# Singularity R91 for ComfyUISingularity is a public-alpha ComfyUI custom node for Wan I2V / V2V cascade continuation.It is built for users who want longer Wan videos without manually extracting a last frame, loading it again, making another short clip, and stitching everything later.## What It DoesSingularity turns this manual loop:```textgenerate a short clipfind a good tail frameload that frame as the next sourcegenerate the next cliprepeatstitch everything later```into one controlled node workflow:```textgeneratepausechoose a tail framecontinue the same runsave one stitched final video```## Main Features- One visible ComfyUI node: <code>Singularity</code>.- Wan I2V / V2V cascade continuation.- Up to five cascade segments in this public alpha.- Optional pause after cascade 1, 2, 3, and 4.- Detached Source / Tail 1 / Tail 2 / Tail 3 / Result panel under the node.- Manual tail-frame selection before continuation.- Same-run resume instead of unrelated separate renders.- Final stitched video output.- Markdown report output with cascade, motion, runtime, delta, and Strategy diagnostics.- Clean public defaults: no bundled image, empty positive prompt, built-in Wan-style Chinese negative prompt.- Native ComfyUI image upload button is preserved.## New In R91R91 is a stabilization update for modern ComfyUI Desktop.- Visible title now shows <code>Singularity R91</code>.- The public node surface is clean for fresh installs.- The <code>math_control_mode</code> dropdown now shows only the clean public modes.- Older saved workflow values are still normalized by the backend.- Pause/continue UI remains detached below the node.- The native upload button is preserved.- Strategy Control Surface diagnostics are included in reports.- Experimental math remains neutral by default.## About The MathSingularity studies generation as an event:```textOutcome(t-1) + ObservedBehavior(t-1)= Strategy(t)= ObservedBehavior(t+1) + Outcome(t+1)```For normal users, this means the node helps you inspect whether the source image, prompt meaning, model behavior, sampler route, motion, and final video are staying connected.This is not a magic quality button. It is a control and diagnostics layer.Use <code>OBSERVE_ONLY</code> for a safe baseline.Use <code>STRATEGY_PRESSURE_WINDOW</code> only when you intentionally want to test bounded experimental math. With <code>high_delta_strength = 1.0</code> and <code>low_delta_strength = 1.0</code>, it stays neutral.Use <code>LATENT_DELTA_SCALE</code> and <code>DEEP_STEP_DELTA_CONTROL</code> only for research.## Recommended First Test```textcascade_count = 2pause_after_cascade_1 = trueframes_per_cascade = 49width = 704height = 1280fps = 16seed = 123math_control_mode = OBSERVE_ONLYhigh_delta_strength = 1.0low_delta_strength = 1.0save_video = truesave_report = true```Pick a source image, write a clear prompt, generate, choose a tail frame when the panel appears, continue, then inspect the final stitched video and report.## Public Alpha Notes- Models are not included.- This is Wan-first.- Heavy resolutions require more VRAM and time.- <code>CompletionGate = PASS</code> means the route completed and a final video exists. It does not guarantee perfect visual quality.- Always inspect the final video, not only the report.

Wan2.2 I2V V2V Video Extend Quant  14B

# Event Horizon / Event Equality Core - Public Changelog## 0.1.1-r59 Public Alpha CleanCompared with the previous public build:```textprevious public: ComfyUI-Event-Equality-Core_v0.1.1-r56_<a target="_blank" rel="ugc" href="http://PublicAlphaCleanReadme.zip">PublicAlphaCleanReadme.zip</a>current public: ComfyUI-Event-Equality-Core_v0.1.1-r59_<a target="_blank" rel="ugc" href="http://PublicAlphaClean.zip">PublicAlphaClean.zip</a>```Status: public alpha / experimental.The main r59 goal is to keep one clean user-facing node, <code>Event Horizon</code>, while adding stronger diagnostics, safer input handling, and a better placement for the math layer so it does not break the native generation physics.## Short SummaryCompared with r56:- Added r57 motion-math metrics.- Added r58 input-integrity hardening.- Corrected r59 <code>LATENT_DELTA_SCALE</code>: it now uses a safe semantic overlay / native-sampler-preserving path instead of replacing the sampler step loop by default.- Added the separate experimental <code>DEEP_STEP_DELTA_CONTROL</code> mode for deeper math research.- Cleaned the public package: one external node, no debug nodes, no old workflow aliases, no examples, no checklists, no test artifacts.- Rewrote the README as public user documentation instead of an internal handoff note.## Important Workflow Changer56 public exported:```textEventHorizonEventHorizonR56PublicAlphaEventDebugPingEventSaveReportToFile```r59 public clean exports only:```textEventHorizon```In the ComfyUI interface this appears as:```textEvent Horizon```If an old workflow was saved with a versioned class id such as <code>EventHorizonR56PublicAlpha</code>, remove that old node and add a fresh <code>Event Horizon</code> node. This is intentional: the public package should stay clean and not accumulate internal development aliases.