Singularity R178

If you've ever made a Wan clip longer than ~5 seconds, you know the ritual: render segment one, grab its last frame, feed it back into an I2V node as the new start image, render segment two, and pray the two clips don't visibly jump. It works, and it's miserable. Singularity ("Singularity R178") is a public-alpha ComfyUI node that wraps that whole loop into one node - and then does one genuinely clever thing on top: at each segment boundary it pauses and shows you the last five frames so you can pick which one the next segment continues from, instead of chaining off whatever frame the model left you with.

It's Wan-first, one node, and early. The whole ComfyUI-Event-Equality-Core pack registers exactly one visible node; everything else is internal machinery. Treat it as an experiment you can actually run, not a production tool.

How it works

Under the hood the node runs a cascade: it samples a segment, extracts five tail candidate frames, and (when a pause is enabled) stops the run and shows a detached panel - Source | Tail 1 | Tail 2 | Tail 3 | Tail 4 | Tail 5 | Result. Click the frame with the best pose, face, or motion direction, hit Resume Cascade / Continue, and that frame becomes the source state for the next segment. Same run, no manual file juggling, one stitched MP4 at the end.

Sampling itself follows the Wan 2.2 high/low pattern you already know: a high-noise stage on primary_model (the structure/motion pass, default steps 0→1) and a low-noise refinement stage on secondary_model (default steps 1→4). If you don't wire a secondary_model, it falls back to the primary for the second stage. The high/low terminology maps directly onto Wan's own two-expert setup.

Here's the honest part: the project dresses this up in a lot of "event equality formula" language. In the public default (math_control_mode = Observe Only) that layer mostly records evidence into a report instead of mutating tensors. The author says it themselves: the formula is not a magic quality button. You're getting a convenient cascade loop plus diagnostics, not a new sampling algorithm.

The inputs that matter

The node has a big surface, but most of it hides in LAB/research groups. For your first run, set these:

• primary_model, clip, vae - standard Wan wiring. Models are not included, so you need a Wan 2.2 checkpoint, its CLIP, and VAE.
• source_image_file - the reference image for I2V. It uses ComfyUI's native image upload, so you can pick from your input folder directly.
• cascade_count (default 2, max 5 in this alpha) and pause_after_cascade_1..4 - where the Tail 5 gate appears. Keep cascade_count = 2 and pause_after_cascade_1 = true for your first test.
• frames_per_cascade (default 49), width/height (default 704×1280), fps, seed.

The default negative prompt is a Chinese Wan-style one; leave it. math_control_mode has 27 options, but only Observe Only, Latent Delta Scale, and Tail Source Reconstruction are billed as public-safe - the rest are research modes that can change output quality.

Outputs

You get status, saved_video_path, saved_report_path, and report as STRINGs, plus tail_frame_0 through tail_frame_4 as IMAGEs you can route to a preview. When save_report is on, a Markdown report lands next to the video with CompletionGate, cascade_progress, seam diagnostics, and more. Read it, but remember CompletionGate = PASS is structural proof, not a quality verdict - always inspect the MP4.

Install

ComfyUI Manager is the easy route - search "Singularity - Tail 5 Continuation Gate" and hit install. Or go manual:

cd ComfyUI/custom_nodes
git clone https://github.com/DIIIUA/ComfyUI-Event-Equality-Core

Then restart ComfyUI. The pack's requirements.txt declares no extra Python dependencies - it runs on ComfyUI's own runtime. One real gotcha the README doesn't advertise: video saving prefers VHS_VideoCombine, i.e. the VideoHelperSuite pack. There's an emergency fallback writer if VHS is missing, but installing VHS through Manager gives you clean MP4 output.

Troubleshooting

• OOM at 704×1280? That's the default and it's heavy. Drop to 416×608 for a fast debugging pass - the README recommends it explicitly.
• The pause panel didn't appear? Check pause_after_cascade_1 is on. The panel stays detached under the node while paused and vanishes once the run finishes - easy to miss.
• Weird output? If you flipped math_control_mode away from Observe Only, or enabled strategy_field_mode, sampler_trace_mode, or auto_calibration_mode, revert them. They're real research modes, not safe defaults.
• Fuzzy stitches between segments are what picking a better tail frame is for. The report's cascade_seam_impulse and tail_next_source_continuity fields tell you whether the boundary jumped.

It's alpha and niche, but the Tail 5 picker is the part that's hard to fake with stock nodes - and it's genuinely the right fix for the 5-second problem.