Composing EBM models at inference time: AND, OR, NOT with energy arithmetic
daibond_alpha · x · 2026-11-20
A five-part thread explaining inference-time composition of energy-based models (EBMs):
- Logical operators on energies: AND = sum energies, OR = log-sum-exp, NOT = negate energy. Learned models — from diffusion models to attribute generators — can be combined without retraining.
- Demonstrations (from the figure): composing two diffusion models, and Boolean composition of face attributes such as Smiling AND Male or Smiling AND NOT Male.
- Unified view: best-of-N sampling, verifiers, and RL fine-tuning all target one KL-regularized EBM, differing only in whether you pay the cost at training time or at inference time.
More from Multimodal
- Microsoft launches Canvas in MAI Playground powered by MAI Image 2.7 Flash — testingcatalog · 2026-11-20
- Ideogram 4.6 lands in HeyGen with precise edits, pairs with Avatar 6 — HeyGen · 2026-11-20
- Building a WebGPU pinball machine with Gemini: 480Hz physics, zero audio samples — AI_Andrew · 2026-11-20
- Same Prompt, Different Images? Fix the Seed for Reproducible Generation — Al_Grigor · 2026-11-20
- GPT 6.2 Astra paints a photoreal landscape entirely with code, no image generator — mark_k · 2026-11-20
- 4-Step MOSS SFT Distill LoRA Workflow Delivers Stunning Results — aimasterguru · 2026-11-20