## What Was Added After r56### r57 - Motion Math Metrics Diffr57 adds more detailed motion math after decode.New metrics:- <code>frame_delta_norm_p25</code>- <code>frame_delta_norm_p50</code>- <code>frame_delta_norm_p75</code>- <code>frame_delta_norm_p90</code>- <code>frame_delta_norm_p95</code>- <code>frame_delta_norm_iqr</code>- <code>frame_delta_norm_cv_ratio</code>- <code>frame_delta_p95_to_p50_ratio</code>- <code>frame_delta_jerk_abs_mean</code>- <code>frame_delta_jerk_ratio</code>- <code>frame_motion_stability_score</code>- <code>frame_motion_profile</code>What this gives you:- faster fixed-seed comparison between runs;- clearer visibility into smoother, sharper, unstable, or spiky motion;- less accidental mixing between single-run and cascade-run baselines because the runtime signature is stricter;- more useful reports for behavior analysis, not just pass/fail status.Important: these metrics are observer-only. They do not mutate tensors and do not improve video quality by themselves.### r58 - Input Integrity Hardeningr58 adds central input normalization before generation.Values now normalized and protected:- <code>cascade_count</code>- <code>frames_per_cascade</code>- <code>width</code>- <code>height</code>- <code>fps</code>- <code>seed</code>- sampler windows: <code>primary_start_step</code>, <code>primary_end_step</code>, <code>secondary_start_step</code>, <code>secondary_end_step</code>- <code>math_control_mode</code>- <code>high_delta_strength</code>- <code>low_delta_strength</code>- <code>decode_tile_size</code>- <code>decode_overlap</code>- <code>decode_temporal_size</code>- <code>decode_temporal_overlap</code>- <code>image_upscale_method</code>- <code>image_crop</code>- <code>cleanup_timing</code>- <code>video_format</code>- <code>save_prefix</code>- <code>sampler_trace_mode</code>- <code>sampler_trace_max_steps</code>What this gives you:- stale or partially corrupted widget values are less likely to break a run;- width, height, and decode values are constrained into safe ranges;- overlap cannot exceed safe tile limits;- empty or unsafe <code>save_prefix</code> values are converted into safe file names;- sampler and scheduler values are checked against available <code>KSamplerAdvanced</code> options when ComfyUI exposes them;- the report records <code>EventInputNormalization</code> and <code>EventInputNormalizationAdjustments</code>.UI hardening:- prompt boxes are wider and taller;- positive and negative prompt widgets are protected from overlap during resize/configuration cycles;- widget order is not sorted, because ComfyUI stores widget values positionally.### r59 - Strategy Math Native Loop / Overlay Correctionr59 adds a deeper split between safe public math behavior and experimental deep-step math behavior.Main correction:```textLATENT_DELTA_SCALE no longer replaces the native sampler loop by default.```Why:A rough step-by-step sampler loop replacement can turn the output into noise even when the structural gate still shows PASS. That means the math started overriding generation physics instead of helping the model hold the intended semantic strategy.What changed:- <code>LATENT_DELTA_SCALE</code> now uses the safe <code>semantic_overlay_native_sampler</code> path.- The model's native sampler remains the main generator.- Delta math is applied as a controlled semantic overlay / post-window relation layer.- The report now includes <code>EventMathSamplerPathPolicy</code>, so the selected math path is visible.- Fixed false huge <code>relative_delta</code> when the high-stage baseline norm is zero.- High-to-low coupling is no longer incorrectly suppressed by a zero-latent start.- Added the separate research mode <code>DEEP_STEP_DELTA_CONTROL</code>.## Math Control Modes### <code>OBSERVE_ONLY</code>Safe baseline mode.What it does:- runs generation through the normal path;- records math/report records;- should not intentionally mutate generation tensors;- is suitable for baseline comparison.When to use it:- first test after installation;- checking that model, workflow, image, VAE, CLIP, and save path all work;- baseline before delta or motion comparisons.### <code>LATENT_DELTA_SCALE</code>The public default working mode.What it does:- measures latent movement as <code>ObservedBehavior</code>;- applies <code>high_delta_strength</code> and <code>low_delta_strength</code>;- preserves the native sampler path;- writes math policy records;- helps study how changing high/low latent delta strength affects the strategy carrier between stages.Practical meaning:- <code>1.0</code> = neutral;- below <code>1.0</code> = reduce this stage's delta contribution;- above <code>1.0</code> = amplify this stage's delta contribution.Formula meaning:```textdelta strength changes ObservedBehaviorObservedBehavior changes OutcomeOutcome becomes StrategyCarrier for the next stage```So this is not just a quality slider. It is a controlled change to the latent route that gets passed forward.