Research map / Large language models / Reasoning and chain of thought

Test-Time Instance-Specific Parameter Composition: A New Paradigm for Adaptive Generative Modeling

CVPR 2026 · Large language models › Reasoning and chain of thought · approach: Test-time compute scaling · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes Composer, a transformer meta-generator that produces per-input low-rank weight updates for frozen diffusion/VAR backbones at inference, improving FID with negligible overhead versus test-time training.

Problem

Generative models use a single fixed set of weights for all inputs, and test-time training to adapt per instance is prohibitively expensive for high-resolution diffusion and VAR backbones.

Approach

A lightweight transformer generator conditioned on the prompt/class and the backbone's own weight-derived tokens outputs low-rank updates (AB) added to W_Q and W_V in one pass before multi-step generation, trained with a context-aware batch sampling scheme that mixes similar and dissimilar conditions.

Key results

• On ImageNet 256x256 VAR d-16, Composer reaches FID 3.15 vs 3.55 standard and 3.22 test-time training, with time +5% vs +10,030% for TTT.
• On SD2.1 COCO 2014, Composer lowers FID-30K from 13.45 to 13.07 and raises CLIP-30K from 0.30 to 0.32 at +2% inference time.
• On 2/8-bit Q-Diffusion, Composer raises IS from 49.08 to 78.21 and lowers FID from 43.36 to 35.26.

Surprising findings

• Composer's gains appear as large improvements for extremely low-bit quantized models, where baseline quality collapses.
• Test-time training with random similar samples performs worse than Composer while costing over 100x more time on VAR backbones.

Limitations

• Adds small inference time and memory overhead, larger relative overhead for few-step models such as one-step diffusion.
• Results are reported only on image generation; extension to other tasks is future work.

Experimental setup

Datasets: ImageNet MS-COCO 2014

Benchmarks: ImageNet 256x256 ImageNet 512x512 MS-COCO 2014 30K

Metrics: FID IS precision recall CLIP score inference time peak memory GenEval

Baselines: Test-time Training Q-Diffusion CTEC ORM PARM

Builds on: LoRA test-time training Q-Diffusion CTEC

Closest papers by meaning

• Dual Prototype Evolving for Test-Time Generalization of Vision-Language Models (NeurIPS 2024 · similarity 0.85)
• Generative Adapter: Contextualizing Language Models in Parameters with A Single Forward Pass (ICLR 2025 · similarity 0.85)
• Noise Hypernetworks: Amortizing Test-Time Compute in Diffusion Models (NeurIPS 2025 · similarity 0.84)
• Why Knowledge Distillation Works in Generative Models: A Minimal Working Explanation (NeurIPS 2025 · similarity 0.84)
• DynaPrompt: Dynamic Test-Time Prompt Tuning (ICLR 2025 · similarity 0.84)
• Adaptive Divergence Regularized Policy Optimization for Fine-tuning Generative Models (NeurIPS 2025 · similarity 0.84)