A custom node for ComfyUI that allows you to easily iterate through all of your downloaded checkpoints sequentially or dynamically. This is extremely useful for testing prompts or settings against a large batch of different models without manually changing the checkpoint loader node each time.

Installation

Navigate to your ComfyUI custom_nodes directory.
```
cd ComfyUI/custom_nodes
```
Clone this repository (or copy the folder into the custom_nodes directory).
```
git clone https://github.com/Ni-zav/ComfyUI-Checkpoint-Iterator.git
```
Restart ComfyUI.

Usage

In the ComfyUI canvas, right-click and go to Add Node -> loaders -> Checkpoint Loader Iterator.
Connect MODEL, CLIP, and VAE outputs into your normal pipeline (KSampler, encoders, etc.).
(Optional) Connect CKPT_NAME to your filename prefix so each saved image includes the model name.
Queue multiple prompts from the main queue controls.

Iteration behavior

This node has no index input now. It advances automatically on each execution.

1st execution -> checkpoint 0
2nd execution -> checkpoint 1
...
wraps with modulo when it reaches the end

Memory behavior

Before loading the next checkpoint, the node unloads previous models and clears cache to reduce VRAM carry-over between checkpoints.

Important note about "batch"

KSampler latent/image batch size is not the same as queue iteration count. To iterate checkpoints, queue multiple executions (prompt queue), not just a larger sampler batch size.

How it works

For each image generated in the batch queue, the Primitive node will increment the batch index by 1. The custom node calculates index % total_checkpoints to select the corresponding checkpoint from your models folder and loads it.

Outputs

MODEL, CLIP, VAE: The standard outputs required for ComfyUI workflows.
CKPT_NAME: (String) The exact filename of the checkpoint currently loaded. You can route this string into a custom "Save Image" node to automatically include the model name in your generated image filenames for easy comparison.

Paired checkpoint evaluation (companion CLI)

The included checkpoint_eval.py provides reproducible operational comparisons between checkpoints using the ComfyUI HTTP API. It does not rate image quality or claim one model is better without a separately designed evaluation.

Export a complete API-format ComfyUI workflow (not a regular UI-only JSON export), using a built-in CheckpointLoaderSimple and a sampler with an editable seed input. Identify both numeric node IDs from the exported JSON.

# Offline: confirm every model gets exactly the same seeds
python checkpoint_eval.py --workflow workflow_api.json \
  --loader-id 4 --sampler-id 8 \
  --checkpoints model-a.safetensors model-b.safetensors \
  --seeds 101 202 --dry-run

# Live: requires a locally running ComfyUI server and the named checkpoints
python checkpoint_eval.py --workflow workflow_api.json \
  --loader-id 4 --sampler-id 8 \
  --checkpoints model-a.safetensors model-b.safetensors \
  --seeds 101 202 --out results.jsonl

# Offline summary of completed, failed, and wall-clock execution times
python checkpoint_eval.py --report results.jsonl

The benchmark deliberately uses CheckpointLoaderSimple, not the automatic iterator: model selection must be explicit and paired to compare runs fairly. It deep-copies the source workflow, pins model and seed per trial, queues jobs, polls history, writes a JSONL row even on failure, and summarizes execution latency and output counts. These are operational metrics, not FID, CLIP, human preference, or model-quality scores. Queue contention, GPU model, caches, warmup runs, backend versions, prompts and generation parameters matter; report them alongside results. Do not publish generated files or results containing private prompts without reviewing them.

python -m unittest discover -s tests -v

No Python dependencies are needed for the evaluator itself; executing real generation still requires ComfyUI and its model dependencies.

License

MIT License