| --- |
| language: en |
| license: apache-2.0 |
| base_model: black-forest-labs/FLUX.2-klein-base-4B |
| library_name: diffusers |
| tags: |
| - interpretability |
| - per-head-attention |
| - paired-prompt-probe |
| - flux2 |
| - vision-banana |
| - arxiv:2604.20329 |
| pipeline_tag: image-to-image |
| --- |
| |
| # ghost-plantain |
|
|
| A per-head attention probe of FLUX.2 Klein 4B testing whether the base model represents amodal completion (the entire object including its occluded extent) as a separable axis from modal description (visible portion only) on identical occluder relationships. |
|
|
| ## Thesis |
|
|
| Amodal completion — the perceptual operation of inferring an object's full extent from its partial visibility — is a primitive of biological vision long predating deep learning. Whether image-generation models recover an explicit modal/amodal distinction at the representational level is a more specific question than whether they can render hidden portions of objects on demand. ghost-plantain tests whether Klein has a per-head representational axis that systematically responds to the modal-vs-amodal scope of a request, on otherwise identical scene descriptions. |
|
|
| ## Method |
|
|
| Twenty-five paired prompts holding the depicted scene constant. The A condition (modal) describes the visible portion only ("show only the visible part of the apple behind the cup"). The B condition (amodal) requests the entire object including the occluded extent ("show the entire apple including the part hidden behind the cup"). Per-head capture protocol identical to the rest of the plantain probe family. |
|
|
| Rigor add-ons: per-head Cohen's d effect size; split-half consistency via 100 random 50/50 stimulus splits. |
|
|
| ## Results |
|
|
| | Metric | Value | Significance | |
| |--------------------------------|-----------------|---------------------------| |
| | Heads with \|t\| > 3 | 6,399 (39.2%) | 8.7× empirical null p99 | |
| | Heads with \|t\| > 5 | 2,537 (15.6%) | 507× empirical null p99 | |
| | Heads with \|d\| > 0.8 (large) | 4,128 (25.3%) | — | |
| | Split-half r (median) | 0.833 | [0.82, 0.84] IQR | |
| | Max \|t\| | 29.63 | — | |
|
|
| **Top blocks by max \|t\|:** |
| - joint[4]: max\|t\|=29.63, 132/192 heads at \|t\|>3, median \|d\|=1.06 |
| - single[0]: max\|t\|=24.98, 406/768 heads at \|t\|>3, median \|d\|=0.65 |
| - joint[2]: max\|t\|=24.97, 127/192 heads at \|t\|>3, median \|d\|=1.21 |
| - joint[3]: max\|t\|=23.95, 139/192 heads at \|t\|>3, median \|d\|=1.10 |
| - joint[1]: max\|t\|=22.52, 130/192 heads at \|t\|>3, median \|d\|=0.92 |
|
|
| **Interpretation.** The axis is strong (507× null at |t|>5) and highly stable (split-half r=0.83). Signal concentrates in joint MMDiT blocks (4 of the top 5 are joint), the cross-attention surface where text-image fusion occurs — consistent with the modal/amodal distinction being routed primarily through how the request modifies the text-image binding, not through a downstream image-only computation. The joint-block median Cohen's d ≥ 1.0 across the top blocks indicates that within those blocks the modal/amodal distinction is the dominant feature partition, not just one signal among many. A quarter of all 16,320 attention heads in the model show large-effect-size selectivity for this axis. |
|
|
| ## Status |
|
|
| Probe complete. No LoRA training; this is a base-model interpretability finding. |
|
|
| ## Limitations |
|
|
| The amodal phrasing ("the entire object including the hidden part") is linguistically marked; the modal phrasing ("the visible portion only") is a partial description. A residual contributor to the per-head signal could be the model encoding "complete vs. partial scope of request" rather than amodal completion specifically. A follow-up could pair an amodal request against a different complete-scope description (e.g., "the visible portion of the apple from a different angle") to disentangle. |
|
|
| Twenty-five pairs is small; the per-head t-vector reproducibility (r=0.83) is high but a larger pair count would tighten estimates. |
|
|
| The probe is correlational. |
|
|
| ## License |
|
|
| Apache 2.0 — matches base FLUX.2 Klein 4B. |
|
|
| ## References |
|
|
| - Gabeur, V., Long, S., Peng, S., et al. *Image Generators are Generalist Vision Learners.* [arXiv:2604.20329](https://arxiv.org/abs/2604.20329) (2026). |
| - Black Forest Labs. *FLUX.2 Klein.* https://bfl.ai/models/flux-2-klein (2025). |
|
|
