ComfyUI-CachedSearch

ComfyUI-CachedSearch

Best-of-N seed search for video (and image) models at about 60% of the cost. The N drafts are sampled with ComfyUI's native EasyCache, a verifier scores them, and only the winning seed is sampled again at full compute. Sampling is seed-deterministic, so the latent you get is exactly what full best-of-N would have delivered whenever both pick the same seed.

Paper: CachedSearch: Training-Free Cached Exploration for Test-Time Search in Video Diffusion (Saini, Birkbeck, Wang, Adsumilli, Bovik). Python package for diffusers: pip install cachedsearch. Blog: Explore every candidate under caching, spend full compute only on the winner.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/shreshthsaini/ComfyUI-CachedSearch
pip install -r ComfyUI-CachedSearch/requirements.txt   # image-reward, for the default verifier

Restart ComfyUI. Requires a ComfyUI build that ships the native EasyCache node (comfy_extras/nodes_easycache.py).

Nodes

CachedSearch Sampler (best-of-N, cached drafts), category sampling/CachedSearch. Drop-in for KSampler, with these extra inputs:

| Input | What it does | |---|---| | vae | Decodes each draft so the verifier can score it | | prompt | The text the verifier scores against (the same prompt you encoded into positive) | | verifier | A CS_VERIFIER, for example from the ImageReward node below | | search_width | N. Candidates use seed, seed+1, ..., seed+N-1 | | cache_threshold, cache_start_percent, cache_end_percent | EasyCache settings for the drafts. Higher threshold skips more steps. 0.2 is the EasyCache default; calibrate per model family | | commit | Re-sample the winner with caching off (recommended). Off returns the cached draft itself |

Outputs: the winning LATENT, the winner_seed (feed it to a plain KSampler to reproduce the result), and draft_scores as a JSON string.

CachedSearch ImageReward Verifier: the paper's verifier, ImageReward averaged over uniformly spaced frames. Any node that outputs a CS_VERIFIER object with a score(images, prompt) -> float method works in its place.

Wiring

Model loader, CLIP text encodes, and an empty latent as usual, then:

CachedSearch ImageReward Verifier ──┐
Model / positive / negative / latent / VAE ──> CachedSearch Sampler ──> VAE Decode ──> Save

Type the prompt into the sampler's prompt box as well, so the verifier sees it. Do not attach the EasyCache node yourself; the sampler applies it to the drafts and samples the winner without it.

Choosing settings

• search_width 4 to 8. Recommitting the winner pays off from about 4 candidates.
• cache_threshold: the paper calibrates one threshold per model family. Start at the EasyCache default (0.2) for Wan; lower it if the draft ranking disagrees with your eye, raise it for cheaper drafts.
• Every draft is VAE-decoded for scoring, which costs time and memory on video models. On small GPUs reduce search_width rather than frames_to_score, since the VAE decodes the whole clip either way.

Citation

@article{saini2026cachedsearch,
  title   = {CachedSearch: Training-Free Cached Exploration for Test-Time Search in Video Diffusion},
  author  = {Saini, Shreshth and Birkbeck, Neil and Wang, Yilin and Adsumilli, Balu and Bovik, Alan C.},
  journal = {arXiv preprint arXiv:2607.23159},
  year    = {2026}
}