Shima Model Citizen

Shima Model Citizen is the pack's unified model loader: it loads a checkpoint plus its CLIP and VAE, and stacks up to three LoRAs on top, in one node. Instead of CheckpointLoader, three separate LoraLoader nodes, and a VAE loader strung together with six wires, you get a single node with MODEL, CLIP, and VAE outputs - plus a bonus name_string that tells you what's actually loaded.

The mechanism is exactly what a good loader should be: it calls ComfyUI's own load_checkpoint_guess_config with output_vae=True and output_clip=True, so it works across SD1.5 and SDXL checkpoints alike and keeps the baked VAE unless you tell it otherwise. Then it applies your three LoRA slots through load_lora_for_models in order, and finally builds a display string like "juggernautXL + Detailer + Lighting" so you can see at a glance what's in the pipe.

The inputs that matter

• ckpt_name - your checkpoint. vae_name - defaults to "Baked VAE"; pick an external VAE from the dropdown if you want to override the baked one (some photoreal models like a separate fp16 VAE).
• lora_1_name / lora_2_name / lora_3_name with strength each (range −10 to 10) - your LoRA stack, right on the node. Set a name to "None" or a strength to 0 and it's skipped.
• filter_by_model_type - a dropdown (All/SDXL/SD1.5/SD3/Flux/SD2.1) that trims the checkpoint list so you don't scroll past 80 SD1.5 models looking for your SDXL one. Pure convenience, doesn't enforce anything.
• lora_stack - an optional chained stack input, so you can add more LoRAs with the pack's Lora Stack node and have them appended after the built-in three.

Outputs: MODEL, CLIP, VAE, name_string, and modelcitizen.bndl (a bundle the panel system and Master Prompt can consume). The bundle is the interesting one - in an island workflow you run one wire instead of three.

Install

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

then restart, or ComfyUI Manager → search "Shima". The node needs no downloads itself, though the pack's official-dependencies bundle lists checkpoints like AbsoluteReality and JuggernautXL for the official "Finals" workflows if you want the full experience.

Gotchas

The classic trap: use_commonparams defaults to true, so if you connect a Shima.Commons wire it can override your ckpt/LoRA choices from the bundle - when your carefully picked checkpoint "keeps resetting," that's why. Also, negative LoRA strengths are allowed and occasionally intentional, but mostly a way to shoot yourself in the foot; if a LoRA looks "wrong," check you didn't leave a −1 from a previous experiment. filter_by_model_type filters the dropdown, not your actual selection - if you switch the filter, your currently loaded model stays loaded until you pick again. For a template workflow or an island, one-node loading genuinely is easier to keep consistent than a seven-node loader sprawl. For a one-off, the stock loaders are fewer dependencies. Your call.