ComfyUI / redux_experiment_final /claude_memory /user-workflow-preferences.md
aleph65's picture
Archive flux-redux experiment: final writeup, workflows, inputs, scripts, research, v4 outputs, comparison sheets, Claude memories
478cb8f verified
|
Raw
History Blame Contribute Delete
1.6 kB
metadata
name: user-workflow-preferences
description: How plx (this pod's user) likes ComfyUI workflow projects run
metadata:
  node_type: memory
  type: feedback
  originSessionId: 749cf677-e729-4f5c-b2c7-375816bfaee4
  modified: 2026-08-04T14:17:45.482Z

The user (plx, RunPod ComfyUI power user, plx9790@gmail.com) runs iterative workflow R&D.

Why: established over the Aug 2026 flux-redux project ([[flux-redux-experiment]]), where this pattern produced excellent results and explicit praise.

How to apply:

  • "Don't start yet / just tell me" is a hard stop — present analysis or a plan and wait for an explicit green light. They often think out loud between green lights.
  • Research before fixing: when outputs look wrong, they want root-cause research (web agents + local diagnostics) before parameter fiddling.
  • Prompts: they want individually handwritten, per-run prompts (target content + style aesthetic fused), not one generic boilerplate prompt.
  • Versioned outputs: outputs_v2/v3/v4 folders per round, seed 777 for comparability, full r×t
    • ordered r×r matrices, audit saves (crops, depth maps), labeled comparison contact sheets.
  • Everything durable goes to HF aleph65/ComfyUI (workflows/, wip// for drafts, models/ mirror); workflows carry MarkdownNotes documenting models, knobs, and gotchas.
  • Standardize on portrait 3:4 (896x1152 bucket, curated inputs at 1152x1536); no quantized checkpoints — bf16 unets + fp16 T5; maximum quality over speed.
  • Adult-content glamour/swimwear generation is the use case; lawful, their own reference sets.