Kforge Labs automations

Turn model tests into evidence.

The automation loop keeps ComfyUI renders, workflow JSON, tracking cards, iteration logs, and promoted findings moving through one repeatable path.

Mental model

One project folder, one repeatable loop.

ComfyUI still owns creative iteration. The automation layer gives every render a traceable card and a durable log, so the next decision is based on what actually happened.

1

Tweak

Change one knob in ComfyUI and update the run notes.

2

Save

Export workflow JSON into the working folder.

3

Render

Write the video beside the JSON with the same run identity.

4

Card

Run /make-card to create the visual run summary.

5

Promote

Move stable patterns from the iteration log into findings.

Data contract

The workflow tells the script what matters.

Three Text Multiline nodes are enough to connect ComfyUI, the repo tools, and the per-project evidence folder.

Per-project working folder
<working folder>/
  LTX-Director_v04.json
  LTX-Director_v04.mp4
  more JSONs and MP4s
  _AI/
    card.png
    iterations.md
Required Text Multiline nodes

Working_Folder_PATH

Absolute path to the project folder. This wins over the repo .env fallback.

OVERLAY - INFO

Run summary plus a BASE (notes): section for what changed this time.

OVERLAY - FEEDBACK

Post-render observations and scores like motion: 4/5, sync: 5/5, sharp: 4/5.

Stable card filename

_AI/card.png is overwritten each run. Wire your NLE to it once and keep comparing.

Append-only log

_AI/iterations.md grows one row per unique workflow save, with duplicate runs filtered.

Repo stays clean

Per-project JSON, MP4, cards, and logs live outside the repo. Docs and findings live here.

Model folders scale

Each model gets its own docs folder under docs/automations/; shared rules stay in CONVENTIONS.md.

Inside ComfyUI

Write the run context where the workflow can carry it.

Before rendering, update OVERLAY - INFO. After watching the video, update OVERLAY - FEEDBACK. The saved JSON becomes the source of truth for the card.

Workflow notes
OVERLAY - INFO
Phase 2 Scheduler: linear_quadratic Steps: 8 Denoise: 0.40 BASE (notes): Lower denoise from 0.80 to preserve stage 1.
OVERLAY - FEEDBACK
Cleaner mouth movement on fast turns. Background retains 2K structure. motion: 4/5 sync: 5/5 sharp: 4/5

In the agent

One command turns the latest JSON into a review artifact.

/make-card finds the newest non-loop workflow JSON, extracts node parameters, writes the card, and appends the iteration log.

Command output
$ /make-card
Reading newest workflow JSON...
Writing _AI/card.png
Appending _AI/iterations.md

Card ready:
/JOBS/scene/ComfyUI-working-folder/_AI/card.png

When evidence settles

Promote patterns so they stop being rediscovered.

When a knob behaves the same way across several runs, move the conclusion into the model's findings.md. The loop gets smarter because the docs shrink toward truth.

LTX 2.3 findings snapshot
Scheduler linear_quadratic with 8 steps is the current locked path.
Sampler / CFG euler and CFG 1.0 stay fixed while sweeping other knobs.
Resolution 2K is the recommended target for multi-segment work; 4K transitions currently dissolve.
Audio mode Model-generated audio makes Phase 2 denoise visible as lip-sync drift.