Struct Sam3Out 

pub struct Sam3Out {
    pub text_embeds: Vec<f32>,
    pub text_feats: Vec<f32>,
    pub trunk_feats: Vec<f32>,
    pub fpn: Vec<Vec<f32>>,
    pub enc_hidden: Vec<f32>,
    pub dec_hidden: Vec<f32>,
    pub pred_boxes: Vec<f32>,
    pub pred_masks: Vec<f32>,
    pub pred_logits: Vec<f32>,
    pub presence_logit: f32,
}

One forward’s outputs (batch 1), flat row-major f32.

Fields

text_embeds: Vec<f32>

CLIP text projection of the pooled token, [512].

text_feats: Vec<f32>

Resizer output fed to the DETR stack, [32 * 256].

trunk_feats: Vec<f32>

Trunk features, [5184 * 1024].

fpn: Vec<Vec<f32>>

FPN levels 288/144/72/36, each [256 * H * W].

enc_hidden: Vec<f32>

DETR encoder output, [5184 * 256].

dec_hidden: Vec<f32>

Stacked decoder layer outputs, [6 * 200 * 256].

pred_boxes: Vec<f32>

Final boxes (xyxy), [200 * 4].

pred_masks: Vec<f32>

Instance masks, [200 * 288 * 288].

pred_logits: Vec<f32>

Per-query classification logits from the dot-product scorer (det.scoring.*), [200] — the final score of a query is sigmoid(pred_logits) * sigmoid(presence_logit) (HF’s post_process_object_detection). Empty when SAM3_ONLY skips the decoder.

presence_logit: f32

Presence-token logit of the final decoder layer (clamped ±10). 0.0 when the decoder did not run.