| --- |
| 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 |
| - human-pose |
| - flux2 |
| - vision-banana |
| - arxiv:2604.20329 |
| pipeline_tag: image-to-image |
| --- |
| |
| # dense-plantain |
|
|
| A per-head attention probe of FLUX.2 Klein 4B testing whether the base model represents canonical body pose as a separable axis from expressive pose deformation, parallel to the canonical-vs-deformed coordinate distinction in dense human-pose models (DensePose, SMPL). |
|
|
| ## Thesis |
|
|
| Dense human-pose representations canonically encode body shape in a fixed reference pose (T-pose) plus a deformation that maps to the actual pose. Whether image-generation models develop an analogous axis without explicit supervision is the question. dense-plantain pairs descriptions of a body subject in canonical T-pose against the same body subject in an expressive pose, and tests whether per-head attention shifts systematically. |
|
|
| ## Method |
|
|
| Twenty-five paired prompts holding body subject and setting constant; only pose configuration varies (canonical T-pose vs. expressive pose). 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 | 8,263 (50.6%) | 6.2× empirical null p99 | |
| | Heads with \|t\| > 5 | 4,614 (28.3%) | 659× empirical null p99 | |
| | Heads with \|d\| > 0.8 (large) | 6,164 (37.8%) | — | |
| | Split-half r (median) | 0.863 | [0.85, 0.87] IQR | |
| | Max \|t\| | 22.45 | — | |
|
|
| **Top blocks by max \|t\|:** |
| - single[0]: max\|t\|=22.45, 283/768 heads at \|t\|>3, median \|d\|=0.45 |
| - single[3]: max\|t\|=22.06, 389/768 heads at \|t\|>3, median \|d\|=0.61 |
| - single[9]: max\|t\|=21.66, 232/768 heads at \|t\|>3, median \|d\|=0.39 |
| - joint[3]: max\|t\|=20.18, 138/192 heads at \|t\|>3, median \|d\|=1.02 |
| - joint[4]: max\|t\|=19.62, 131/192 heads at \|t\|>3, median \|d\|=1.01 |
|
|
| **Interpretation.** The body-pose axis is strong (659× null at |t|>5) and highly reproducible (split-half r=0.86). Over half of all 16,320 attention heads register the canonical-vs-expressive pose distinction at |t|>3, and over a third reach Cohen's d > 0.8. Localization is mixed: max signal is in early single blocks (single[0]/[3]/[9]), but joint-block median |d| is highest (≥1.0 in joint[3] and joint[4]), suggesting pose information is initially routed through cross-attention text-image fusion (joint blocks) and then maintained through early single-block transformer processing. |
|
|
| The result places Klein among the models that have an internal canonical-pose representation without explicit pose supervision, parallel to but more diffuse than the structured axis encoded in DensePose-trained networks. |
|
|
| ## Status |
|
|
| Probe complete. No LoRA training; this is a base-model interpretability finding. |
|
|
| ## Limitations |
|
|
| The T-pose vs. expressive-pose pair has unmatched articulation complexity — expressive poses involve more anatomical regions deviating from canonical position than T-pose does. A residual contributor to the signal could be "amount of articulation" rather than canonical-vs-deformed status specifically. A follow-up controlling articulation (e.g., A-pose vs. expressive pose with matched limb configurations) would tighten the claim. |
|
|
| Body-type and clothing variations are not factored out across the 25 pairs. |
|
|
| 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). |
|
|
