ECCV 2026 3D Vision Human Understanding hand–object interaction generation dynamic contact map diffusion model text-driven motion generation physical plausibility

JointHOI: Jointly Generating Contact Maps Enhances Hand Object Interaction Generation¶

Conference: ECCV 2026
Paper: ECCV 2026 / Project Page
Area: 3D Vision / Human Understanding
Keywords: hand–object interaction generation, dynamic contact map, diffusion model, text-driven motion generation, physical plausibility

TL;DR¶

JointHOI treats contact as an inner modality of hand–object motion: a single-stage diffusion model denoises bimanual MANO poses, the object trajectory, and per-hand frame-wise contact distance maps in one shared sequence, then uses the agreement between the self-generated contact maps and the contact implied by the synthesized geometry to guide sampling — improving both action semantics (ARCTIC Top-1 0.948) and physical plausibility (interpenetration volume less than half that of Text2HOI).

Background & Motivation¶

Hand–object interaction (HOI) is defined by contact: where the hand touches the object, when contact forms or breaks, and how the contact region migrates as the motion unfolds. This is what separates HOI generation from generic human motion synthesis — it is not enough for the motion to look right, the contact must be spatio-temporally self-consistent, because humans are extremely sensitive to artifacts such as interpenetration, floating fingers, and unstable grasps (a few millimetres of contact error is enough). Existing text-driven methods (Text2HOI, DiffH2O, LatentHOI, HOI-GPT) all try to inject interaction modelling, but each pays a price. Text2HOI explicitly predicts contact maps and uses them as the conditioning signal for motion generation, yet its contact representation is static and binary: it discards the frame-by-frame evolution of contact and coarsens proximity into a touched/not-touched label, and it still needs a post-hoc refinement module afterwards. DiffH2O factorizes generation into two stages (grasp planning, then motion), while LatentHOI and HOI-GPT hide the interaction in a latent space or an autoregressive sequence — effective, but leaving no explicit, inspectable contact signal that can be enforced at inference time, and multi-stage pipelines propagate stage-wise errors. In the end, these designs either drop the temporal dimension or bury the geometry in a latent code, collapsing the spatio-temporal structure of contact.

The core tension is that contact is inherently a spatio-temporal signal — it varies across the object surface and evolves continuously over time (approach, slide, roll, release) — whereas the representations used on the generative side are typically static, binary, or implicit. The consequence of this mismatch is exactly the failure mode we observe: every frame looks acceptable in isolation, yet the full sequence plays back with floating and interpenetration. The goal of this paper is therefore to make contact dynamic, dense, and explicit during generation, produced jointly with the motion rather than patched in afterwards as a constraint or guessed from a latent code.

The angle of attack comes from a recent observation in generative modelling: treating an inner modality — a signal directly derivable from the primary output — as an additional generation target can in turn improve the primary output (VideoJAM co-generates optical flow for temporal coherence; Redi co-generates DINOv2 features for image synthesis). In HOI, the signal that best fits "derivable from the primary output and central to the task" is the contact map. Core idea: treat the contact map as an inner modality of hand–object motion, use a single-stage diffusion model to jointly denoise motion and a frame-wise contact distance field within one sequence so that motion and contact dimensions condition each other through self-attention, and at inference steer sampling with the consistency gradient between the explicitly generated contact and the contact implied by the motion — reducing penetration and floating without any post-processing or multi-stage training.

Method¶

Overall Architecture¶

The input is a text prompt plus the object's canonical point cloud (and the target sequence length \(L_{\max}\)); the output is a bimanual hand–object interaction sequence — per-frame MANO parameters for both hands, the object pose (plus articulation for articulated objects), and per-frame dynamic contact maps for the left and right hands. The pipeline contains a single generative model and a single diffusion process. At training time, ground-truth motion is first parsed into frame-wise contact maps that serve as supervision targets; motion states and contact maps are concatenated into per-frame tokens and fed to a transformer diffusion model trained with clean-sample prediction. At inference time, the whole token sequence starts from pure noise and is denoised jointly; each reverse step applies one contact-consistency guidance (CIG) update to the current noisy variable, and the final clean estimate is sliced dimension-wise into motion and contact. The three stages below correspond to the three key designs: how contact is represented, how motion and contact are generated jointly, and how the generated contact is used to correct geometry during sampling.

