Trajectory Branch
Branch a rectified-flow sampling trajectory mid-flight and resume it with new settings, for ComfyUI (Anima family)
ComfyUI-TrajBranch
Branch a rectified-flow sampling run mid-flight — and resume the second half with different settings.
Two nodes for Anima-family flow models (ComfyUI 0.44 sampler contract, original clean-room code):
Nodes
Traj Branch Sampler (capture)
Runs a plain flow-Euler generation. At capture_step it snapshots
the latent — the branch point — and outputs all three:
latent— the finished generationbranch_latent— the state at the branch pointbranch_sigma— the noise level at the branch point
An optional sigmas input replaces the built-in linear-shift
grid (shift 3.0; do not stack with an AuraFlow patch).
Traj Resume Sampler (branch)
Continues from a branch latent. Wire branch_sigma into
start_sigma, set the remaining steps, and the second half runs
on the tail of the original grid. The node takes its own
model / positive / negative / cfg, so at the branch you can change
the prompt, the CFG — even the checkpoint — and only the second
half of the trajectory feels it.
The contract
Capture at step k, then resume with the remaining step count and
the same shift, reproduces the uninterrupted run (the resume grid
is the original grid's tail, recovered by inverting the
linear-shift parameterisation at the branch sigma; agreement is at
float32 grid precision, ~1e-7 in the sigma values — verified by a
numpy mirror test in tests/). Forking is therefore a controlled
act: keep the settings and nothing changes; change one and only
the tail moves.
Euler only, for now: momentum and AB2 carry velocity history, which a branch latent alone does not hold. Resuming those exactly would need the history exported alongside the latent.
Install
cd ComfyUI/custom_nodes
git clone https://github.com/seren-alegro/ComfyUI-TrajBranch
# restart ComfyUI
Or copy this folder as ComfyUI-TrajBranch into
ComfyUI/custom_nodes manually, then restart ComfyUI.
Example workflow
examples/traj_branch_identity_test.json - a minimal core-nodes-only
workflow: the Branch sampler captures at step 15 of 30 and the
Resume sampler finishes the remaining 15 steps with the same
settings, saving both results (traj_full / traj_resumed). If
capture + resume is working, the two images match. The checkpoint
name in the loader is a placeholder - point it at any Anima-family
checkpoint, and if your setup loads the text encoder / VAE from
separate files, swap that part of the chain accordingly. To fork
instead of verify, change the prompt or CFG on the Resume side only.
Tests
tests/test_traj.py — numpy mirror of the grid/capture/resume
math (no torch needed): grid-tail reproduction, capture-state
identity, capture+resume vs uninterrupted run.
License
MIT
한국어 요약
샘플링 궤적을 중간에서 분기하는 노드예요. Branch Sampler가 지정한 스텝에서 latent를 캡처하고, Resume Sampler가 그 지점부터 이어 뽑는데 — 뒤쪽 절반은 프롬프트·CFG·체크포인트까지 바꿀 수 있어요. 설정을 그대로 두면 무중단 실행과 같은 그림이 나오는 것을 수치로 검증했어요.