Stepwise Latent Loader

StepwiseLatentLoader is the quiet utility of the AsymmetricQuantSampler pack: it reads a .pt file that StepwiseKSampler saved and hands you back a LATENT, so a latent from any denoising step is back inside your graph. That's the whole job, and it's a useful one - it's what lets you take step 3 of one run and continue sampling from there with a different model, a different sampler, or a different switch point, without re-running the first three steps.

The inputs are three. save_dir is the folder (wire it to StepwiseKSampler's save_dir output, or paste the path). step_number picks the step, default 1 - the node looks for step_0001.pt-style files, so step 5 means step_0005.pt. And load_type is a two-way choice: noisy_sample loads the actual latent the sampler was carrying at that step, while x0_prediction loads the model's denoised prediction for that step instead. For continuing a generation you want noisy_sample - that's the state you'd feed forward. For inspecting what the model thought the image was at that point, x0_prediction. If the exact step file doesn't exist it falls back to step_final.pt, and if that's missing too, you get a clear error. Output is a single latent you can wire into any node that takes one.

The natural move: run StepwiseKSampler once with save_steps="all", then feed StepwiseLatentLoader's output into a DualModelKSampler and swap in a different model for the remaining steps. You've turned a one-shot analysis run into a jumping-off point. A couple of honest caveats: the loaded latent comes with no conditioning attached - that's still your job - and resuming isn't a free lunch. The step index and the noise schedule interact, so continuing from "step 7" means telling the next sampler to start there, and it's easy to get subtly wrong results if the schedules don't line up. For casual inspection and single-step experiments it's great; for perfect resumption you'll want to think about what denoise and start_step actually do.

It's the pack's least glamorous node and it knows it. Install is identical to the rest: ComfyUI Manager → search "Asymmetric Quantization Sampler", or:

cd ComfyUI/custom_nodes/
git clone https://github.com/lee09lee26/ComfyUI-AsymQuantSampler.git

Restart and it's registered. No extra pip dependencies - ComfyUI built-ins only - so there's nothing to go wrong on install. If you're not doing step-level analysis you'll never touch it; if you are, it's the missing link that keeps your experiments in the graph instead of in a Python script.