Research map / Image generation / Fast diffusion sampling
SD3.5-Flash: Distribution-Guided Distillation of Generative Flows
Hmrishav Bandyopadhyay, Rahim Entezari, Jim Scott, Reshinth Adithyan, Yi-Zhe Song, Varun Jampani
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 |
|---|---|---|---|
| share DMD timesteps with student trajectory timesteps (timestep sharing) (ablation) | SD3.5-Flash 4-step · ablation (qualitative) | image quality | Improved |
| add adversarial objective (ablation) | SD3.5-Flash 4-step · ablation (qualitative) | image quality | Improved |
| add trajectory-guidance pre-training of the student (ablation) | SD3.5-Flash 4-step · ablation (qualitative) | colour and composition | Improved |
| add periodic discriminator head refresh (ablation) | SD3.5-Flash 4-step · ablation (qualitative) | composition and smoothness | Improved |
| drop T5-XXL text encoder (null embeddings) (ablation) | SD3.5-Flash 16-bit 4-step · COCO-30K | GenEval: 0.70 -> 0.68 | Hurt |
| drop T5-XXL, 16-bit (ablation) | SD3.5-Flash 4-step · peak VRAM | peak VRAM (GiB): 17.58 -> 8.71 | Improved |
| quantize MM-DiT 16-bit to 8-bit (no T5) (ablation) | SD3.5-Flash 4-step · COCO-30K | GenEval: 0.68 -> 0.68 | No clear effect |
| distill SD3.5M to 4 steps vs 50-step teacher | SD3.5-Flash · COCO-30K | GenEval: 0.64 -> 0.70 | Improved |
| distill SD3.5M to 4 steps vs 50-step teacher | SD3.5-Flash · COCO-30K | FID: 20.06 -> 29.80 | Hurt |
| split-timestep fine-tuning with 3:7 branch merge | SD3.5-Flash 4-step · GenEval prompt alignment | GenEval | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 0 of 1 on most shared results.
- SANA-Sprint 0.6B — SANA-Sprint: One-Step Diffusion with Continuous-Time Consistency Distillation (ICCV 2025) · better on 0 of 4 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Align Your Flow: Scaling Continuous-Time Flow Map Distillation (NeurIPS 2025 · similarity 0.87)
- Self-Corrected Flow Distillation for Consistent One-Step and Few-Step Image Generation (AAAI 2025 · similarity 0.87)
- Flash Diffusion: Accelerating Any Conditional Diffusion Model for Few Steps Image Generation (AAAI 2025 · similarity 0.86)
- Hyper-SD: Trajectory Segmented Consistency Model for Efficient Image Synthesis (NeurIPS 2024 · similarity 0.86)
- DreamFlow: High-quality text-to-3D generation by Approximating Probability Flow (ICLR 2024 · similarity 0.85)
- Simple and Fast Distillation of Diffusion Models (NeurIPS 2024 · similarity 0.85)
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.
