self-conditioning
Imagine you're trying to draw a perfect circle by starting with a rough scribble. Self-conditioning means that at each step, the model looks at its own previous guess and uses that information to refine the next guess. In diffusion models, this lets the system gradually improve its predictions during generation, using its own outputs as a guide rather than needing extra components.
The papers show that self-conditioning can be the backbone of high-performing generative models, not just an add-on. It provides coherent spatial guidance in 3D objects, preserves uncertainty in discrete text generation, and even works across very different data types, images, speech, and proteins, with a single training approach. Across all cases, self-conditioning turns a simple denoising process into a powerful, versatile generation engine.
- Papers
- 5
- Released code
- 0
- First seen
- Oct 2026
- Latest
- Nov 2026
5 papers in the last two months, against 0 in the two before.
Start here
Three papers to read first, in the order they came out- 1Position-based self-conditioning improves single-stage 3D generation without extra stages
Position Forcing improves single-stage 3D generation by using recovered token positions as progressive spatial guidance during denoising.
Oct 2026 - 2A language model using only self-conditioning outperforms existing diffusion models
A continuous generative model can be built purely on self-conditioning by propagating the expected embedding across denoising steps and decoding once at the end.
Oct 2026 - 3Image generators can improve their own samples without external judges
A training recipe, Scout-Refine-Correct, lets flow-matching image generators refine their own samples during inference, matching an oracle-based best-of-N baseline on ImageNet without a verifier.
Nov 2026
How fast it is moving
Often studied with
diffusion transformerA neural network architecture combining transformer layers with diffusion processes for high-quality image and video generation.2 togetherdiffusion language modelA language model that generates text by iteratively denoising a latent or discrete representation.2 togethercoarse-to-fine generationGeneration process that first produces a rough structure and then refines details in subsequent stages.1 togetherThe papers
Most central to this idea first, not most recent.
- Big Techcs.LG
A language model using only self-conditioning outperforms existing diffusion models
A continuous generative model can be built purely on self-conditioning by propagating the expected embedding across denoising steps and decoding once at the end.
University of Basel, Microsoft AI · Oct 2026
- Industrycs.CV
Image generators can improve their own samples without external judges
A training recipe, Scout-Refine-Correct, lets flow-matching image generators refine their own samples during inference, matching an oracle-based best-of-N baseline on ImageNet without a verifier.
FusionBrain, AXXX, HSE Unifersity · Nov 2026
- Independentcs.CV
Position-based self-conditioning improves single-stage 3D generation without extra stages
Position Forcing improves single-stage 3D generation by using recovered token positions as progressive spatial guidance during denoising.
Oct 2026
- Big Techcs.LG
Keeping token uncertainty improves discrete diffusion generation and fast sampling
Simplex Diffusion Models represent beliefs over categories on the probability simplex, mitigating information collapse and achieving competitive performance on text and code generation.
Google DeepMind, EPFL · Oct 2026
- Big Techcs.CL
One Bitwise Diffusion Framework Generates Across Three Multimodal Domains
CoBit, a single continuous bitstream diffusion protocol, handles joint and conditional generation across image-text, speech-text, and protein sequence-structure pairs.
Devotion AI Labs, University of Cambridge · Oct 2026
Concepts are extracted from each paper and reused across the corpus, so this page grows on its own as the desk reads.