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
    A["Text prompt + object point cloud"] --> B["Dynamic contact map<br/>anchor → hand-surface min distance"]
    B -->|supervision target at training| C["Per-frame joint sequence<br/>motion token + L/R contact token"]
    C --> D["Single-stage joint diffusion<br/>one transformer denoises both"]
    D --> E["Contact Inner Guidance<br/>contact-consistency gradient"]
    E --> F["Bimanual/object motion + dynamic contact maps"]

Key Designs¶

1. Dynamic contact map: writing "contact" as a frame-wise continuous surface distance field

The pain point is specific: a binary contact map is unstable around the threshold and cannot express "the hand is approaching but has not touched yet" — precisely the intermediate state that matters most for physical plausibility — while a static representation discards how the contact region migrates over time. JointHOI builds its contact representation on anchors on the object surface: for each object, farthest point sampling (FPS) selects a fixed set of \(N_c=1024\) anchors that uniformly covers the surface in the canonical frame. Anchors live in the object's own coordinate frame, so they are decoupled from global motion: when the object translates and rotates, the anchors move with it, and contact can be tracked consistently. This is also why this parameterization is better suited to guidance than recomputing contact in the world frame each frame — anchor identities are stable, so frame-to-frame comparison is meaningful in the first place.

Per frame, contact is computed separately for each hand: the anchors are transformed into the world frame with the current object pose, giving \(\{a_{l,n}\}\), and each anchor takes the minimum Euclidean distance to all mesh vertices of that hand

\[\hat{C}^{H}_{l}[n]=\min_{u\in\mathcal{U}^{H}_{l}}\lVert u-a_{l,n}\rVert_{2},\qquad n=1,\dots,N_c,\ H\in\{L,R\}\]

where \(\mathcal{U}^H_l\) is the set of MANO mesh vertices for that hand at that frame (778 vertices). Using the minimum over vertices instead of an exact surface distance is a differentiable and efficient approximation; using a continuous distance instead of a binary label removes threshold sensitivity and keeps the signal smooth near contact boundaries, which is exactly the prerequisite for gradient-based guidance later. The ablation confirms that "separate maps for the two hands" is not a detail: merging both hands into one unified contact representation degrades FID on ARCTIC from 0.038 to 0.145, indicating that bimanual contact dynamics are asymmetric and loosely coupled, and that forcing a shared representation mixes the two hands together.

2. Single-stage joint diffusion: motion and contact condition each other within one sequence

This is the heart of the paper. Each frame's motion state and contact maps are concatenated into a single token:

\[y_l=\big[x^{L}_{l},\,x^{R}_{l},\,x^{O}_{l},\,C^{L}_{l},\,C^{R}_{l}\big]\in\mathbb{R}^{d_y},\qquad d_y=99+99+10+2N_c\]

Here a single hand \(x^H_l\in\mathbb{R}^{99}\) is global translation plus 6D rotations for 16 joints (MANO parameters), and the object \(x^O_l\in\mathbb{R}^{10}\) is translation, 6D rotation, and one articulation scalar (for rigid-object data such as GRAB the articulation term is dropped, giving 9 dimensions); with \(N_c=1024\), \(d_y\) is 2256. The sequence is stacked over time and the diffusion process is defined directly on \(y\); the denoiser is a transformer that takes the noisy sequence \(y_t\) and predicts the corresponding clean sequence \(\hat y_0\), trained by clean-sample regression:

\[\mathcal{L}_{\mathrm{joint}}=\mathbb{E}_{y_0,t,\epsilon}\Big[\big\lVert y_0-D_\theta\big(y_t,o,e,\tau(t)\big)\big\rVert_2^2\Big]\]

Conditioning is injected through prefix conditioning: a text token (CLIP embedding \(e\)), an object token (global/local PointNet features plus scale and centroid, \(o\in\mathbb{R}^{1088}\)), and a timestep token (\(\tau(t)\)) are prepended to the sequence, after which standard full self-attention is applied over \([z;\phi(y_t)]\), with no separate cross-attention module.

The crucial question is where "joint" actually lives. Because the motion dimensions and the contact dimensions sit in the same token and the same attention operation, \(\mathcal{L}_{\mathrm{joint}}\) forces the model to learn cross-modal dependencies: motion channels can inform contact channels, and contact channels can inform motion channels in return. This coupling is bidirectional, not the serial dependency of "generate motion, then compute contact" or "generate contact, then use it as a condition". Physically grounded co-variation patterns — distances shrinking smoothly during approach, growing during release, contact regions changing in coordination during sliding or rolling — are therefore internalized into the model weights as a prior. By contrast, Text2HOI feeds a static binary contact map as an input condition and then applies post-hoc refinement, so errors propagate across stages; LatentHOI / HOI-GPT hide the interaction in a latent code or autoregressive token, making contact neither inspectable nor enforceable during sampling; contact-constraint methods such as HOIDINI / CODA require hundreds of noise-optimization steps at inference. JointHOI produces motion and contact together with a single sequence, a single model, and a single diffusion process.

