Add optimized workflow notes layer

2026-08-08 21:35:04 +00:00
parent e9efbc4616
commit 5cdfcad277
5 changed files with 147 additions and 0 deletions
@@ -0,0 +1,50 @@
+# FLUX.2-klein Notes
+
+## Current default assumptions
+
+- Chris uses `FLUX.2-klein 9B`.
+- Prefer `9B fp8` paths first.
+- `GGUF` is not the default path unless it is the only strong or practical workflow route.
+
+## Strong current source workflows
+
+- Official Comfy `image edit` templates:
+  - `image-edit-9b-base`
+  - `image-edit-9b-distilled`
+  - `image-edit-4b-base`
+  - `image-edit-4b-distilled`
+- Current best outpainting starter:
+  - `models/flux2klein/image-edit-4b-distilled/`
+  - Reason: official, current, and easiest first test even though Chris's main stack is `9B`
+- Community Hugging Face workflow worth remembering:
+  - `models/flux2klein/community-gguf-kv-yarvixpa/`
+  - Interesting because of KV-cache / GGUF optimization ideas, not because it is the default path
+
+## Strong current heuristics
+
+- Prefer official Comfy templates as the baseline unless a community workflow clearly improves something real.
+- For `9B`, watch closely for text-encoder mismatches.
+- Community `9B` optimization notes often revolve around:
+  - low-step usage
+  - `CFG=1`
+  - cleaner model/text-encoder pairing
+  - avoiding broken loader substitutions
+
+## Current weak areas
+
+- We still do not have a clean, trusted, explicitly labeled `9B fp8 outpainting` community workflow import.
+- Some public workflow pages mention outpainting but do not expose a trustworthy raw workflow asset.
+
+## Best current optimization direction
+
+- Build future `optimized/image-edit/flux2klein-9b-master/` from the official `9B` edit template family first.
+- Borrow only proven improvements from community sources:
+  - cleaner graph organization
+  - low-step settings when they are clearly validated
+  - better reference/outpaint control if portable
+
+## What needs Monday proof
+
+- Whether `9B distilled` or `9B base` is the better real default for edit/outpaint on Chris's setup
+- Whether a `4B distilled` outpaint starter still beats `9B` in practical iteration speed enough to matter
+- Which custom-node optimizations are real versus noise