### <code>DEEP_STEP_DELTA_CONTROL</code>Experimental research mode.What it does:- enables the deeper native step-loop control path;- can intervene more strongly in step dynamics;- is meant for studying where the math belongs inside the intersection of model output, sampler state, latent state, and prompt strategy.Risk:- it can produce noise;- it requires sweeps over <code>high_delta_strength</code> / <code>low_delta_strength</code>;- it is not recommended as a normal public-user mode.When to use it:- only for controlled experiments;- only with a fixed seed;- always save the report;- always inspect the actual video, not just PASS/BLOCKED status.## Drift: What It Means In This NodeDrift is measurable separation between the intended route and the actual generation trajectory.Drift types:- Source drift: the video stops respecting the source image.- Prompt drift: the result moves away from the positive prompt or toward the negative prompt.- Identity drift: the main object or character changes more than intended.- Motion drift: movement becomes jerky, spiky, reversed, or too chaotic.- Cascade drift: the next cascade segment does not continue the previous segment cleanly.- Route drift: runtime route, stage order, or boundary records do not match the expected logic.- Math drift: the selected math starts overriding generation instead of refining strategy.Important:The report does not remove drift by itself. It shows where drift appears so fixed-seed runs can be compared with evidence.## Full Node Functionality### External Node```textEvent Horizon```This is the only visible node in the public clean package.It combines:- source image loading;- prompt encoding;- high sampler stage;- low sampler stage;- optional cascade continuation;- VAE decode;- video saving;- report saving;- runtime diagnostics;- Event Equality records.### Required Inputs<code>primary_model</code>Main model branch. Used for the high/main sampling stage.<code>clip</code>Text encoder. Converts positive/negative prompts into conditioning.<code>vae</code>Used to decode latent frames into images/video.<code>source_image_file</code>Image picker for the ComfyUI input folder. If the external <code>image</code> socket is not connected, the node uses this file picker.<code>positive_prompt</code>What the model should generate. In the project logic, this is not the whole Strategy; it is a carrier / part of the StrategyCandidate.<code>negative_prompt</code>What the model should avoid. This is a negative constraint carrier.<code>temporal_texture_lock</code>Flag for temporal/texture continuity intent. In public alpha it is part of route intent and report context.### Optional Inputs<code>secondary_model</code>Optional low/secondary model branch. If it is not connected, the node can fall back to a single branch and record a conflict/warning.<code>image</code>Direct IMAGE input when the workflow passes an image instead of using the file picker.<code>mask</code>Optional mask input. Currently used as report/context input for future or partial masked-route scenarios.## Generation Controls<code>cascade_count</code>Number of generation segments.- <code>1</code> = single video;- <code>2-5</code> = cascade continuation.<code>frames_per_cascade</code>Number of frames per cascade segment.<code>width</code>, <code>height</code>Video dimensions. Normalized into a safe step-aligned range.<code>fps</code>Output video frame rate.<code>seed</code>Seed for reproducible comparisons. For metric research, keep the same seed across runs.## Sampler Controls<code>sampler_name</code>Sampler name. r58+ tries to validate it against available ComfyUI sampler options.<code>scheduler</code>Scheduler for the sampling path. Also validated/fallbacked when ComfyUI exposes allowed values.<code>global_steps</code>Total sampling steps.<code>primary_cfg</code>CFG for the primary/high branch.<code>secondary_cfg</code>CFG for the secondary/low branch.<code>primary_start_step</code>, <code>primary_end_step</code>Step window for the high sampler stage.<code>secondary_start_step</code>, <code>secondary_end_step</code>Step window for the low sampler stage.## Delta Controls<code>math_control_mode</code>Selects the math mode:- <code>OBSERVE_ONLY</code>- <code>LATENT_DELTA_SCALE</code>- <code>DEEP_STEP_DELTA_CONTROL</code><code>high_delta_strength</code>Scales high-stage latent delta.<code>low_delta_strength</code>Scales low-stage latent delta.Practical recommendation:Start with:```texthigh_delta_strength = 1.0low_delta_strength = 1.0```For comparison after baseline:```texthigh_delta_strength = 0.988low_delta_strength = 1.0```## Decode Controls<code>decode_tile_size</code>Tile size for VAE decode.<code>decode_overlap</code>Overlap between tiles. r58+ prevents it from becoming unsafe relative to tile size.<code>decode_temporal_size</code>Temporal tile size for video decode.