{"id":"99bfab40-1820-463f-be47-9b5864466804","arxiv_id":"2507.05678","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"LiON-LoRA adds a learned scaling token to video-diffusion LoRA adapters, enabling linear and independent control of camera trajectory and object motion strength.","lead":"Researchers trained small LoRA adapters for a video diffusion model to control camera movement and object motion separately, using a new control token that adjusts motion strength. The method claims higher control accuracy and better video quality than prior camera-control systems while needing far less training data.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Decoupled linear scalability is not established: Section 3.5's averaged hidden-state fusion lacks a quantitative multi-scale test, so the central claim that LiON-LoRA linearly and independently controls camera and object motion is unsupported.","rationale":"The reader's conditional verdict is appropriate, and the weakest assumption identified is in fact the load-bearing one. My stress-test pass adds specificity: the averaging in Section 3.5 is applied to hidden states that then pass through further nonlinear transformer blocks, so the assumption is not merely 'averaging might mix signals' but 'the fused model's output is not generally the linear combination of individual LoRA outputs.' No experiment in the paper isolates this property. The quantitative camera-control tables compare fixed trajectories to baselines but never sweep scale coefficients to verify linearity and independence; the motion-strength Pearson correlation in Figure 6 is reported only qualitatively. Thus the central contribution, training-free fusion with decoupled linear scalability, is currently supported only indirectly. This strengthens the need for the conditions the reader already stated, but it does not move the verdict: a focused ablation plus code release could resolve it. I therefore keep CONDITIONAL-equivalent UNCHANGED rather than escalating to REJECT, because the described mechanism is plausible and the missing experiment is well-defined.","tokens_in":18001,"tokens_out":5782,"duration_ms":70963,"concrete_test":"Run a two-primitive fusion experiment on fixed scenes, e.g., 'forward' and 'orbit-left' LiON-LoRAs. First, run single-primitive inference at scales S1, S2 in {0.1, 0.3, 0.5, 0.7, 1.0} and compute the estimated camera trajectory (via GLOMAP poses) for each primitive, giving T1(S1) and T2(S2). Then run the Section 3.5 fusion at all scale pairs and compare the fused trajectory to the linear superposition T1(S1) + T2(S2). Report per-frame rotation/translation error and RMSE, and check whether varying S1 at fixed S2 changes only the forward component while leaving the orbit component unchanged, and vice versa. If the fused trajectory deviates substantially from linear superposition, or the two scales show cross-dependence, the decoupled linear scalability claim fails. For motion strength, similarly measure optical-flow magnitude versus S at fixed camera scale to verify independence.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is the abstract's 'linearly control both camera trajectory and object motion' achieved through training-free fusion. The fusion described in Section 3.5 splits k scaling tokens, runs each LoRA in its own attention subspace, then averages the k sets of visual/textual tokens into H_out while concatenating the k scaling tokens. All subsequent transformer blocks propagate this common averaged H together with all scaling tokens. Averaging is linear, but the transformer blocks are not: softmax attention and LoRA updates do not generally commute with averaging, so the fused model's output is not necessarily the linear combination of the k individual LoRA-controlled outputs. The paper provides no argument, theoretical or empirical, that averaging preserves each LoRA's independent effect. The orthogonality measurement in Section 3.3 (cosine similarity approximately 0.06 +/- 0.06 in shallow layers) characterizes raw LoRA outputs, not the coupling introduced by averaging plus later nonlinear attention layers. Norm consistency (Eq. 4) normalizes per-LoRA outputs before averaging but cannot prevent coupling in subsequent blocks. Crucially, the claimed linear scalability is never quantitatively tested for multi-LoRA fusion: Table 2 reports fixed complex-pose metrics against baselines, Figure 10 compares the scaling token to adapter scaling for a single primitive, and Figure 6 reports only a qualitative Pearson-correlation claim for motion strength without a number or error bars. The headline property therefore rests entirely on the unvalidated averaging assumption identified by the reader; if averaging mixes control signals, the claimed decoupled linear control fails.