Research map / Image generation / Autoregressive and latent generation
Latent Forcing: Reordering the Diffusion Trajectory for Pixel-Space Image Generation
Alan Baade, Eric Ryan Chan, Kyle Sargent, Changan Chen, Justin Johnson, Ehsan Adeli, Li Fei-Fei
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| denoise DINOv2 latents earlier than pixels (alpha shift 16 vs 1/64) (ablation) | Multi-Schedule LF-DiT, DINOv2-B+Reg · ImageNet FID-10K without guidance | FID: 55.35 -> 18.65 | Improved |
| order Data2Vec2 latents before pixels (alpha shift 16 vs 1/64) (ablation) | Multi-Schedule LF-DiT, Data2Vec2 · ImageNet FID-10K without guidance | FID: 55.19 -> 23.61 | Improved |
| order 64x64 downsampled pixels before full-resolution pixels (alpha 16 vs 1/64) (ablation) | Multi-Schedule LF-DiT · ImageNet FID-10K without guidance | FID: 44.51 -> 42.35 | Improved |
| replace JiT+REPA distillation with Latent Forcing DINOv2 ordering | JiT, conditional ImageNet 80 epochs · ImageNet 256 FID-50K unguided | FID: 18.60 -> 9.76 | Improved |
| replace JiT+REPA with Latent Forcing in unconditional generation | JiT, unconditional ImageNet 80 epochs · ImageNet 256 FID-50K guided | FID: 24.40 -> 13.36 | Improved |
| use cascaded schedule instead of variance shift or linear offset (ablation) | Single-Schedule LF-DiT-L · ImageNet FID-10K unguided | FID: 13.48 -> 12.42 (var shift); 20.98 -> 12.42 (linear offset) | Improved |
| add output expert layers (split last 4 layers into latent and pixel experts) (ablation) | LF-DiT-L · ImageNet FID-10K unguided | FID: 12.93 -> 12.42 | Improved |
| sample 10% of pixel timesteps from U[0,0.5] instead of logit-normal (ablation) | LF-DiT-L · ImageNet FID-10K unguided | FID: 12.98 -> 12.42 | Improved |
| add up to 25% noise to latents during pixel timesteps in training (ablation) | LF-DiT, 200 epochs · ImageNet FID-10K unguided | FID: 16.47 -> 10.93 | Improved |
| add up to 50% training noise to latents during pixel timesteps (80 epochs) (ablation) | LF-DiT · ImageNet FID-10K unguided | FID: 13.48 -> 14.64 | Hurt |
| add noise to latents at inference (ablation) | LF-DiT, 200 epochs, trained with 25% noise · ImageNet FID-10K | FID: 10.93 -> 12.55 | Hurt |
| raise probability of sampling latent timestep p_latent from 0.4 to 0.8 (ablation) | LF-DiT cascaded training · ImageNet FID-10K without guidance | FID: 12.42 -> 16.13 | Hurt |
Closest papers by meaning
- Simpler Diffusion: 1.5 FID on ImageNet512 with Pixel-space Diffusion (CVPR 2025 · similarity 0.91)
- PixelDiT2: Representation-Grounded Pixel Diffusion Transformers (NeurIPS 2026 · similarity 0.90)
- OSCAR: One-Step Diffusion Codec Across Multiple Bit-rates (NeurIPS 2025 · similarity 0.89)
- Smooth Diffusion: Crafting Smooth Latent Spaces in Diffusion Models (CVPR 2024 · similarity 0.89)
- Dual-Rate Diffusion: Accelerating diffusion models with an interleaved heavy-light network (NeurIPS 2026 · similarity 0.88)
- Reconstruction vs. Generation: Taming Optimization Dilemma in Latent Diffusion Models (CVPR 2025 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
