Singularity is a public-alpha ComfyUI custom node for Wan image-to-video cascade continuation.
It is built for users who want longer Wan I2V / V2V videos without manually extracting a last frame, loading it again, generating another short clip, and stitching everything later.
The main feature in this release is the Tail 5 Continuation Gate.
When a cascade segment finishes, Singularity can pause the workflow and show a detached panel:
Source | Tail 1 | Tail 2 | Tail 3 | Tail 4 | Tail 5 | Result
You choose the tail frame that should become the source for the next segment, then press Resume Cascade / Continue.
This gives you direct control over the continuation point. Instead of blindly continuing from a weak last frame, you can pick the frame with the best pose, face, motion direction, or visual stability.
- Adds one visible ComfyUI node:
Singularity. - Generates Wan I2V cascade segments.
- Supports up to five cascade segments in the current public alpha.
- Can pause after cascade 1, 2, 3, or 4.
- Shows five tail-frame candidates at each pause.
- Lets you manually choose the continuation frame.
- Continues the same running workflow after your choice.
- Saves one stitched final video.
- Saves a Markdown runtime report with cascade, motion, seam, and completion evidence.
- Keeps the public default conservative: observe and report first, no active generation math by default.
The pause-and-select layer. It shows Source, five tail candidates, and Result. The selected tail frame becomes the next segment source.
The continuation layer. The node keeps the cascade route inside one run and saves one final stitched video instead of leaving you with separate clips.
The default public mode. Observe Only records evidence and reports what happened without intentionally mutating tensors.
Prompt analysis stays report-side. Singularity can inspect prompt structure and relation pressure, but the public default does not inject formula prose into the model-facing prompt.
The report layer. CompletionGate = PASS means the requested route completed and a final video exists. It is structural proof, not a guarantee that the video is visually perfect.
The diagnostics layer. Reports can show cascade seam pressure, visible motion spikes, tail/source continuity pressure, and whether the next segment looks like it re-entered the same event or drifted.
Version: 0.1.1-r178
Visible title: Singularity R178
Release name: Tail 5 Continuation Gate
R178 is a continuation-gate stabilization release.
Highlights:
- Friendly public mode names:
Observe OnlyLatent Delta ScaleTail Source ReconstructionSource Noise Field Shaping- and other research modes.
- Backend still normalizes friendly names to stable canonical values.
- Public controls are grouped visually:
SOURCEPROMPTTIMELINESAMPLERMATHDECODEOUTPUTLAB
- Internal transport widgets are hidden from the public first surface.
- Native ComfyUI image upload is preserved.
- The Tail 5 panel stays detached under the node and disappears when the run finishes.
- Reports remain enabled by default.
- Public baseline uses
Observe Onlywith neutral delta values.
Fresh-node defaults are intended to be safe starter settings:
source_image_file = none
positive_prompt = empty
negative_prompt = built-in Wan-style base negative prompt
cascade_count = 2
pause_after_cascade_1 = true
frames_per_cascade = 49
width = 704
height = 1280
fps = 16
seed = 123
math_control_mode = Observe Only
high_delta_strength = 1.0
low_delta_strength = 1.0
strategy_field_mode = OFF
image_crop = wan_native
save_video = true
save_report = true
sampler_trace_mode = OFF
prompt_transcode_mode = Report Only
formula recommendation = off
If 704 x 1280 is too heavy for your GPU, lower the resolution before rendering. A fast comparison size such as 416 x 608 is useful for debugging.
- Add a fresh
Singularitynode. - Connect your model, CLIP, VAE, and source image.
- Enter your positive prompt.
- Start with the default two-cascade setup.
- Queue the workflow.
- When the Tail 5 Continuation Gate appears, choose the best tail frame.
- Press
Resume Cascade / Continue. - Wait for the final stitched video and report.
primary_modelclipvaeimagesource_image_file
Optional:
secondary_modelmask
In a Wan High/Low setup, primary_model is usually the high-noise / structure model and secondary_model is usually the low-noise / refinement model. If secondary_model is not connected, the node can fall back to the primary model for the second stage.
How many segments to generate.
1 = one normal clip
2 = one clip plus one continuation
5 = current public-alpha maximum
Controls where the Tail 5 Continuation Gate appears.
For the first test, keep:
cascade_count = 2
pause_after_cascade_1 = true
Frames generated per segment. The public default is 49.
Default: Observe Only.
Public-safe modes:
Observe Only: report evidence, no intentional tensor mutation.Latent Delta Scale: explicit delta testing mode.Tail Source Reconstruction: report-only tail/source continuation evidence.
Research modes are available for experiments, but they can change output quality and should not be treated as safe defaults.
Neutral value is 1.0.
In Observe Only, these values are recorded but not used as active generation control.
Default: Report Only.
Report Only keeps the user prompt clean and records prompt/strategy evidence in the report.
Default: OFF.
Enable Shadow Step Trace only for diagnostics.
Singularity can save a Markdown report beside the video.
Useful report signals:
result_status: whether a final output was produced.CompletionGate: whether the requested route completed.cascade_progress: which cascade segments completed.cascade_seam_impulse: whether the stitch boundary showed a visible jump.tail_next_source_continuity: whether the next segment needs better source/tail inheritance.seam_phase_classifier: why the boundary may have jumped.PublicReleaseReadiness: whether the current settings look public-safe or research-like.
Important: CompletionGate = PASS means the structural route completed. It does not mean the video is visually perfect. Always inspect the MP4.
The project reads generation as an event:
Outcome(t-1) + ObservedBehavior(t-1)
=
Strategy(t)
=
ObservedBehavior(t+1) + Outcome(t+1)
For normal users, this means:
source image
+ prompt meaning
+ model interpretation
+ sampler route
+ latent motion
+ visible video
should keep describing the same event as the cascade continues.
The formula is not a magic quality button. In the public build, it mostly provides diagnostics and safer structure for experiments.
- This is a public alpha, not a final production release.
- Current public cascade limit is 5.
- Infinite cascades and prompt-per-cascade scheduling are future work.
- The node is Wan-first.
- Models are not included.
- Heavy resolutions require more VRAM and time.
- Always inspect the final video.
- Research modes can produce artifacts.
Place this folder in your ComfyUI custom_nodes directory:
ComfyUI/custom_nodes/Singularity
Then restart ComfyUI.
cascade_count = 2
pause_after_cascade_1 = true
frames_per_cascade = 49
fps = 16
seed = 123
math_control_mode = Observe Only
high_delta_strength = 1.0
low_delta_strength = 1.0
sampler_trace_mode = OFF
prompt_transcode_mode = Report Only
save_video = true
save_report = true
Queue the workflow, choose a tail frame at the pause, continue, then inspect the final MP4.