CoolFace
Datasetpublic

Perflow-Shuai/cfg-sensitive-video-prompts

CFG-Sensitive Video Prompts A provenance-first, text-only prompt suite for comparing classifier-free guidance behavior in video generation. It contains 22 English prompts: existing_visualized: 6 prompts used by the local Wan2.2-TI2V-5B CFG-only LoRA scale-sweep report. cfg_sensitive_candidate: 16 additional prompts selected because color/attribute binding, numeracy, readable text, spatial relations, temporal change, fluid motion, weather, camera motion, multi-subject… See the full description on the dataset page: https://huggingface.co/datasets/Perflow-Shuai/cfg-sensitive-video-prompts.

sourceHugging Facecc-by-nc-sa-4.0updated 17d agoView on Hugging Face
0likes102downloads
Dataset Card

CFG-Sensitive Video Prompts

A provenance-first, text-only prompt suite for comparing classifier-free guidance behavior in video generation. It contains 22 English prompts:

  • existing_visualized: 6 prompts used by the local Wan2.2-TI2V-5B CFG-only LoRA scale-sweep report.
  • cfg_sensitive_candidate: 16 additional prompts selected because color/attribute binding, numeracy, readable text, spatial relations, temporal change, fluid motion, weather, camera motion, multi-subject consistency, style, transparency, or symmetry may make guidance differences easier to inspect.

The 16 candidates are hypotheses, not empirical winners. They have not yet been inferred in this package, so the dataset does not claim that CFG=1 and CFG=5 differ visibly on every row.

Recommended comparison protocol

Hold prompt, negative prompt, seed/noise, resolution, frame count, sampler, shift, and step count fixed. Compare the same base model at CFG=1 versus its documented default CFG (Wan2.2-TI2V-5B uses CFG=5), and keep any LoRA multiplier labeled separately: a LoRA weight scale is not itself a calibrated CFG value.

Provenance

The six existing prompts were selected from local val_prompt1.txt non-empty lines 1, 3, 6, 21, 23, and 32 (SHA256 611a1292a450e2933a46b76549b7d1a78211951b5c8db2398dd94af6523fedd2). Four of them match the pinned public vidprom_filtered_extended.txt byte-for-byte:

  • golfers -> upstream line 1
  • coastal cottage -> upstream line 3
  • spaceship/meteorite belt -> upstream line 9
  • milk splatter -> upstream line 972

The eagle and rose-bloom prompts exist in the local validation file but have no exact match in the pinned 248,221-row source; their earlier upstream authorship is therefore marked unresolved rather than guessed.

All 16 new candidates are exact rows from `Efficient-Large-Model/LongLive-1.3B` at `cda9138d9387`. That file's SHA256 is 7896742f468bc8aef9e4547424d1ce0a951acdb2a82233790155401a99bf5aa5. An identical public dataset mirror is available at `Ingra14m/vidprom_filtered_extended` at `48f3fac42d2e`. Candidates were checked for exact-prompt exclusion from both the local 32-prompt validation file and the 64-prompt CFG teacher trajectory selection.

See provenance.json for revisions, hashes, source lines, selection policy, and package checksums.

Fields

  • id, group, prompt, suggested_seed
  • cfg_sensitivity_axis, selection_reason, empirical_status
  • per-prompt SHA256 and local/upstream source coordinates
  • existing report build/case IDs where applicable

Scope and license

This repository contains prompts and metadata only, not generated videos, latents, or model weights. It uses the conservative cc-by-nc-sa-4.0 license carried by the pinned LongLive-1.3B repository that hosts the source prompt file.