treat the exotic list as aspirational, not a guarantee. Also note use_commonparams defaults to true, so if you have a Shima.Commons wire attached it overrides model_type - disconnect it if your model type keeps "reverting." Compared to stock dual CLIP encoding, this is a genuine quality-of-life win for SDXL/Flux users, which is most people by now.