File size: 4,109 Bytes
48cbcaa | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 | ---
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).
|