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes LiON-LoRA, a parameter-efficient framework for controlling camera trajectories and object-motion strength in video diffusion models. It trains separate LoRA modules for discrete camera primitives on rendered DL3DV videos, normalizes per-layer LoRA output norms, and injects a dedicated scaling token (with Fourier positional embedding) to encode the desired motion amplitude. For fusing multiple LoRAs, it runs each LoRA in an attention subspace attending only to its own scaling token and averages the resulting visual/textual hidden states. The same scaling-token mechanism is applied to motion-strength control using static-camera videos, claiming unified spatial and temporal control. Experiments report lower RotErr, TransErr, ATE, and FVD than several baselines, plus a Pearson-correlation analysis for motion strength.","tokens_in":18290,"tokens_out":5841,"duration_ms":66967,"significance":"If the central claims held, LiON-LoRA would be a useful and practical contribution: it achieves strong camera-control metrics with only 100 training scenes and 4,000 fine-tuning steps, keeps the base VDM unchanged at inference, and proposes a clean, training-free fusion scheme. The paper deserves credit for including an ablation on dataset size and training iterations (Table 3), for re-implementing DimensionX-S on the same CogVideoX backbone as a controlled baseline, and for explicitly separating scaling-token control from adapter-scale control in Figure 10. However, the evidence as presented does not yet establish the headline property of decoupled linear scalability for multi-LoRA fusion, and several evaluation choices (train/eval scene separation, error bars, baseline comparability) need to be addressed before the quantitative gains can be taken at face value.","major_comments":[{"comment":"The central claim of decoupled linear scalability is not quantitatively supported. The fusion averages k sets of visual/textual tokens into H_out while concatenating the scaling tokens; however, all subsequent transformer blocks operate on this common averaged H together with all scaling tokens, and softmax attention and LoRA updates do not commute with averaging. Hence there is no reason, without further argument, that the fused output equals a linear combination of the individual LoRA-controlled outputs. The paper provides no multi-LoRA experiment that sweeps the scaling values S1...Sk and checks whether camera trajectory and motion strength respond linearly and independently. Table 2 evaluates fixed complex poses, Figure 10 tests only a single camera primitive, and Figure 6 tests motion strength without fusion. Please add a controlled multi-LoRA scalability experiment (e.g., independent sweeps of camera S and motion S, with the ideal linear prediction overlaid) and, if possible, a diagnostic that measures cross-LoRA coupling.","section":"Section 3.5, Figure 8"},{"comment":"The evaluation split between training and test scenes is unclear. The paper states that 100 DL3DV scenes are reconstructed and rendered for each primitive, and later that 'we selected 100 samples from DL3DV for each motion primitive' for quantitative evaluation, with no statement that these are held-out scenes. If the evaluation scenes overlap with the 100 training scenes, the trajectory metrics in Tables 1-2 may reflect memorization rather than generalization. Please specify the exact train/eval scene split and report metrics separately for seen and unseen scenes.","section":"Section 4.1"},{"comment":"The numerical comparisons lack error bars and are partly across uncontrolled settings. Tables 1 and 2 report a single run per method, so differences such as TransErr 0.197 vs. 0.201 and ATE 0.345 vs. 0.398 in Table 2 may be within sampling noise; multiple seeds and confidence intervals are needed. In addition, FVD is compared across different base models and sequence lengths (e.g., CamI2V is a different architecture and uses shorter sequences), which makes the reported 37.8% FVD improvement not directly attributable to LiON-LoRA. Only DimensionX-S* is a same-backbone re-implementation; please either re-run the other baselines under CogVideoX/49-frame evaluation or clearly restrict the FVD comparison to matched settings.","section":"Tables 1-3, Section 4.3"},{"comment":"The Pearson-correlation result is partially self-consistent by construction. For camera control, S directly defines the number of frames sampled from a rendered trajectory (first 600*S frames), so larger S corresponds to greater motion by design; the analogous construction holds for the 240-frame object-motion videos. Demonstrating that the generated outputs correlate with S is therefore evidence that the model follows its training signal, but it is not independent evidence of linear scalability. Please include held-out, continuously varied S values that are not derived from the frame-subsampling protocol, and compare against adapter scaling with identical normalization, reporting the fitted slope and correlation with error bars.","section":"Section 3.4, Figure 6"}],"minor_comments":[{"comment":"The definition of alpha is ambiguous: 'alpha = P(||Delta W_i||)/k' does not state what P denotes. Please specify whether P is the product, mean, or another aggregation, and clarify that the norm is computed per layer and per sample.","section":"Equation (4)"},{"comment":"The phrase 'we use specific project linear to encode S for different camera LoRAs' should be rewritten. Also specify the exact Fourier dimension J and the minimal scaling values s used in the experiments.","section":"Section 3.4"},{"comment":"The spelling 'Cami2V' appears in the text while the tables use 'CamI2V'; please unify the notation throughout.","section":"Section 4.2"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is promising and the core idea is within scope, but the central claim of decoupled linear scalability needs a dedicated multi-LoRA experiment, and the evaluation needs held-out scenes and error bars. The authors should also release code and the exact evaluation protocol to enable reproducibility. I would not reject, but I cannot recommend acceptance in the current form."