Research map / Image generation / Text-to-image diffusion

DifFRACT: Diffusion Feature Reconstruction and Attribution for Circuit Tracing

NeurIPS 2026 · Image generation › Text-to-image diffusion · approach: Multi-concept composition · 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.

Extends transcoder-based circuit tracing to the double-stream MLP sublayers of FLUX.1[schnell] using timestep-conditioned transcoders, enabling exact feature-to-feature attribution graphs and circuit-guided interventions for diffusion transformers.

Problem

Multimodal diffusion transformers (MM-DiT) for image generation remain opaque; existing tools (attention maps, SAEs) do not reveal how features are transformed through nonlinear MLP sublayers or how semantics propagate across denoising steps and text/image streams.

Approach

Train timestep-conditioned (FiLM-modulated) transcoders per stream and block that approximate MLP sublayers of FLUX double-stream blocks, then build a local replacement model with frozen normalization and attention so that feature-to-feature attributions are exact and can drive circuit-guided steering.

Key results

• Transcoders achieve comparable or modestly better sparsity-faithfulness Pareto frontiers than SAEs at layers 6, 12, 18 in both streams.
• Text-stream contribution to the image-stream target falls from 89.9% at step 0 to 5.4% at step 3, and cross-modal edge fraction drops from 14.9% to 2.0%.
• Feature + context suppression of the red circuit yields the requested color across 30 seeds more reliably than baseline or single-feature suppression (exact numbers in figure only).

Surprising findings

• Amplifying the numeral 'five' shifts the mean apple count upward (6.30) rather than toward five, indicating diffuse rather than sharp numeral representations.
• Suppressing semantic text-stream features only affects generation at early denoising steps.

Limitations

• Analysis restricted to double-stream blocks of FLUX.1[schnell]; single-stream blocks not studied.
• Circuit steering results vary across seeds; some failures where spurious features are absent from the graph.
• Pruning and aggregation trade off graph completeness (about 30% higher conservation error).

Experimental setup

Metrics: L0 sparsity normalized MSE (nMSE) attribution sum conservation error fraction of seeds with correct color/count

Baselines: sparse autoencoders single-feature steering

Builds on: transcoders circuit tracing sparse autoencoders FiLM FLUX.1

Closest papers by meaning

• Why DiT Models Underperform as Representation Learners without Long Skip Connections (NeurIPS 2026 · similarity 0.90)
• Transcoders find interpretable LLM feature circuits (NeurIPS 2024 · similarity 0.89)
• Faster Diffusion: Rethinking the Role of the Encoder for Diffusion Model Inference (NeurIPS 2024 · similarity 0.89)
• Rare Text Semantics Were Always There in Your Diffusion Transformer (NeurIPS 2025 · similarity 0.89)
• DiCo: Revitalizing ConvNets for Scalable and Efficient Diffusion Modeling (NeurIPS 2025 · similarity 0.88)
• Dual Diffusion for Unified Image Generation and Understanding (CVPR 2025 · similarity 0.88)