<code>decode_temporal_overlap</code>Temporal overlap. r58+ constrains it relative to temporal size.## Image Controls<code>image_upscale_method</code>Source image resize method:- <code>nearest-exact</code>- <code>nearest</code>- <code>bilinear</code>- <code>area</code>- <code>bicubic</code>- <code>lanczos</code><code>image_crop</code>Crop mode:- <code>disabled</code>- <code>center</code>## Cleanup / Barrier Controls<code>cleanup_timing</code>Controls cleanup/barrier timing:- <code>NONE</code>- <code>BEFORE_GENERATION</code>- <code>BETWEEN_SAMPLERS</code>- <code>AFTER_GENERATION</code>- <code>BEFORE_AND_AFTER</code>- <code>ALL</code>What it does:- records where cleanup boundaries were called;- attempts to release disposable memory;- should not destroy tensors needed by the low sampler or final output;- writes Smart Branch Barrier records into the report.Formula meaning:The barrier must not break the StrategyCarrier between the high and low stages.## Save Controls<code>save_video</code>Saves the video to ComfyUI output.<code>video_format</code>Save format:- <code>video/h264-mp4</code>- <code>video/h265-mp4</code>- <code>image/webp</code>- <code>image/gif</code><code>save_report</code>Saves the markdown report.<code>save_prefix</code>Output file prefix. r58+ sanitizes unsafe characters and empty values.## Sampler Trace Controls<code>sampler_trace_mode</code>Step trace mode:- <code>OFF</code>- <code>SHADOW_STEP_TRACE</code><code>SHADOW_STEP_TRACE</code> writes shadow records for step-level behavior analysis, but it should not replace the main sampler output.<code>sampler_trace_max_steps</code>Limits the number of traced steps so reports do not grow without control.## Outputs<code>status</code>Short execution status.<code>saved_video_path</code>Path to the saved video when <code>save_video = true</code> and saving succeeds.<code>saved_report_path</code>Path to the saved report when <code>save_report = true</code> and saving succeeds.<code>report</code>Report text returned directly from the node output.## What The Report Can IncludeThe report can include:- runtime version/name;- generation settings snapshot;- source image load status;- text encode records;- high/low sampler records;- latent boundary math;- frame motion math;- cascade boundary math;- RouteMemory timeline;- S-Wire state;- SState snapshots;- EventConflict records;- CompletionGate;- RuntimeMonitor summary;- optional sidecar paths;- sampler trace summary;- input normalization adjustments.## CompletionGate<code>EventCoreBodyCompletionGate = PASS</code> means structural integrity of the runtime route:- required stages were recorded;- stage order was valid;- report/finalize logic reached completion;- the node returned the expected outputs.It does not mean:- the video is visually good;- the prompt was understood perfectly;- drift is absent;- deep math mode is safe for all cases.Correct verification:```textresult_status = VIDEOEventCoreBodyCompletionGate = PASSvideo visually inspectedreport metrics compared against baseline```## Recommended Test Matrix### First Run```textmath_control_mode = OBSERVE_ONLYhigh_delta_strength = 1.0low_delta_strength = 1.0cascade_count = 1save_video = truesave_report = true```Goal: check that model, VAE, CLIP, image input, and save path work.### Public Default Run```textmath_control_mode = LATENT_DELTA_SCALEhigh_delta_strength = 1.0low_delta_strength = 1.0```Goal: check the safe overlay path.### Delta Comparison```textmath_control_mode = LATENT_DELTA_SCALEhigh_delta_strength = 0.988low_delta_strength = 1.0same seed as baseline```Goal: compare motion metrics and visual output.### Cascade Test```textcascade_count = 5frames_per_cascade = 49same seed```Goal: check cascade continuation and boundary records.### Experimental Deep Math Test```textmath_control_mode = DEEP_STEP_DELTA_CONTROLhigh_delta_strength = start below 1.0low_delta_strength = 1.0same seed```Goal: study deep intervention. If noise appears, return to <code>LATENT_DELTA_SCALE</code> or lower strength values.## Known Limits- Public alpha, not production-stable.- The main testbed is currently Wan I2V.- Step-level formula coverage is still not fully closed.- Shadow trace is a diagnostic observer, not a full native sampler hook.- Deep step control is research mode.- CompletionGate PASS does not replace visual inspection.- Reproducibility depends on ComfyUI, torch, GPU, drivers, model build, and sampler implementation.## Package Cleanup In r59 PublicRemoved from the public package:- old examples;- internal release checklists;- debug node exports;- versioned public aliases;- pycache;- generated output artifacts;- handoff/transfer files;- private workflows;- old reports and videos.Kept:- working node code;- imported internal modules;- web UI helper;- README;- CHANGELOG;- license;- requirements;- pyproject;- version file;- formula integrity note.

Files with SHA256: 1791648ef58626f9764e1cdf211bb2051f972ebf8d0649ffa5e57272826fb8a0