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a useful incremental paper with one solid new idea—the scaling token—and one unproven headline claim. The authors show that a learned input token can control motion amplitude more reliably than the usual LoRA adapter scale, and that normalizing LoRA output norms helps stabilize fusion of camera-control LoRAs. That part is genuinely new and worth taking seriously.\n\nWhat the paper does well: the norm-consistency observation is simple and credible, and the ablation showing 100 rendered scenes beats 7k samples after 4k steps is a real result if it holds. The qualitative results are noticeably cleaner than the baselines. The orthogonality measurement (cosine similarity around 0.06 in early layers) is a reasonable empirical observation, though the 'theoretically sufficient' phrasing oversells it.\n\nThe main soft spot is the central claim about linear scalability in the multi-LoRA setting. The fusion scheme in Sec 3.5 runs each LoRA with its own scaling token in a separate attention subspace and then averages the visual/textual hidden states. Averaging is linear, but the transformer blocks that follow are not, so there is no reason to expect the fused output to equal the linear combination of the individual controls. The linearity evidence in Fig 10 and the Pearson correlation in Fig 6 is all for a single scaling token. Nothing quantitatively tests whether two simultaneous controls (forward + orbit, say) are actually decoupled and linear. That is the load-bearing claim, and it is currently supported only by qualitative examples.\n\nThe evaluation has other issues: no error bars over seeds, no explicit train/eval split for DL3DV (the text suggests the 100 rendered scenes might also be the evaluation set), FVD compared across different base models and frame counts, and no comparison to Direct-a-Video even though it is cited as related work. The motion-strength Pearson correlation is also a self-consistency check rather than an external validation, since S is defined by frame subsampling.\n\nNone of this sinks the paper. The scaling token idea is portable, and the norm normalization is a clean fix. It deserves a serious referee and a major-revision request: add a proper multi-scale fusion experiment with two or three simultaneous controls, report variance, clarify the split, add the missing baseline, and release code. I would bring it to reading group to discuss the fusion assumption.","headline":"Useful scaling-token idea for controllable video generation, but the paper's headline claim of decoupled linear multi-LoRA control is untested.","tokens_in":18843,"tokens_out":4099,"would_cite":true,"duration_ms":42755,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that LoRA fusion for video diffusion can be made linear, orthogonal, and norm-consistent, so a single scaling token linearly controls camera trajectory and object motion with far less training data.","keywords":["LiON-LoRA","video diffusion models","LoRA fusion","camera trajectory control","motion strength control","scaling token","norm consistency","orthogonality"],"falsifier":"Run a fusion of a forward LoRA and an orbit-left LoRA, fix one scaling token, sweep the other through its full range, and recover the resulting camera poses with GLOMAP; if the trajectory does not vary monotonically with the swept token while the fixed motion stays constant, the decoupled linear-scalability claim is false.","tokens_in":17790,"feed_emoji":"🎥","tokens_out":7106,"duration_ms":73562,"temperature":0.7,"pith_summary":"LiON-LoRA is proposed to establish that low-level camera and object-motion control in a video diffusion model can be handled by rethinking how LoRA adapters are fused, rather than by injecting explicit camera poses. It rests on three principles: camera LoRAs trained from basic motion primitives are nearly orthogonal in the shallow transformer blocks; normalizing each layer's LoRA output makes combinations smooth; and a dedicated scaling token gives a linear lever over motion amplitude. The paper demonstrates this with a CogVideoX-based system trained on 100 rendered scenes for camera primitives and static-camera videos for object motion, using about 4,000 fine-tuning steps. If the claims hold, trajectory accuracy and motion-strength adjustment improve over explicit-control baselines while using far less annotated data.","feed_headline":"Scaling token makes LoRA video control linear","feed_subtitle":"New LoRA scheme steers camera and object motion with a single token, trained on 100 scenes.","key_machinery":"The engine is the scaling token $E = \\mathrm{linear}(\\gamma(S))$, a trainable embedding of the scalar amplitude $S$ (Fourier positional features encoding $S$ from a small minimum up to $1.0$) concatenated to the visual and textual token sequence $H$. During fusion, $k$ tokens are concatenated to form $H' = [H; E_1; \\dots; E_k]$, and each LoRA computes its own $H_i' = [H; E_i]$ inside its own attention subspace; the $k$ sets of visual and textual tokens are averaged into $H_{\\mathrm{out}}$ while the scaling tokens are concatenated. The other two pillars prepare the ground: per-layer norm normalization $\\Delta \\hat{W}_i = \\alpha \\|\\Delta W_i\\|^{-1} \\Delta W_i$, and the orthogonality observation that shallow-layer camera LoRA outputs have low cosine similarity, so their controls do not interfere.","core_discovery":"The paper's central discovery is that the three failure modes of vanilla LoRA fusion for camera control—correlated features, mismatched magnitudes, and non-linear intensity—all have vector-space fixes. In