4-Prompt Model Cycler Loader (Nukun)

If you read the 4-Prompt Checkpoint Cycler Loader article, you know the shape: cycle four prompts, step through every model, keep the seeds honest. This is the same machine for a different model type - it cycles diffusion models out of the diffusion_models/ folder, with the CLIP and VAE as separate, fixed inputs. It's the batch-comparison node for FLUX, SD3, and every workflow where the weights live in separate files.

The two loaders look nearly identical because they're siblings sharing a design. Pick the checkpoint version for monolithic .safetensors, this one for split-model workflows.

How it works

cycle_index (keep it on increment) walks through four multiline prompts - text_1 through text_4 - for the first naturally-sorted model in your chosen unet_folder, then moves to the next model and restarts at prompt 1. Empty text fields are legitimate cycle slots, not mistakes. prompt_index, model_index, and model_count come out alongside the active text, so you can encode position in filenames.

The difference from the checkpoint version is in the fixed companions:

• clip_name - one text encoder shared by every model in the cycle.
• vae_name - defaults to pixel_space, shared across the run.
• weight_dtype (default, or the three fp8 options) - how the core loader handles quantization for whatever model is currently selected.
• clip_type - architecture of the CLIP file (stable_diffusion, sd3, ltxv, wan, ...).
• clip_device - normal or forced CPU for the encoder.

The seed modes (same as the checkpoint version)

seed_mode = increment, fixed, or random. The honest summary: increment advances the seed once per model (same seed across that model's four prompts), fixed pins one seed to everything, and random rolls a fresh seed per model-group - with the caveat that the fresh rolls happen when the frontend queues a new group, so API-only queues use your entered seed as-is.

That seed discipline is the actual point of the node. A "does this model suite handle this prompt set?" test is only meaningful if each prompt gets a fair shake per model, and these three modes are the three standard ways of defining "fair".

Outputs

text, MODEL, CLIP, VAE, modelname, filename_modelname, unet_name, folder, prompt_index, model_index, model_count, seed. Wire text to your text encoder, everything else to the sampler and filename builder.

Installing it

cd ComfyUI/custom_nodes
git clone https://github.com/OnekoSL/Nukun_ComfyUI_Nodes.git

Restart ComfyUI or install via Manager ("Nukun").

Where it sits in your workflow

This one is aimed squarely at people running multi-model comparisons on FLUX/SD3-class architectures, where "just batch over checkpoints" isn't an option. If you're on SDXL/Pony/Illustrious and only use monolithic checkpoints, you don't need this - the checkpoint 4-prompt cycler is the simpler fit. But if you maintain a folder of quantized diffusion models and want to sweep prompts across all of them with clean per-model seeds, this is the node that does the bookkeeping for you.