Root cause: IPAdapter CLIP Vision embeds from a character with saturated
red/copper hair tint the entire scene warm/pink, regardless of background
location. Tested systematically: with all IPA weights=0, backgrounds are
correctly varied (dark city, grey ruins, blue night). With any IPA > 0,
all backgrounds shift warm.
WF21 (template): reduce IPA PLUS 0.5→0.3, FACE 0.75→0.4, blend 0.45→0.3;
add color bleed troubleshooting note to in-canvas instruction node.
WF24 (Evi Magini specific): disable all IPA (weights=0.0); detailed text
description maintains character consistency without IPA color bleed.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- Replaces API-format JSON (no visual properties) with full LiteGraph format
- All nodes now have pos, size, color, bgcolor fields
- Nodes arranged left-to-right: user edits → IPA → CLIP → KSampler → VAE+Florence2 → Save
- 5 group boxes per workflow with color coding and labels
- PrimitiveStringMultiline instruction note per workflow with in-canvas guidance
- WF21 (char): 21 nodes, 34 links — 3-stage IPA face chain
- WF22 (outfit): 18 nodes, 27 links — single IPA PLUS
- WF23 (loc/obj): 18 nodes, 27 links — single IPA PLUS, location/object wildcards
- All tested: validate + queue successfully against ComfyUI 192.168.50.199
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Previous weights (w=1.0) caused color bleeding and body distortion.
Values calibrated empirically against a real character reference:
IPA PLUS (body/style): w 1.0→0.5, end_at 0.65→0.55
IPA FACE (style xfer): w 1.0→0.75, end_at unchanged 0.15
IPA FACE (blend): w 0.7→0.45, end_at 0.65→0.50
KSampler: steps 20→25, cfg 4.5→3.5
Result: clean cel-shaded full-body with correct hair/outfit/accessories.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Fixes from code review:
- Replace PixaromaNote (rare custom node) with PrimitiveStringMultiline (ComfyUI core)
so instruction notes work without extra dependencies
- WF21: Florence2 fp32→fp16, keep_model_loaded false→true,
detailed_caption→more_detailed_caption, max_new_tokens 512→1024, do_sample false→true
(aligns WF21 with WF22/23 — faster, less VRAM, richer captions)
- WF22: fix default description that ended with incomplete "wearing [nothing]"
- wildcards.py: add outfit_background list (simple/neutral bg) used in outfit variety prompt
so strój stays in focus; add "white background, studio lighting" suffix for object type
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