shallow layers, different camera LoRAs have average cosine similarity of $0.06\\pm0.06$, so they can be treated as nearly orthogonal and hence decoupled; deep-layer correlations are less relevant because low-frequency camera control is encoded early. Norms vary enough across trajectories and blocks that a single adapter scale cannot balance them, so each LoRA output is renormalized to a common norm. Linear amplitude control comes from a Fourier-embedded scaling token appended to the token sequence, letting the model see the desired scale $S$ explicitly instead of tuning an adapter coefficient. At fusion time, each LoRA operates in its own attention subspace with its own scaling token, and the visual and textual outputs are averaged, which the paper argues keeps controls independent without joint training.","pith_inferences":["If the shallow-layer orthogonality observed here is a general property of DiT video models rather than an artifact of CogVideoX, the same three fixes could be dropped into other transformer-based video generators without retraining the base model.","Because the scaling token is just an extra sequence element, one natural extension is to use several tokens per LoRA, such as one for trajectory shape and one for speed, to obtain even richer continuous controls; the paper does not explore that.","Norm-consistency normalization might also stabilize high-level LoRA composition, such as style plus subject, where the paper only tests camera and motion LoRAs; an explicit transfer study would be a cheap check.","A practical product consequence is a two-slider interface for camera path and motion speed, since the reported Pearson correlation suggests the mapping is near-linear; this is an extension, not a result the paper proves."],"forward_implications":["Basic camera-trajectory errors drop by roughly a quarter on rotation error and one-fifth on absolute trajectory error relative to the strongest prior controller tested, with lower FVD on the same base model.","A single camera primitive can be trained with about 100 rendered scenes and 4,000 fine-tuning steps, and the reported ablations show that increasing to 7,000 samples at 4,000 steps does not improve accuracy, so minimal data suffices.","Because fusion is training-free, combining several camera primitives, or camera with object-motion LoRA, requires no joint fine-tuning; the scaling tokens stay separate through attention.","The same scaling token used for camera trajectory also linearly controls object-motion strength, measured by Pearson correlation with optical flow, so spatial and temporal controllability share one mechanism."],"supporting_citations":[{"why":"Defines LoRA low-rank adaptation, the mechanism LiON-LoRA rethinks and fuses.","marker":"[32]"},{"why":"CogVideoX is the base video diffusion model into which the LoRA modules are injected.","marker":"[107]"},{"why":"CameraCtrl supplies the camera-pose evaluation protocol and is a main baseline for trajectory control.","marker":"[23]"},{"why":"MotionCtrl is the unified motion-and-camera baseline used for comparison.","marker":"[91]"},{"why":"CamI2V is a camera-controlled image-to-video baseline and the source of the GLOMAP pose-recovery protocol.","marker":"[119]"},{"why":"DimensionX is the LoRA-based spatial and temporal control baseline whose S-Director is re-implemented for comparison.","marker":"[77]"},{"why":"3D Gaussian Splatting renders the fixed-trajectory training videos from DL3DV scenes.","marker":"[42]"},{"why":"DL3DV provides the 100 reconstructed scenes used to render camera-primitive videos.","marker":"[50]"}],"fun_headline_variants":["LoRA fusion made linear with a scaling token","One token linearly scales camera and object motion","LiON-LoRA: orthogonal LoRA for decoupled video control","Unified spatial-temporal video control via LoRA rethinking","Norm-consistent LoRA stabilizes camera trajectory control"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The argument depends on the averaging of the $k$ attention outputs keeping each LoRA's control signal separate; if hidden states mix, the claimed decoupled linear scalability for camera and object motion fails.","fun_headline_variants_meta":{"raw":{"variants":["LoRA fusion made linear with a scaling token","One token linearly scales camera and object motion","LiON-LoRA: orthogonal LoRA for decoupled video control","Unified spatial-temporal video control via LoRA rethinking","Norm-consistent LoRA stabilizes camera trajectory control"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000516,"raw_usage":{"total_tokens":2520,"prompt_tokens":979,"completion_tokens":1541,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":595,"completion_tokens_details":{"reasoning_tokens":1463}},"tokens_in":595,"tokens_out":1541,"duration_ms":10800,"temperature":1.0,"reasoning_tokens":1463,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T19:20:02.229872+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a fusion of a forward LoRA and an orbit-left LoRA, fix one scaling token, sweep the other through its full range, and recover the resulting camera poses with GLOMAP; if the trajectory does not vary monotonically with the swept token while the fixed motion stays constant, the decoupled linear-scalability claim is false.","supporting_citations":[{"cited_title":"Motionctrl: A unified and flexible motion controller for video generation","cited_arxiv_id":null,"evidence_quote":"MotionCtrl is the unified motion-and-camera baseline used for comparison."}],"review_version":1}