ECCV 2026 · Project Page

Reconstruction-Anchored Diffusion Model for Text-to-Motion Generation

Yifei Liu^1,2· Changxing Ding¹· Ling Guo¹· Huaiguang Jiang¹· Qiong Cao²

¹ South China University of Technology · ² Joy Future Academy

A motion-centric latent space and reconstructive error guidance for realistic, semantically aligned human motion generation.

Paper ↗ arXiv ↗ Supplementary ↗ Code · coming soon

0.032FID on HumanML3D

56.1%R-Precision @1

20denoising steps

RAM uses motion and text encoders to guide a diffusion model toward a final motion output — RAM anchors text-to-motion generation to a motion-centric latent manifold.

Abstract

Text-to-motion diffusion models must bridge an abstract language representation with continuous, high-dimensional motion, while errors introduced early in denoising can propagate through the rest of the sampling process. RAM addresses both issues. It learns a motion-centric latent space through a reconstruction branch, uses self-regularization and motion-centric latent alignment to map text into that space, and introduces Reconstructive Error Guidance (REG) to expose and amplify improvements during inference.

Video Presentation

Text descriptions become coherent motion

The supplementary video walks through the motivation, method, comparisons, and results.

Method

Two anchors stabilize text-to-motion diffusion

RAM uses motion reconstruction during training and reconstructive guidance during inference.

RAM training and inference framework with motion-centric latent alignment and REG — Training learns a motion-centric latent space; inference uses its reconstruction branch as a negative reference for guidance.

Reconstruct motion

A motion encoder summarizes an input sequence into a latent representation that preserves the dynamics needed by the diffusion decoder.

Align text to motion

Self-regularization makes the motion manifold discriminative, while motion-centric latent alignment maps text conditions onto that manifold.

Correct errors early

REG reconstructs the previous estimate, then amplifies the residual between that reconstruction and the current text-driven prediction.

Motion-centric latent space

Text follows the structure of motion

RAM keeps the motion space as the anchor and pulls text latents toward it. On paired HumanML3D test examples, the resulting representation forms tighter interleaving between text and motion latents than the two-stream baseline and a contrastive variant.

0.579paired cosine similarity for RAM in the joint t-SNE comparison

Joint t-SNE comparison of paired text and motion latents for three variants — Paired text and motion latents become more tightly interleaved with RAM.

previous
estimate

reconstruction

text-driven
prediction

final
prediction

The reconstruction branch reproduces prior error patterns so the current prediction can move away from them.

Reconstructive Error Guidance

Use the model’s self-correction ability

Early denoising steps recover motion from nearly pure noise and are especially prone to error. At each step, REG encodes the previous estimate, reconstructs it, and compares that reconstruction with the current text-conditioned prediction. Amplifying this residual highlights the improvement made by the current step.

final prediction = text-driven prediction + w₁ · REG residual + w₂ · CFG residual

Benchmarks

Strong alignment with fewer denoising steps

Results reported on the standard HumanML3D and KIT-ML test sets.

14,616HumanML3D motions

44,970HumanML3D descriptions

3,911KIT-ML motions

20RAM inference steps

Selected text-to-motion results. Lower FID and higher R-Precision are better.
Method	HumanML3D		KIT-ML
	FID ↓	R@1 ↑	FID ↓	R@1 ↑
MDM	0.489	0.418	0.547	0.404
MotionLCM v2	0.056	0.553	—	—
Salad	0.076	0.581	0.296	0.477
RAM (ours)	0.032	0.561	0.172	0.464

On HumanML3D, RAM reaches an FID of 0.032 and an R-Precision@1 of 0.561 with 20 inference steps. On KIT-ML, it remains competitive on both motion realism and semantic alignment.

Qualitative comparison

Long action sequences stay on track

RAM better preserves the action sequence and fine-grained motion cues in challenging prompts.

Qualitative comparison of MDM, MoMask, Salad, and RAM across four text prompts — Examples on multi-action prompts from the HumanML3D test set.

Efficiency and ablations

REG improves realism where it matters most

Guidance can be applied selectively to trade a small amount of time for a large FID improvement.

0.398 sRAM AITS with full REG20 steps, compared with 0.490 s for MDM at 50 steps

0.038FID with REG on early stepsEarly guidance captures most of the realism gain

0.032FID with full REGFull guidance gives the best reported HumanML3D result

RAM

Removing motion-centric latent alignment raises FID from 0.032 to 0.424. Removing self-regularization raises it to 0.064, while removing REG raises it to 0.132. The three components contribute complementary gains.

User study

Viewers preferred the most coherent motion

Participants ranked samples by considering motion quality and semantic alignment together.

User study ranking distribution for RAM, Salad, and MoMask — RAM receives the largest first-place share in the perceptual comparison.

The study compares RAM with MoMask and Salad on HumanML3D prompts. Participants jointly considered whether the motion looked natural and whether it matched the text.

Read the supplementary material ↗

Additional material

Method details and extended presentation

Supplementary Material ↗

Supplementary presentation

Slides covering the motivation, method, latent space, and ablations.

Motion reconstruction

The reconstruction branch supplies the motion-centric anchor used to supervise text-to-motion generation.

Reconstructive guidance

REG uses the previous estimate as a negative reference and amplifies the improvement in the current prediction.

Limitations

One text token summarizes the description

RAM currently represents the entire textual description with a single token. Finer-grained multi-token conditions could support more detailed control for long-horizon motion generation.

Acknowledgement

This work was supported by the National Natural Science Foundation of China under Grants 62476099 and 62076101, Guangdong Basic and Applied Basic Research Foundation under Grants 2024B1515020082 and 2023A1515010007, and the TCL Young Scholars Program.

BibTeX

Citation

@inproceedings{liu2026ram,
  title     = {Reconstruction-Anchored Diffusion Model for Text-to-Motion Generation},
  author    = {Liu, Yifei and Ding, Changxing and Guo, Ling and Jiang, Huaiguang and Cao, Qiong},
  booktitle = {European Conference on Computer Vision (ECCV)},
  year      = {2026}
}