Rootport/HareSkip-calibration

HareSkip Calibration Dataset

1. Summary

This dataset contains all images and measurements generated during a July–September 2026 recalibration and method-comparison experiment for HareSkip, a step-skipping inference-acceleration extension for DiT-based image diffusion, implemented on top of Forge neo (repository: Rootport-AI/forge-neo-Anima-HareSkip).

The experiment measured how much visual damage results from skipping specific denoising steps, across single-step sweeps, multi-step interaction patterns, out-of-sample validation patterns representative of HareSkip's actual runtime behavior, a fourth-stage comparison of skip-selection methods at matched skip counts, and a fifth-stage response-surface scan of the ResRefine prediction formula's own parameters. The full write-up, methodology, and analysis are in the experiment report:

docs/recalibration-2026-07/REPORT.md in the forge-neo-Anima-HareSkip repository.

This dataset consists of:

• 11,128 PNG images (1536×1536), covering reference (no-skip) images and skip-pattern candidates across 45 conditions (3 samplers × 5 prompts × 3 seeds), plus a step-count anchor set, a Shift-transfer spot check, a Stage 4 method-comparison set (TeaCache vs. sigmoid_band_v0.2 vs. monotone_saturate, at matched skip counts of 5, 10, and 15, including a supplemental TeaCache run forced to a true 15 realized skips), and a Stage 5 response-surface scan of the ResRefine prediction formula's strength × smoothing grid across 4 sampler/scheduler families (15 conditions each) plus a 2-prompt, 3-seed verification experiment (V1).
• Measurements: 14 Judge comparison runs (LPIPS-VGG/AlexNet, SSIM, PSNR, etc.) as CSV, a frozen pre-registered prediction table, and a full response table for the TeaCache threshold-calibration search.
• Patterns: the skip-pattern text files used to drive generation, including the frozen out-of-sample holdout patterns, the frozen Stage 4 comparison manifest, and the Stage 5 pattern-reuse tables (Stage 5 draws its skip patterns from the frozen Stage 4 monotone_saturate-r2 N=15 set rather than generating new ones — see §2 below).
• Generation provenance (generation/): per-image generation records (condition, source/destination path, SHA-256, sampler/seed/shift, acceptance status) for the Stage 5 campaigns.
• Analysis (analysis/): the Stage 5 findings write-ups, aggregate tables, and response-surface/scatter figures, plus the scripts used to produce them.
• Provenance: a fixed-commit statement, frozen-file hashes, and a full SHA-256 manifest.

The purpose of publishing this dataset is reproducibility: every image is traceable to an exact commit, an exact skip pattern, and an exact set of generation parameters, so that the report's findings can be independently re-derived or re-analyzed.

2. Dataset structure

images/
measurements/
patterns/
generation/
analysis/
provenance/

images/ (11,128 PNG files, 1536×1536)

Directory	Images	Content
Reference-images02/	45	No-skip reference images, one per condition (3 samplers × 5 prompts × 3 seeds), regenerated on the fixed commit
Calibration-phase1-02/	1,305	Stage 1: single-step skip sweep (steps 2–30, 29 positions × 45 conditions)
Calibration-stage2/	1,260	Stage 2: interaction patterns (adjacent pairs, distance scans, cross-band pairs, chains, triples — 28 patterns × 45 conditions)
Stage3-Calibration/	450	Stage 3: out-of-sample validation against HareSkip's actual runtime patterns (10 patterns × 45 conditions)
StepAnchor-images/	90	±1-step anchor: no extension, total step count varied to 29/31 only (45 conditions × 2)
ShiftSpotcheck-Reference/	6	Reference images for the Shift-transfer spot check (Shift=1)
ShiftSpotcheck-Calibration/	30	Shift=1 candidates at 5 characteristic curve points × 6 conditions
Stage4-Calibration/TeaCache/	135	Stage 4: TeaCache control, threshold-calibrated per sampler family, at nominal skip targets 5/10/15 (45 conditions × 3 targets)
Stage4-Calibration/sigmoid-v02/	405	Stage 4: sigmoid_band_v0.2 candidate, exact-target skip patterns at N=5/10/15 (45 conditions × 3 targets × 3 replicates)
Stage4-Calibration/monotone/	25	Stage 4: monotone_saturate candidate, N=15 only, new images only
Stage4-Calibration/TeaCache-15forced/	30	Stage 4 supplement: TeaCache re-run with the threshold raised until 15 skips were actually realized (2 sampler families × 15 conditions)
Stage5-Scan1/	6,660	Stage 5: ResRefine strength × smoothing response-surface scan, 4 sampler/scheduler families (ER SDE-Beta, ER SDE-Simple, Euler-Beta, Euler-Simple) × 15 conditions × 111 cells
Stage5-EulerSimple-Reference/	15	No-skip reference images for the Euler-Simple family (15 conditions × 1)
Stage5-V1/	666	Stage 5 verification experiment (V1): Euler-Beta × 2 new prompts × 3 seeds = 6 conditions × 111 cells
Stage5-V1-Reference/	6	No-skip reference images for the V1 conditions (6 conditions × 1)

Step 1 is excluded throughout: Manual Skip mode rejects step 1 (its residual is not yet held, so it is physically unskippable).

Why monotone/ has only 25 images, not 135: at N=15, the monotone_saturate and sigmoid_band_v0.2 skip patterns are identical for 110 of the 135 measured (condition, replicate) cells — the fall-off ("taper") term is the only structural difference between the two models, and it activates outside the effective skip window for all but 25 cells, so it changes nothing there. Where the patterns are bit-identical, the images are bit-identical, so those 110 measurements reuse the corresponding sigmoid-v02/ N=15 image rather than storing a duplicate. Only the 25 cells where the two patterns actually differ were generated separately and are stored here. The full mapping (which condition/replicate is shared vs. distinct) is recorded in patterns/stage4-comparison/stage4-comparison-manifest.csv.

Stage 5 design facts:

• Frozen-pattern reuse: Stage 5 does not generate new skip patterns. It reuses the frozen Stage 4 monotone_saturate-r2 N=15 skip set (see patterns/stage5-*/ below) and scans only the ResRefine prediction formula's own strength × smoothing parameters on top of that fixed skip set. This isolates the prediction-formula question from the skip-position question.
• 111 cells per condition: an 11×11 grid of strength ∈ {0.0, …, 1.0} × smoothing ∈ {0.0, …, 0.9, 0.99}, except that at strength=0 the prediction term is algebraically absent and smoothing has no effect — so only one image (reuse.png) was generated for that row and is replicated across all 11 smoothing columns during analysis. Realized image count per condition is therefore 10×11 + 1 = 111, not 121.
• Reuse = strength=0 baseline: the reuse cell is the no-prediction baseline against which every "improvement" figure in the Stage 5 analysis is measured.
• ema-099 naming: the top smoothing grid level is written ema-099 (not ema-10) because the extension's UI clamps the internal STATE value to 0.99 — 1.0 is not representable and would fail the infotext realized-value check by design. ema-00 … ema-09 correspond to smoothing 0.0 … 0.9; ema-099 corresponds to smoothing=0.99, the nearest representable value to 1.0.
• Seed correspondence: seed short-codes are consistent across all stages — 5193 = 3000995193, 6111 = 2455776111, 9676 = 2767019676. Because skip patterns are determined by scheduler × seed only (not by prompt), rows sharing a seed short-code across different prompts use the identical skip set — this is what makes the V1 verification experiment (new prompts, same seeds) a direct pattern-controlled comparison against the existing Euler-Beta conditions.

measurements/

Path	Content
judge-results/<run>/merged-results.csv + run-config.json (+ job-status.csv + raw/ for the 4 Stage 5 runs) (14 runs total)	Nz DoppelPix Judge comparison output: LPIPS-VGG (primary metric), LPIPS-AlexNet, SSIM (win 7 / win 11), PSNR, per candidate-vs-reference image pair
predictions/stage3-predictions.csv	Frozen pre-registered prediction table for the 450 Stage-3 out-of-sample points (predicted before generation; SHA-256 below)
predictions/fit_params.json	Frozen saturation-link fit parameters used to produce the predictions
tea-threshold-calibration.csv	Full response table for the TeaCache rel_l1-threshold calibration search (both the initial Stage 4 calibration and the supplemental forced-15 calibration): threshold tried, realized skip count, realized skip positions, and an adopted-run flag. The source images from this search are exploratory calibration probes, not part of the released image set — this table is the complete record of the search.

The 14 included runs are: Calibration-phase1-02_20260718-150151, Calibration-phase1-02-remainder_20260719-160107, StepAnchor_20260727-075911, Calibration-stage2_20260727-081812, ShiftSpotcheck_20260729-084448, Stage3_20260729-085016, Stage4-TeaCache_20260806-094737, Stage4-sigmoid_20260806-101101, Stage4-monotone_20260806-111759, Stage4-Tea15forced_20260901-082119, Stage5-Scan1_20260905-113330, Stage5-Families_20260907-235041, Stage5-EulerSimple_20260909-221014, Stage5-V1_20260914-190316. All completed with zero metrics errors.

An earlier measurement run (Calibration-phase1_20260716-234016) and its associated images/analysis are excluded from this dataset: they were generated before a Forge-neo version confound was discovered (see §4) and would mislead reproducibility checks if published alongside the corrected data.

patterns/

Skip-pattern definitions used to drive Manual Skip generation, plus the prompt set:

File	Content
layer1-single-skip-sweep-30steps.txt	Stage 1 patterns (29 lines)
stage2-interaction-patterns.txt	Stage 2 patterns (28 lines)
stage3-holdout-patterns.txt	Stage 3 out-of-sample holdout patterns (10 lines, frozen; SHA-256 below)
stage3-holdout-patterns-REPORT.md	Explanation of how the Stage 3 patterns were derived and verified
stage3-holdout-simulate.py	Offline simulator used to enumerate the Stage 3 patterns from HareSkip's probability model
stage4-comparison/stage4-comparison-manifest.csv	Stage 4 comparison manifest (540 rows: sigmoid_band_v0.2 N=5/10/15 × 45 conditions × 3 replicates = 405 rows, monotone_saturate N=15 × 45 conditions × 3 replicates = 135 rows). Each row records the exact-target-drawn skip set, its provenance, and the frozen pre-registered prediction (pred_main, frozen before generation). SHA-256 394EA341...A6B6 (full hash in provenance/manifest.sha256).
prompt.txt	The 5 prompts used across all conditions
stage5-scan1/stage5-scan1-patterns.csv, stage5-families/stage5-families-patterns.csv, stage5-eulersimple/stage5-eulersimple-patterns.csv, stage5-v1/stage5-v1-patterns.csv	Stage 5 pattern-reuse tables (condition, pattern_hash8, skip_set, source). Each row's source column points back to the frozen Stage 4 monotone_saturate-r2 N=15 pattern it reuses — Stage 5 generates no new skip patterns (see §2 above, "Stage 5 design facts")
stage5-v1/prompts-v1.json	The 2 new prompts (Prompt006-alchemist, Prompt007-butterfly) added for the Stage 5 verification experiment (V1)

generation/ (Stage 5 only)

Per-image generation provenance for the Stage 5 campaigns — not present for Stages 1–4, whose equivalent records are summarized in provenance/PROVENANCE.md instead.

Path	Content
stage5-scan1/manifest.json, stage5-scan1/status.csv, stage5-scan1/reference-eulersimple-status.csv, stage5-scan1/acceptance-chunk01.md…acceptance-chunk05.md	Allowed-condition list, grid spec, frozen commits, per-image source/destination path and SHA-256, and chunk-wise acceptance-test records for the Stage 5 response-surface scan
stage5-v1/manifest.json, stage5-v1/status.csv, stage5-v1/reference-status.csv, stage5-v1/acceptance-v1.md	The same, for the Stage 5 verification experiment (V1)

Run-time operational records (heartbeat/watchdog logs, the journal, dry-run previews, per-chunk stdout/stderr, calibration smoke-test probe images) are excluded — they are execution logs, not measurement data, and remain archived locally as an audit trail.

analysis/ (Stage 5 only)

Findings write-ups, aggregate tables, and figures for the Stage 5 campaigns, plus the scripts that produced them — included because a reader of the source report is likely to want these as a direct reference.

