AI Paper Atlas

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

ControlNet++: Improving Conditional Controls with Efficient Consistency Feedback

Ming Li, Taojiannan Yang, Huafeng Kuang, Jie Wu, Zhaoning Wang, Xuefeng Xiao, Chen Chen

ECCV 2024 · Image generation › Text-to-image diffusion · approach: Text-driven synthesis and semantic guidance · 97 citations (OpenAlex)

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 changedWhereMeasuredEffect
add pixel-level consistency reward fine-tuning to ControlNetControlNet (SD1.5) · ADE20K (seg mask condition)mIoU: 32.55 -> 43.64Improved
add consistency reward fine-tuningControlNet (SD1.5) · MultiGen-20M depth mapRMSE: 35.90 -> 28.32Improved
add consistency reward fine-tuningControlNet (SD1.5) · MultiGen-20M canny edgeFID: 14.73 -> 18.23Hurt
add consistency reward fine-tuningControlNet (SD1.5) · ADE20K, COCO, hed, lineart, depthFID: 33.28 -> 29.49 (ADE20K)Improved
add consistency reward fine-tuningControlNet (SD1.5) · CLIP score (several conditions)CLIP score: 32.15 -> 31.87 (canny)No clear effect
use a stronger text-to-image backbone (SDXL instead of SD1.5)ControlNet / T2I-Adapter · MultiGen-20M depth mapRMSE: 35.90 (SD1.5) vs 40.00 (SDXL)No clear effect
add real data augmented with ControlNet++ generated images for segmentation trainingDeepLabv3-MBv2 on ADE20K · ADE20KmIoU: +1.76 mIoUImproved
augment real data with ControlNet-generated imagesDeepLabv3-MBv2 on ADE20K · ADE20KmIoU: 34.11 vs 34.08No clear effect
train with consistency loss together with the diffusion loss rather than consistency loss only (ablation)ControlNet++ · ADE20Kimage quality / controllabilityImproved
fine-tune only a subset of timesteps with reward, evaluated by crossing models across timesteps (ablation)ControlNet++ · ADE20KmIoU: 32.55 -> 38.03Improved
use a stronger reward model (UperNet-R50 instead of DeepLabv3-MBv2) (ablation)ControlNet++ · ADE20KmIoU: 31.96 -> 43.64Improved
use the weakest reward model (DeepLabv3-MBv2) (ablation)ControlNet++ · ADE20KmIoU: 32.55 -> 31.96Hurt

Later papers that compared with this method

77 later accepted papers report ControlNet as a baseline in their results tables; 65 of them report a better number on most of the results they share with it.

69 more not listed.

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.

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

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy