TF Sweep Edit

The question this node exists to answer is the one every edit session eventually asks: was that change my edit, or just the seed? A single before/after pair can't tell you - re-sampling from an edited canvas is still sampling. TF Sweep Edit runs one edit once per value of a single axis - a seed list, a set of levels, or a strength ramp - with everything else pinned, and tabulates the results. That's what makes it an experiment rather than a loop: no two arms differ in more than one way.

The design that earns the word "experiment" is the per-arm baseline. Every arm gets the same trajectory resumed at the same level with the same seed and no edit, and each arm is measured against its own baseline. Without that, a seed sweep mostly measures the seed - "the image changed a lot" becomes indistinguishable from "the seed changed a lot." baseline (on by default) is what buys you the honest number; turning it off folds the resume seed into every figure, which the tooltip admits is only defensible for a strength sweep. Cost scales honestly: two re-samples and one decode per arm, so four arms is a few seconds and forty is a coffee break.

Inputs break into the edit itself and the sweep axis. axis is the one thing that varies - seed, level (l*), or strength - and values supplies it as comma/space-separated numbers with a-b inclusive ranges (1,2,3,4 is the default; duplicates are dropped). Everything else is pinned to its widget below: level, seed, strength, source_mode, target_tokens, source_tokens. arm_limit (default 12) refuses to start if values expands past it - a genuine guard against a mistyped 0-1000 pinning your GPU for an hour, not a real ceiling. decode (on) renders each arm's final level for the contact sheet; off uses free PCA tiles instead and the table is unaffected. output_arm names which arm leaves on the levels output so you can save or compare one result.

Wire a TF Region Map into regions and the node sweeps a shape edit instead - target_tokens are handed to the region named by source_tokens. One constraint follows from geometry: a region map describes a single level, so only seed and strength are available as axes and level is refused with that reason. Sweeping l* is the one axis that can't hold everything fixed - a selection snapped to one level's regions isn't a whole region at another - so the node keeps the token set fixed, which is the only sane meaning of "the same edit at every level," and says so in the report rather than silently doing something else.

The number to actually read is the spread across arms: mean pairwise cosine distance at the final level. Near zero means the edit decides the outcome, so a single-seed result was trustworthy. Large means the seed does, and it wasn't. The report output is the whole table, and the sheet output is the contact sheet with the no-edit baseline first - both exist to feed the save nodes, because a sweep's numbers outlive the tab.

Install

Pack-standard: Manager → search Trajectory Forcing, or git clone https://github.com/korayulusan/ComfyUI-TrajectoryForcing into custom_nodes/. JAX inside ComfyUI needs its own Python 3.11 env on CUDA 12; requirements.txt is empty on purpose, install.py adds the JAX stack or declines with a reason (normal), and README's env/setup.sh is the fallback. Health check:

cd ComfyUI/custom_nodes/ComfyUI-TrajectoryForcing
python -m tf_nodes.doctor

A sweep is only as meaningful as its inputs are pinned - and the pack's own example workflow 05 is exactly this graph, ready to load and run if you want to see the spread before you trust a single-seed conclusion.