Eric Qwen Direction Average From Folder

The bulk version of Direction Average, and the primary tool the docs point at for building a direction library. Where Direction Average handles 2–4 live directions, Direction Average From Folder sweeps up every saved direction file in a directory, averages them, normalizes the result, and hands you one clean direction - or saves it straight to your library.

This is the "14 pairs of Silvermax conversions" workflow in a single node. You save each pair's direction as silvermax/01, silvermax/02, …, then point this node at the silvermax folder and get silvermax_final back.

The inputs

• folder - the dropdown. It lists subfolders of models/qwen_directions/ and refreshes on every run. The (root folder) option averages everything in the top level.
• prefix - optional filename filter, for averaging only prefix_* files inside a folder.
• normalize - on by default here (unlike the 2–4 node). Averaging shrinks magnitude, so rescaling is the sane default.
• target_rms - default 2.0. The rescale target: 1.0 unit, 2.0 moderate, 3.5 ≈ original single-pair strength.
• save_as - optional filename to persist the averaged result right there (subfolders supported, same as Direction Save).
• save_as_bf16 - halve the saved file size.

Output is the averaged direction.

The console numbers that tell you it worked

This is one node where the console output is genuinely informative:

[EricQwenDir]   mean individual rms=3.8320
[EricQwenDir] averaged 14 directions | rms=1.0625
[EricQwenDir] normalized: rms 1.0625 → 2.00 (scale=1.883)

The ratio of averaged RMS to mean pair RMS is your purity measure. 1.06 / 3.83 = 28% - high purity, meaning most of the magnitude was image-specific and cancelled. If the ratio comes out near 1.0, your pairs weren't diverse enough and you're averaging near-identical directions.

The library-building pattern

# 14 pairs: encode color ↔ B&W, compute + save each
Direction Average From Folder(folder="silvermax", normalize=True, target_rms=2.0, save_as="silvermax_final")
Direction Load("silvermax_final") → Direction Apply(base_cond, scale=1.0) → Conditioned Edit

Diversity guidance from the docs: mix skin tones, hair colors, lighting, backgrounds, focal distances - every covered dimension is one that won't dominate the average.

Installing it

ComfyUI Manager → "Eric Qwen-Edit", or:

cd ComfyUI/custom_nodes
git clone https://github.com/EricRollei/Eric_Qwen_Edit_Experiments.git

Restart. Loading saved direction files needs no model; the full graph needs a Qwen-Image-Edit pipeline via the pack's loader (Qwen/Qwen-Image-Edit-2511, ~54 GB).

Gotchas

• Empty folder in the dropdown? You haven't saved anything into that path, or the files are elsewhere - it only scans models/qwen_directions/.
• Zero-pad your filenames. 01, 02, ... 10 sorts correctly; 1, 2, ... 10 puts 10 next to 1. The docs flag this for a reason.
• Purity ratio near 1 = wasted run. Rebuild the pairs with more diverse images rather than trusting the average.