3. Contact Inner Guidance: using the self-generated contact map to correct geometry during sampling

Joint generation does not guarantee that the sampled motion actually honours the contact the model itself predicted — over long, highly dynamic interactions, iterative denoising still accumulates geometric inconsistency. The idea behind CIG is that the model already provides two versions of contact: the contact map it explicitly generates, \(\hat C^H\), and the contact map analytically derived from the hand–object geometry it generates, \(\bar C^H\) (anchors transformed by the predicted object trajectory, hand vertices from \(\mathrm{MANO}(\hat x^H)\), then the same minimum distance). The two should agree; where they disagree is where the artifacts are.

At each reverse step, the current clean estimate \(\hat y_0\) is parsed into motion and contact, the motion-implied \(\bar C^H\) is computed, and a log-scale energy measures their discrepancy. The gradient of that energy with respect to the current noisy variable is then applied:

\[\mathcal{L}_{\mathrm{cig}}=\sum_{H\in\{L,R\}}\Big\lvert\log\big(\hat C^{H}+\varepsilon\big)-\log\big(\bar C^{H}+\varepsilon\big)\Big\rvert,\qquad y_t\leftarrow y_t-w\,\nabla_{y_t}\mathcal{L}_{\mathrm{cig}}\]

The log scale is deliberate: it concentrates weight on the near-contact, small-distance regions, which are precisely the regions that determine how real an interaction looks, whereas far-field distances carry little information; the constant \(\varepsilon\) keeps \(\log\) from blowing up as distances approach zero. The gradient has to flow back to \(y_t\) through two differentiable operators, MANO and the minimum-distance computation, so the guidance acts on the entire joint sequence (the object trajectory included); after the update, the standard diffusion reverse transition is applied. In family it is classifier-based guidance, but the "classifier" here is a closed-form energy rather than an extra network, so it requires neither additional training, nor a grasp-planning stage, nor post-processing — the single-stage property is fully preserved, at the cost of one extra backward pass per sampling step (see the efficiency table below). One detail is worth flagging: the main text calls this a log-ratio energy while the guidance ablation names the best variant Log-\(\ell_1\); the formula is corrupted in the cached version, so it is written above as the absolute log-difference, ⚠️ refer to the original paper for the exact form. The guidance weight is \(w=0.05\).

Loss & Training¶

The training objective is \(\mathcal{L}_{\mathrm{joint}}\) above, using clean-sample prediction (regressing the clean sequence) rather than noise prediction, with a linear noise schedule, \(T=1000\) diffusion steps for training and 100 sampling steps at inference. The whole model is trained on a single RTX A6000 (48GB): Adam with learning rate \(10^{-4}\), batch size 128 on GRAB and 64 on ARCTIC, up to 380K and 60K iterations respectively, with early stopping once the training loss plateaus. CIG takes no part in training and only acts at inference with \(w=0.05\); the weight sweep shows it is stable over a wide range (accuracy essentially unchanged, penetration depth consistently reduced), so this hyper-parameter needs no fine tuning. Generated sequences use the same number of frames as the corresponding ground-truth clip so that all comparisons follow one protocol.

Key Experimental Results¶

Main Results¶

Evaluation uses GRAB (51 rigid objects, 29 action categories, 30 fps, up to 196 frames) and ARCTIC (11 articulated objects, 10 action categories, 30 fps, up to 64 frames), with the text prompts released by Text2HOI. Neither dataset provides an official split in the text-to-HOI setting, so the authors build one under a shared protocol: a balanced test set is formed by uniformly sampling object–action pairs, and the remaining sequences are used for training; all methods are trained and evaluated on the same splits. Evaluation covers two sides. On the semantic side, an RNN action classifier trained per the IMOS protocol provides motion embeddings, and Acc (Top-1/Top-3) and FID are reported. On the physical side, the interpenetration volume IV (cm³) and maximum penetration depth ID (cm) between hand and object meshes are reported together with the Contact Ratio CR (the fraction of frames, among those where the object translates or articulates, that maintain valid contact below 5 mm).

