Sample Predictions

Everything else in ComfyUI-Prediction is a prediction you build; Sample Predictions is where you actually use one. It's the pack's replacement for the KSampler / SamplerCustom - same job, but instead of taking a model plus positive and negative prompts, it takes a single PREDICTION (your whole prediction graph) and denoises with it. If you're new to this pack, this is the node that wires everything together, and the README's setup steps are worth following exactly once.

Inputs

• model (MODEL) - the checkpoint, same as any sampler.
• add_noise (BOOLEAN, default true) - set false for img2img partial-denoise, exactly like SamplerCustom.
• noise_seed (INT) - the seed for the initial noise.
• sampler (SAMPLER) - from KSamplerSelect, which lives under sampling > custom_sampling > samplers.
• sigmas (SIGMAS) - from a scheduler like BasicScheduler, under sampling > custom_sampling > schedulers. The README is explicit: these are not in the "sigmas" menu. People hunt for them there for ages.
• latent_image (LATENT) - your empty latent or source image.
• noise_prediction (PREDICTION) - the final output of your prediction chain. This is the input that makes the node unique.

Outputs

Two LATENTs, mirroring the core SamplerCustom: output and denoised_output. Wire output to a VAEDecode and you're done. denoised_output is the model's estimate of the clean latent, useful for previews or post-processing.

The canonical wiring

The README's five steps, condensed:

1. Add Sample Predictions as your sampler.
2. sampler ← KSamplerSelect.
3. sigmas ← BasicScheduler (from the schedulers menu).
4. CLIP Text Encode (Prompt) → Conditioned Prediction for each prompt; chain predictions through the pack's nodes (Combine, Switch, CFG Prediction, whatever you need).
5. Final prediction → noise_prediction.

If your chain is just a single Conditioned Prediction, congratulations: you've built a sampler with no negative prompt - effectively CFG at 1. Add a CFG Prediction or the combine-based equivalent and you've rebuilt a KSampler in graph form.

How it works

Under the hood it's the pack's own sampler loop wrapping ComfyUI's model API: it resolves all the conditionings referenced anywhere in your prediction graph, batches the model evals the predictors ask for, runs your chosen sampler over the sigmas, and hands back the latents. That's why your prediction nodes don't each need their own sampler - they're all just describing what to compute at each step, and this node executes it.

Installing it

Ships in the ComfyUI-Prediction pack by @RedHotTensors (Project RedRocket). ComfyUI Manager → search "ComfyUI-Prediction", or:

cd ComfyUI/custom_nodes
git clone https://github.com/redhottensors/ComfyUI-Prediction

Restart ComfyUI. No extra dependencies or model files. It's under Add Node > sampling > prediction.

Common issues

• Feeding it a KSampler-style setup. It doesn't take positive/negative CONDITIONING inputs; you must build a PREDICTION chain first.
• Missing sampler or sigmas. You have to provide them from the custom_sampling menus - the node won't invent defaults.
• ControlNet isn't supported by the pack, so any ControlNet you've wired into the model won't apply. Regional prompting is untested per the README.