Path	Content
stage5-scan1/, stage5-families/, stage5-eulersimple/, stage5-v1/	Each contains FINDINGS*.md (write-up), tables/ (CSV/JSON aggregates), surfaces/ (response-surface heatmap PNGs), scatter/ (scatter-plot PNGs), and the .py scripts used to regenerate them

provenance/

File	Content
PROVENANCE.md	Fixed-commit statement, frozen-file hashes, excluded-data rationale, included-run list
manifest.sha256	SHA-256 checksums for every file in the dataset (sha256sum-compatible format)

3. Self-describing images

Every PNG carries a tEXt chunk keyed parameters, in Forge's standard infotext format, embedding the full generation condition — prompt, negative prompt, seed, sampler, schedule type, CFG scale, Shift, model and hash, HareSkip mode, ResRefine formula, and the realized skipped-step set. For Manual Skip mode this is Manual skipped_steps; for TeaCache mode (used in the Stage 4 comparison) it is Tea skipped_steps, recording the skip positions TeaCache actually took at generation time, not a nominal target. Example (from images/Calibration-phase1-02/ERSDE-Beta_Prompt001-schoolgirl-5193/00000-3000995193.png):

masterpiece, best quality, score_7, safe, a girl with demon horns is sitting indian style on ground, holding a cat. ...
Negative prompt: worst quality, low quality, score_1, score_2, score_3, artist name
Steps: 30, Sampler: ER SDE, Schedule type: Beta, CFG scale: 4.0, Shift: 3.0, Seed: 3000995193, Size: 1536x1536, Model: anima_baseV10, Model hash: bd43b7cffe, Module 1: qwen_image_vae, Module 2: qwen_3_06b_base, RNG: CPU, HareSkip enabled: True, HareSkip mode: Manual Skip, ResRefine formula: Reuse (residual only), Beta schedule alpha: 0.6, Beta schedule beta: 0.6, Manual skipped_steps: 2, Version: neo-2.27

Manual skipped_steps / Tea skipped_steps is written by the patcher from the steps it actually skipped during generation, not from the requested input — it is a realized value, not a nominal one. This makes every image self-describing on its own: the generation condition can be recovered from the file alone, with no dependency on directory naming or an external index. Dataset-wide acceptance testing cross-checked this realized value against file name and pattern canon for all images, including the Stage 4 TeaCache set (Tea skipped_steps verified against generation logs for all 135 + 30 TeaCache images).

4. Reproducibility & integrity

All images through Stage 3 were generated on a single, fixed Forge neo commit: b61642140acb7c2f1c65c5d0f2ab961b7366c02e (reported in PNG metadata as Version: neo-2.27). This fix was adopted after an earlier measurement round was found to be confounded by an undocumented Forge-neo version drift between reference and candidate images (see the source report, §2.2, for the full incident writeup); that earlier round is excluded from this dataset (§2 above).

The Stage 4 comparison set was generated on the same fixed Forge neo commit, with the extension advanced to commit 149325b or later, which added realized-skip-count reporting (Tea skipped_steps) for TeaCache mode — the mechanism this comparison depends on to verify what each generated image actually skipped. All Stage 4 images still report Version: neo-2.27 in infotext.

Stage 4 followed the same pre-registration discipline as Stage 3: the comparison patterns and the predicted-damage values were frozen (patterns/stage4-comparison/stage4-comparison-manifest.csv) before the corresponding images were generated or measured.

