
触发词
文件清单 线路:原始(huggingface.co) · 切换
下载后放进对应的 ComfyUI 目录。models/ 指 ComfyUI 的模型根目录。
说明
VegaNec Attention Director | Eyes That Know What Matters | Krea 2
作者:neculaalexandru86749
VegaNec Attention Director is a positive-direction creator utility LoRA for Krea 2. It is designed to make what one fictional adult is attending to more readable by improving the agreement among eyes, pupils, head orientation, posture, expression, and one visible or implied target.
This is not an identity model and does not create one fixed person. It is a composition-support tool for scenes where a technically attractive subject still looks vacant, stares into the camera by default, or turns the head without a convincing eye line.
The model supports:
- deliberate camera or viewer engagement;
- focused reading, inspection, controls, screens, documents, and hand-guided practical work;
- attention toward one off-camera person, sound, alarm, or event;
- readable left, right, upward, downward, and diagonal glances;
- over-shoulder awareness with visible eyes and target space;
- portrait, waist-up, three-quarter, full-body, and environmental framing;
- fictional adult women and men across varied ages, skin tones, body classes, wardrobes, lighting, and environments.
Trigger: veganec_attention_director
Recommended strength: 0.8
Useful range: 0.8 release-validated; other strengths unverified
For best results, describe exactly one adult and exactly one attention target. Name the target, where it sits relative to the subject, and the observable behavior that connects them:
one fictional adult, eyes and pupils directed upward-right toward one amber cue light, compatible head angle and attentive posture
For an implied off-frame target, state one direction and one cause:
eyes, pupils, head orientation and paused posture turn together toward one unseen alarm beyond the upper-left frame
Keep the scene contract simple. One subject, one target, one direction, and one readable action are more reliable than multiple alternative targets or dense camera language.
Attention Director does not claim medical gaze correction, exact pupil coordinates, perfect eye anatomy, arbitrary vector control, multi-person eye-line choreography, identity replacement, general pose control, or ControlNet replacement. It can still inherit normal generation failures in eyes, hands, anatomy, target count, and scene composition.
The exact epoch-10 adapter at strength 0.8 passed six fixed unseen, identical-seed comparisons covering direct viewer engagement, practical task focus, an implied off-camera alarm, a visible directional target, over-shoulder awareness, and a low-light diagonal cue. Fourteen additional owner-approved launch scenes then demonstrated broader identities, environments, framing, target classes, and attention directions without a recurring shortcut. The larger twenty-holdout/four-strength statistical study remains unperformed and is not claimed.
Every gallery subject is a fictional adult aged 21 or older. The model is not based on an actual person.