Research map / Image generation / Autoregressive and latent generation
Accelerating Controllable Generation via Hybrid-grained Cache
Lin Liu, Huixia Ben, Shuo Wang, Jinda Lu, Junxiang Qiu, Shengeng Tang, Yanbin Hao
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 |
|---|---|---|---|
| apply HGC (block-level and prompt-level cache) instead of no cache | ControlNet · ADE20k segmentation | MACs: 18.22T -> 6.70T | Improved |
| apply HGC instead of no cache | ControlNet · ADE20k segmentation | FID: 38.13 -> 39.44 | Hurt |
| apply HGC to ControlNet++ instead of no cache | ControlNet++ · ADE20k segmentation | FID: 30.51 -> 28.93 | Improved |
| apply HGC instead of DeepCache at similar MACs | ControlNet · COCO-Stuff segmentation | FID: 24.84 -> 20.29 | Improved |
| lower similarity threshold theta from 0.9 to 0 (ablation) | HGC block-level cache · ADE20k segmentation | FID: 28.93 -> 30.92 | Hurt |
| raise threshold theta from 0.9 to 1.0 (ablation) | HGC block-level cache · ADE20k segmentation | FID: 28.93 -> 28.72 | No clear effect |
| increase lambda_intra from 0 to 1.0 (ablation) | HGC generative module · ADE20k segmentation | FID: 28.86 -> 29.04 | Hurt |
| increase lambda_inter from 0.6 to 1.0 (ablation) | HGC generative module · ADE20k segmentation | FID: 28.93 -> 30.49 | Hurt |
| increase lambda_inter from 0 to 0.6 (ablation) | HGC generative module · ADE20k segmentation | FID: 28.44 -> 28.93 | No clear effect |
| cache generative module blocks only (ablation) | ControlNet++ · COCO-Stuff segmentation | MACs: 18.22T -> 12.09T | Improved |
| cache control module blocks only (ablation) | ControlNet++ · COCO-Stuff segmentation | MACs: 18.22T -> 14.95T | Improved |
| apply HGC to video generation with CTRL-Adapter | CTRL-Adapter · DAVIS 2017 | MACs and FID: 48.29T -> 28.88T MACs; FID 11.03 -> 11.70 | Mixed |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- DeepCache — DeepCache: Accelerating Diffusion Models for Free (CVPR 2024) · better on 4 of 6 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
- CCM: Real-Time Controllable Visual Content Creation Using Text-to-Image Consistency Models (ICML 2024 · similarity 0.85)
- CtrLoRA: An Extensible and Efficient Framework for Controllable Image Generation (ICLR 2025 · similarity 0.85)
- FlexControl: Computation-Aware Conditional Control with Differentiable Router for Text-to-Image Generation (ICML 2025 · similarity 0.85)
- Model Tells You What to Discard: Adaptive KV Cache Compression for LLMs (ICLR 2024 · similarity 0.84)
- GEN3C: 3D-Informed World-Consistent Video Generation with Precise Camera Control (CVPR 2025 · similarity 0.84)
- OmniGen2: Towards Instruction-Aligned Multimodal Generation (CVPR 2026 · similarity 0.84)
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.
