Shima Panel Master Prompt

Shima Panel Master Prompt is the conditioning half of Shima's Panel workflow, and it's where the pack's prompt philosophy actually lives. On the canvas it's a compact PCB chassis with one masterprompt.bndl output; double-click it and you get the real interface: positive and negative text fields, a model_type dropdown, and - for the architectures that support them - separate CLIP-L, CLIP-G, and T5 encoder fields with independent weight sliders.

Here's the idea worth caring about. Modern models don't have one text encoder. SDXL uses CLIP-L (great at style and keywords) plus CLIP-G (great at subject and composition); Flux and SD3 add T5, a language-model encoder that actually understands paragraphs. Most workflows just mash one prompt through one encoder and leave quality on the table. This node lets you split the work: main prompt in the box, style keywords in CLIP-L, subject description in CLIP-G, complex scene text in T5 - each with its own weight slider (0–10) so you can dial how much each encoder's voice counts.

The model_type dropdown (SDXL, SD 1.5, SD3, Flux, AuraFlow, HunyuanDiT) decides which encoder fields are active, so you're never staring at T5 inputs on a SD1.5 graph.

How it works

It's the panelized version of the full Master Prompt node. It encodes the prompt fields into conditioning using the CLIP from the connected Model Citizen bundle (it can also pull a CLIP from the optional modelcitizen.bndl input directly), and outputs the resulting positive/negative conditioning. It also passes the vae through as its own output - a small convenience so the sampler can get VAE without a separate wire. It respects Shima's Common Params: turn on use_commonparams (the green/red circle in the topbar) and settings like model type can be driven by the shared Commons node.

The inputs and outputs that matter

• positive, negative - the main prompt fields (always present).
• model_type - picks the encoder set.
• clip_l_weight, clip_g_weight, t5_weight plus the positive_l/negative_l, positive_g/negative_g, positive_t5/negative_t5 text fields - the multi-encoder inputs. Leave the encoder-specific fields empty and only the main prompt feeds that encoder.
• flux_guidance and lumina_sysprompt - architecture-specific extras for Flux-family and Lumina models.
• model_type_override, model_override, clip_override, vae_override - escape hatches to force values when the graph's model_type would otherwise win.
• panelinputs.bndl - lets a Shima Panel BNDLer drive the prompt settings externally.

Outputs: positive, negative (CONDITIONING), vae (VAE), and masterprompt.bndl - the bundle that feeds the Panel Sampler.

How to install it

It's in the Shima pack. ComfyUI Manager → search "Shima", or:

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

restart. Core deps (Impact Pack, cg-use-everywhere) install automatically. No extra models.

Common issues & troubleshooting

Panel renders broken after a ComfyUI update. Frontend-heavy node - the Nodes 2.0 (Vue) frontend breaks these JS chassis. Run the legacy LiteGraph canvas.

CLIP-L/G/T5 fields seem to do nothing. Check model_type - on SD1.5 the extra encoders aren't wired at all. And remember the weight sliders: setting a field to weight 0 makes it contribute nothing regardless of what's typed.