Dataset	Method	Acc Top-1 ↑	Acc Top-3 ↑	FID ↓	IV (cm³) ↓	ID (cm) ↓	CR (%)
ARCTIC	GT	0.966	0.992	–	4.419	0.288	95.56
ARCTIC	MDM	0.739	0.839	0.365	9.542	0.581	61.37
ARCTIC	DiffH2O*	0.636	0.862	0.424	6.491	0.475	94.28
ARCTIC	DiffH2O	0.560	0.823	0.547	6.412	0.527	94.91
ARCTIC	LatentHOI	0.549	0.846	0.301	8.597	0.496	78.81
ARCTIC	Text2HOI*	0.815	0.923	0.152	8.930	0.587	93.28
ARCTIC	Text2HOI	0.816	0.927	0.148	8.281	0.524	97.33
ARCTIC	JointHOI	0.948	0.983	0.033	4.406	0.426	93.71
GRAB	GT	0.779	0.895	–	2.790	0.526	95.02
GRAB	MDM	0.395	0.500	1.203	4.915	0.650	81.22
GRAB	DiffH2O*	0.595	0.737	0.410	3.282	0.610	87.24
GRAB	DiffH2O	0.458	0.526	0.533	3.607	0.846	87.82
GRAB	LatentHOI	0.584	0.721	0.214	4.356	0.570	88.39
GRAB	Text2HOI*	0.716	0.805	0.118	8.114	0.810	97.85
GRAB	Text2HOI	0.711	0.831	0.116	7.790	0.778	98.24
GRAB	JointHOI	0.663	0.847	0.031	3.419	0.525	93.00

Inference cost for a 196-frame sequence on a single A6000:

Metric	MDM	DiffH2O	LatentHOI	Text2HOI	JointHOI
Time (s) ↓	10.61	124.44	15.95	16.79	10.98
FPS ↑	18.47	1.57	12.29	11.67	17.85

On ARCTIC, JointHOI attains the highest semantic fidelity among synthesized methods (Top-1 0.948, Top-3 0.983, close to the 0.966/0.992 of ground-truth motion) and the lowest FID (0.033 versus 0.148 for the runner-up Text2HOI), while cutting interpenetration to 4.406 cm³ — slightly below the 4.419 cm³ of the ground-truth motion. Note, however, that its penetration depth of 0.426 cm is still above the ground truth's 0.288 cm, so this should not be read as "more physical than the ground truth"; the penetration area is simply very small. It also outperforms the oracle-assisted DiffH2O (whose grasp stage is conditioned on ground-truth grasp signals) and the refinement-free Text2HOI. On GRAB, JointHOI obtains the best Top-3 (0.847) and the lowest FID (0.031, versus 0.116 for Text2HOI), but its Top-1 (0.663) is below Text2HOI's 0.711 — the authors argue that the clear Top-3 advantage indicates more reliable handling of ambiguous actions (e.g. use versus inspect), and that its interpenetration volume is a little over half of Text2HOI's (3.419 versus 7.790), with a penetration depth of 0.525 cm essentially matching the ground truth's 0.526 cm. The CR column must not be read as "higher is better": Text2HOI's CR (97.33 on ARCTIC, 98.24 on GRAB) exceeds the ground truth's (95.56 / 95.02) precisely because interpenetration inflates the contact ratio, so the sensible reading is "CR close to GT while IV/ID stay low".

Ablation Study¶

An incremental ablation is run on ARCTIC, where each row adds one component on top of the previous setting:

Config	Acc Top-1 ↑	Acc Top-3 ↑	FID ↓	IV (cm³) ↓	ID (cm) ↓	CR (%)
Baseline (two-stage)	0.520	0.738	0.291	8.096	0.528	73.40
+ One-stage (joint)	0.795	0.919	0.274	7.057	0.548	81.72
+ Dynamic contact map (unified)	0.894	0.948	0.145	7.146	0.432	85.28
+ Dynamic contact map (L/R)	0.935	0.981	0.038	6.988	0.439	92.21
+ Contact Inner Guidance	0.948	0.983	0.033	4.406	0.426	93.71

The CIG loss design and guidance weight are given in Fig. 4 of the original paper (no numeric table for that figure is available in the cache, so only the conclusions are reported here): five consistency forms are compared — None, \(\ell_1\), \(\ell_1\)+mask, cosine, and Log-\(\ell_1\) — and Log-\(\ell_1\), which emphasizes relative errors near contact, gives the best Acc–ID trade-off, with proximity-aware variants such as \(\ell_1\)+mask also performing strongly. The weight sweep shows a broad stable range on both datasets where Acc is essentially unchanged while ID drops consistently, so \(w=0.05\) is used throughout.

Key Findings¶

Single-stage joint generation is the single largest gain: switching from the two-stage baseline to one-stage joint training lifts ARCTIC Top-1 from 0.520 straight to 0.795, the biggest jump in the ablation. This directly supports the motivation that stage-wise pipelines accumulate errors — merging grasp/contact and motion synthesis into one generative process is far more efficient than generating the interaction first and the motion afterwards.
Moving the contact representation from "static/unified" to "dynamic/per-hand" contributes the largest realism gain: introducing the distance-based dynamic contact map drops FID from 0.274 to 0.145, and separating the two hands drops it further to 0.038 — more than doubling the improvement. The chain shows that neither "contact must be a time-varying signal" nor "bimanual contact is heterogeneous" is a nice-to-have.
Interpenetration is mainly suppressed by CIG, and not at the cost of semantics: the dynamic-contact-map step even raises IV slightly (7.057 → 7.146) while substantially improving Acc and FID; the real IV drop (6.988 → 4.406) happens once CIG is added, and Top-1 simultaneously edges up from 0.935 to 0.948. This is evidence that CIG is a training-free inference-time correction rather than yet another semantic model.
The log-scale guidance energy receives direct empirical support: Log-\(\ell_1\), which penalizes relative error only in near-contact regions, is optimal, matching the design intuition that near-contact regions determine realism; CIG's insensitivity to the weight further shows that its benefit is not bought by fine-tuning.
Efficiency pays almost none of the multi-stage penalty: 10.98 s / 17.85 FPS for a 196-frame sequence, on par with the single-stage MDM (10.61 s) and far faster than Text2HOI (16.79 s), which also models contact explicitly, and two-stage DiffH2O (124.44 s). CIG's overhead stays modest because it evaluates the energy directly on the anchor-based contact representation, with no extra planning or optimization problem to solve.
Where qualitative results benefit: bimanual coordination ("inspect the cube with both hands", where baselines often engage only one hand), thin structures ("wear an eyeglasses with both hands", where baselines struggle to align with the frame), and placement actions requiring stable control ("place the scissor with right hand", where baselines show floating or unstable grasps).
One honest caveat: on GRAB, Top-1 loses to Text2HOI. Although it leads substantially on physical metrics and has a higher Top-3, the claim "semantic alignment is best across the board" does not hold.

Highlights & Insights¶

Upgrading contact from a condition to an inner modality is the paper's cleanest move: a static contact map can only serve as an input condition, whereas generating motion and contact together turns "the contact the model itself predicts" into an object that can be self-checked at inference time — and only then does CIG become possible. Contact goes from "a constraint given before generation" to "part of the generation process that can in turn constrain it". The shift is independent of the specific network and is well worth transferring.
Guidance without an extra network: CIG belongs to classifier-based guidance, but the "classifier" is a closed-form energy (the log-difference between the explicit and the analytically derived contact), so it trains nothing, adds no module, and needs no post-processing. Wherever a task has such a redundant pair — an explicitly predicted quantity and the same quantity parsed from the predicted geometry — this self-consistency guidance carries over (e.g. velocity fields in physics simulation, multi-view geometry in depth estimation).
The FID jump from per-hand contact maps is an informative finding: a unified representation forces both hands to share one contact pattern, whereas in real bimanual interaction the contact dynamics of the dominant and assistive hands differ substantially; modelling them separately spares the model the burden of first disentangling and then re-coupling.
Anchor-based contact turns "contact" into a fixed-size table: the 1024 object-surface anchors serve both as supervision targets and as the compared quantity during guidance, with identities that are stable across frames and decoupled from object pose. Compared with recomputing contact in the world frame every frame, this parameterization makes temporal comparison and gradient back-propagation far cleaner.
Where to go one step further from "contact": the inner modality need not be a contact map — contact forces or contact normals capture manipulation semantics even better — and the CIG consistency loss could be replaced by a learned contact discriminator, upgrading physical plausibility from a hand-written log-difference into a learnable prior.

Limitations & Future Work¶

Validated on only two short-clip datasets: GRAB is at most 196 frames and ARCTIC only 64, both desktop-level interactions between one person's two hands and a single object. Longer task-level sequences (pick up, use, put down) and multi-object scenes are not covered, so whether joint denoising can still maintain contact consistency over long sequences remains unanswered.
The data split is not official: no official split exists in the text-to-HOI setting; the self-built balanced test set is consistent across all methods but its numbers are not directly comparable with those reported in other papers.
Contact-map geometry is limited by the 1024 anchors and the "minimum vertex distance" approximation: for thin structures such as eyeglasses or scissors, the anchor density may be insufficient to resolve the critical contact regions; approximating surface distance by the minimum over mesh vertices is differentiable and efficient but deviates in thin-walled or concave parts.
Limited object-side expressiveness: articulated objects are described by a single scalar joint angle, which does not cover multi-DoF or deformable objects; fixing anchors in the canonical frame also means the representation breaks down once the object deforms (cloth, soft bodies).
Physical metrics are coupled: CR is inflated by interpenetration (Text2HOI's CR exceeding the ground truth is the proof), so "physical plausibility" effectively requires manual judgement over the IV/ID/CR trade-off, with no single game-proof physical metric available.
Semantics do not lead across the board on GRAB: Top-1 is below Text2HOI, suggesting that conditioning directly on contact maps still has an edge on datasets with little semantic ambiguity; this paper's relative advantage concentrates in motion realism and physical plausibility.
Directions worth trying: make anchors adaptive in density or learnable query points so thin structures get more capacity; extend the CIG consistency constraint from contact distance to contact normals / local surface orientation to suppress wrong orientations during sliding; replace the single articulation scalar with a multi-dimensional joint vector to generalize to multi-finger, multi-joint objects.

vs Text2HOI: they also use contact maps explicitly, but the maps are static and binary and act as a precondition for motion generation, followed by a post-hoc refinement module; here the maps are frame-wise continuous distance fields generated jointly with motion inside one diffusion process, with no post-processing. The cost is a slightly lower Top-1 on GRAB (0.663 vs 0.711), while physical plausibility (IV 3.419 vs 7.790) and FID (0.031 vs 0.116) lead by a wide margin. This also shows that contact-as-condition and contact-as-joint-target each have their own regime: the former stays "on topic" semantically, the latter is stronger on physical self-consistency.
vs DiffH2O: two-stage (grasp first, motion second), whereas this paper merges grasp and motion synthesis into one diffusion process; even when the grasp stage is fed ground-truth signals (DiffH2O*), ARCTIC Top-1 is only 0.636 versus 0.948 here, and DiffH2O's inference time (124.44 s) is more than ten times longer.
vs LatentHOI / HOI-GPT: both place the interaction in a latent space or autoregressive tokens, which removes the need to design a contact representation but leaves the contact signal uninspectable and unenforceable during sampling; the explicit anchor distance field buys both interpretability and guidability here, at the price of introducing anchor density as a new representation bottleneck.
vs HOIDINI / CODA (inference-time contact-constraint optimization): those methods satisfy contact constraints through noise optimization (e.g. DNO) at inference, often requiring hundreds of steps; this paper generates contact directly so guidance costs only one extra backward pass on sampling steps that would be taken anyway (10.98 s vs Text2HOI's 16.79 s) — a fundamentally different cost structure.
vs BimArt: BimArt also predicts HOI contact maps but assumes the object trajectory is given; this paper targets text-to-HOI where the object motion must be synthesized too, so contact, object, and hands constrain one another.
Methodological takeaway: treating "a quantity derivable from the primary output" as an additional generation target (optical flow in VideoJAM, features in Redi, contact maps here) is a low-cost, direct enhancement; more importantly, such inner modalities naturally provide a set of self-consistency checks at inference that translate directly into training-free guidance. Wherever a task has redundant intermediate quantities, this path can be copied.

Rating¶

Novelty: ⭐⭐⭐⭐ "Contact as an inner modality plus inference-time guidance from self-generated contact" is a clean combination; neither joint generation nor self-consistency guidance is a first, but applied to HOI generation it addresses a genuine pain point.
Experimental Thoroughness: ⭐⭐⭐⭐ Two datasets, six metrics, four baselines including oracle-assisted and refinement-free variants, a component-wise incremental ablation plus guidance-loss and weight analyses; it loses a star for the non-official split, the short sequences, and the absence of variance or significance reporting.
Writing Quality: ⭐⭐⭐⭐ The motivation chain and the causal role of the three designs are clearly laid out, and the design matrix table settles the differences from the four baselines at a glance; but the contact energy is described inconsistently (log-ratio vs Log-\(\ell_1\)), and some information exists only inside figures.
Value: ⭐⭐⭐⭐⭐ Single-stage, no post-processing, fast inference, with a substantial quantitative improvement in physical plausibility — directly useful for robotics manipulation and AR/VR animation, where contact stability is what matters.