The Stage 5 response-surface scan and the V1 verification experiment were generated on the same fixed Forge neo commit (b6164214, Version: neo-2.27) with the extension at commit 7424841. Stage 5 does not introduce new skip patterns to pre-register — it reuses the frozen Stage 4 monotone_saturate-r2 N=15 skip set and scans only the ResRefine strength × smoothing prediction-formula parameters against it (see provenance/PROVENANCE.md §7 for the full design record, including the grid layout, the ema-099 naming note, and the seed-correspondence table).

Integrity verification

Every file in the dataset is listed in provenance/manifest.sha256. To verify the full download:

sha256sum -c provenance/manifest.sha256

Frozen hashes and pre-registration

Stage 3 (out-of-sample validation) and Stage 4 (method comparison) each followed a pre-registration discipline: predictions were frozen before the corresponding images were generated or measured.

File	SHA-256 (first/last 8 hex)
patterns/stage3-holdout-patterns.txt	659A7445...7CD3
measurements/predictions/stage3-predictions.csv	F8495DD1...FA131D
patterns/stage4-comparison/stage4-comparison-manifest.csv	394EA341...A6B6

(Full 64-character hashes are recorded in provenance/manifest.sha256.) The Stage 3 holdout patterns were derived by offline simulation of HareSkip's probability model (sigmoid_band_v0.1) and independently verified against the extension's own code before generation. The prediction table's parameters were fit only on Stage 2 data; re-fitting on Stage 3 data was explicitly disallowed. The Stage 4 manifest's pred_main column carries the corresponding frozen pre-registered prediction for every sigmoid_band_v0.2/monotone_saturate row, likewise frozen before generation. This lets a third party check that the predictions in this dataset were not adjusted after seeing the outcomes they are being judged against.

5. Model and licensing

The images in this dataset were generated with Anima (Anima-Base v1.0) by CircleStone Labs & Comfy Org (https://huggingface.co/circlestone-labs/Anima), a derivative of NVIDIA Cosmos-Predict2-2B-Text2Image ("Built on NVIDIA Cosmos"). The exact model file used has SHA-256 bd43b7cffe1ed1153d9c41e7beb2f18cb1273eafbaa3af3edd6a173dc90a006e, which matches the Model hash: bd43b7cffe recorded in every image's infotext (verified against the dataset).

License note: The Anima model weights themselves are distributed under the CircleStone Labs Non-Commercial License v1.2 (non-commercial use only for the weights). However, that license's §2.e addresses Outputs (generated images) explicitly:

"We claim no ownership rights in and to the Outputs. ... You may use Outputs for any purpose (including for commercial purposes)."

Accordingly, this dataset — the collection of Outputs plus the original measurements produced by this experiment — is released under CC-BY 4.0. This license covers the images and measurement data in this repository. It does not cover the Anima model itself; anyone wishing to use the Anima model weights to generate new images must comply with CircleStone Labs' own license terms.

6. What this data shows

• Damage from skipping a step is concentrated in the early denoising steps (Spearman ρ = −0.784 between step position and damage), with the worst position typically at skip 2–4 rather than the very first skippable step.
• A simple predictive formula — the sum of single-step damages passed through a two-parameter saturation link — passed pre-registered out-of-sample validation on 450 held-out points (Spearman ρ = 0.930), with no need for a streak (consecutive-skip) penalty term.
• Predicted damage carries a consistent safe-side bias (actual damage tends to be ≤ predicted), which is favorable for quality-guarantee use but means absolute-value predictions should not be taken as tight bounds.
• The hypothesis that damage is governed purely by trajectory coordinate z (independent of the Shift parameter) did not hold cleanly: changing Shift from 3 to 1 changes the shape (steepness) of the damage curve, not just its position.
• Stage 4 (method comparison, matched skip counts): at N=10, sigmoid_band_v0.2 clearly outperforms a threshold-calibrated TeaCache control (median LPIPS-VGG 0.162 vs. 0.196 across 45 conditions; anchor-inlier rate 44.4% vs. 28.9%). At N=15, TeaCache's apparent edge in the initial comparison (0.282 vs. 0.348) turned out to be confounded by realized skip count — TeaCache only reached 13–14 of its nominal 15 skips in 30 of 45 conditions. A supplemental run that raised TeaCache's threshold until it genuinely realized 15 skips on every condition reversed the ranking: at true N=15, sigmoid_band_v0.2 is comparable to or better than TeaCache (median 0.348 vs. 0.365, and a clear edge in the tail — q95 0.504 vs. 0.554, 3/45 vs. 6/45 conditions crossing the "different-image" threshold of 0.5). Whether the fall-off ("taper") term is present makes no measurable difference: for 110 of 135 N=15 cells the skip patterns of sigmoid_band_v0.2 and its taper-free variant monotone_saturate are identical, and the 25 cells that do differ show no consistent direction (median difference +0.004, roughly 3% of the anchor scale).
• Stage 5 (ResRefine strength × smoothing response surface, plus the V1 verification experiment): see analysis/stage5-scan1/FINDINGS.md, analysis/stage5-families/FINDINGS-families.md, analysis/stage5-eulersimple/FINDINGS-eulersimple.md, and analysis/stage5-v1/FINDINGS-v1.md for the full per-family and verification results — not summarized here to avoid pre-empting the source report's own discussion.

