ComfyUI-GTE

Generative Trajectory Explorer

A ComfyUI node pack for stopping a generation part-way, looking at it, and branching from there.

A normal generation runs from pure noise to a finished image in one go — you pay full price before you know whether the seed was any good. These nodes let you stop at intermediate checkpoints, inspect the candidates, keep the directions worth continuing, and create controlled variations from them. You only pay full price for the branches you keep.

STEP A   DISCOVER    0 -> A     many seeds, cheap. Pick a direction.
STEP B   BRANCH      A -> B     vary the one you picked. Pick again.
STEP C   BRANCH      B -> C     vary again, more finely. Pick again.
FINAL                C -> end   finish the one you chose.

The unit of work is a partial latent plus the noise level it sits at. A latent without its sigma is meaningless, and that idea is why every node here is built the way it is.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/mgkgng/ComfyUI-GTE

Restart ComfyUI. No extra Python dependencies — everything runs on the torch, numpy and Pillow that ComfyUI already ships.

The example workflow needs three other packs

The nodes above have no dependencies beyond ComfyUI itself. The example workflow is a different matter — it is built with helpers from:

| pack | used for | |---|---| | rgthree-comfy | Any Switch, Power Lora Loader, KSampler Config, the group bypasser and the mute/bypass relays | | ComfyUI-KJNodes | SetNode / GetNode — the virtual wires that carry the shared controls into every stage | | ComfyUI-Easy-Use | easy int, easy seed, easy ifElse (the lazy guider switch) |

Install those if you want the shipped graph. If you are only after the nodes, you need none of them.

The loaders ship empty — pick your own checkpoint, or UNET + CLIP + VAE. Nothing in these nodes is tied to a model family. Set the CLIP type to match whatever you load.

Then load workflows/GTE-example.json and read WORKFLOW_GUIDE.md, or the notes on the canvas itself.

The nodes

| node | what it does | |---|---| | Seed Range Noise | a batch where item i uses seed + i as a full independent reseed, so each candidate has a real, reproducible identity | | Sigma Segment | slices an absolute range out of one immutable schedule: sigmas[start : end+1]. Adjacent segments share their boundary sigma, which is what makes continuation seamless | | Stamp Step / Resume Step | writes the position onto the latent itself, so the next stage reads where the latent actually is rather than where the wiring assumes | | Candidate Select | a cursor over a batch: click a tile, see it large, keep one. Reports that candidate's own seed | | Noise Rotate | controlled descendants of a partial latent — theta sets how far from the parent, spread how widely the siblings fan out | | Schedule Info | how much noise is left at every step, before you spend a run on it | | Show Text · Lines | puts that table on the canvas |

Two ideas worth knowing

Rotate the noise, never add to it. x + d·eps inflates the noise level by sqrt(1+d^2) — +41% at d=1 — while the scheduler still believes the old sigma. The sampler then under-denoises, and variation strength and quality loss end up tangled in one dial with no error anywhere. Rotating preserves ||x - x0|| exactly, so theta moves the direction and never the amount.

Ask the model, never name it. x - x0 is the noise only for additive models (SD1.5/SDXL). Rectified flow (Flux, SD3, krea) mixes noise in instead, so the same subtraction yields noise and signal. There is no list of model families anywhere in this pack: Noise Rotate probes the model's own noise scaling twice to recover its noise coordinate, which works for every parameterisation ComfyUI ships and for ones that do not exist yet.

Constraints, measured

• euler only. Splitting a trajectory works only for samplers whose step depends on (x, sigma) alone. dpmpp_2m carries derivative history and degrades; dpmpp_2m_sde injects its own noise and breaks outright. euler, heun and ddim are safe.
• Branch where there is noise left to branch on. Read the Schedule Info table rather than trusting step numbers — the same step means very different things under different schedulers.
• Keep the variation seeds different from the origin seed. If they collide, the "new" noise is the tensor the candidate was born from, and both the angle and the noise level quietly stop meaning what they say.

Tests

ComfyUI/venv/bin/python tests/test_gte.py     # standalone, no pytest needed
pytest tests/                                  # if you have it

Every test corresponds to something that actually broke.

A note on node IDs

Node IDs are prefixed GTE_ (GTE_NoiseRotate, GTE_SigmaSegment, …) and the shipped workflow refers to them by those IDs. If you have graphs saved against an earlier build of these nodes under a different prefix, they will not resolve — open them, and re-add the affected nodes.

License

MIT