See docs/recalibration-2026-07/REPORT.md for full methodology, statistics, and discussion.

7. Citation

@misc{rootport2026hareskipcalibration,
  author       = {Rootport},
  title        = {HareSkip Calibration Dataset},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/Rootport/HareSkip-calibration}},
}

---

日本語要約

本データセットは、DiTベース画像生成の高速化拡張「HareSkip」（Forge neo拡張、リポジトリ: Rootport-AI/forge-neo-Anima-HareSkip）の較正実験・方式比較実験・応答曲面走査実験で生成した全画像・測定値です。固定コミット（b6164214、neo-2.27）で生成した11,128枚のPNG（1536×1536）と、LPIPS/SSIM等の測定CSV、パターン正典、生成来歴、分析結果、出自記録一式を収録しています。各PNGのinfotextに実現スキップ位置を含む生成条件が完全に埋め込まれており、画像単体で再現性を検証できます。第4段階では、同一スキップ本数（5/10/15）でTeaCache（既存の高速化手法）とHareSkipの確率モデル（sigmoid_band_v0.2 / monotone_saturate）を比較しており、N=10ではHareSkipが明確に優位、N=15は本数を揃えて比較するとHareSkipが同等以上（特に裾の損傷でHareSkipが優位）という結果が得られています。

第5段階（images/Stage5-Scan1/・images/Stage5-V1/ 等、計7,347枚追加）は、ResRefine予測式（Linear extrapolation + EMA平滑化）自身の strength × smoothing パラメータに対する応答曲面を走査した実験です。スキップ位置そのものは第4段階で凍結した monotone_saturate-r2 N=15 パターンを再利用しており、新規パターン生成は行っていません（予測式の効き方だけを単独で評価するための設計）。ER SDE-Beta・ER SDE-Simple・Euler-Beta・Euler-Simple の4系×15条件、および新規2プロンプトによる検証実験V1（Euler-Beta×6条件）を収録し、各条件111セル（11×11格子のうち strength=0 行はreuse基準1点の複製）で構成されます。格子の smoothing 最上位値は拡張UIのクランプ制約により ema-099（=0.99）と命名されています。シードは全プロンプト共通で 5193=3000995193 / 6111=2455776111 / 9676=2767019676 に対応します。第5段階の詳細な結果は analysis/stage5-*/FINDINGS*.md に譲り、本カードでは結論を先取りしません。

生成モデルはAnima（CircleStone Labs & Comfy Org、NVIDIA Cosmos派生）で、モデル本体は非商用ライセンスですが、同ライセンス§2.eにより生成画像（Outputs）の商用利用を含む自由な利用が明記されているため、本データセット（Outputs＋独自測定値）はCC-BY 4.0で公開します。詳細な分析結果（損傷の序盤集中、予測式のout-of-sample検証、Shift依存性、第4段階の方式比較、第5段階の応答曲面など）は元リポジトリの実験レポートおよび analysis/ 配下の各FINDINGSを参照してください。