{"id":"73f37b2d-8df7-494e-85a9-fda53bd2258a","arxiv_id":"2608.12107","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":9,"one_line_summary":"Decoupled parallel training of few-step distillation and rollout-based long-horizon adaptation, together with chunk-wise history feature caching, enables real-time 768x512 infinite audio-driven avatars at 27.2 FPS.","lead":"Avatar-Forever trains fast and stable avatar video generation in two separate branches that are merged at deployment, reaching 27.2 frames per second on one H100 GPU. It also caches history features during streaming to cut redundant computation while keeping identity and lip-sync over long conversations.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Additive merge in Eq. (6) is untested: RRT LoRA trained on 30-step base rollout is added to a 4-step DMD student, with no sequential or step-matched ablation.","rationale":"The reader's synthetic-data concern is legitimate but only partially load-bearing: RRT is a robustness correction, not a full appearance model, and the external real-video evaluation on EMTD, HDTF, and TalkVid provides direct evidence of transfer. The merge assumption in Eq. (6) is more central to the paper's methodological novelty: few-step efficiency and long-horizon robustness are claimed to be learnable in parallel and composable additively. If the base-trained RRT LoRA does not transfer to the DMD student, the framework's advantage over sequential training is not demonstrated. Table 3 gives prima facie support, but it lacks the critical control that would isolate the transfer, and the step-count mismatch between 30-step RRT training and 4-step DMD deployment is unaddressed. This concern does not warrant rejection, because the merged model does improve over DMD-only and the method is described in enough detail to re-implement. It does strengthen the case for the conditional verdict already given, so no verdict change is needed.","tokens_in":20048,"tokens_out":8468,"duration_ms":79500,"concrete_test":"Add one ablation row to Table 3: train the RRT LoRA directly on the four-step DMD student (θ0 + Δθ_DMD) using the same K=4 rollout, and use a 4-step denoising schedule during RRT training. Evaluate this sequential baseline and the Eq. (6) merged model on the 30s EMTD split with at least three seeds and bootstrap CIs on LLM Overall. If the sequential adapter is not statistically worse than the merged model (or is better), the additive merge in Eq. (6) is unsupported; if the merged model matches sequential, the decoupling survives this test.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing condition for the central claim is the additive composition in Eq. (6): θ* = θ0 + Δθ_DMD + Δθ_RRT. The paper states that both branches start from θ0 (Sec. 3.2, Deployment), and that RRT rollout uses the default 30-step denoising schedule (Sec. 5.1), while the efficiency branch produces a four-step DMD generator. For Eq. (6) to be valid, two things must hold: (i) a LoRA delta computed against the base model remains a correct robustness correction when added to the DMD-updated weights θ0 + Δθ_DMD, and (ii) a correction calibrated under 30-step base-model rollout error dynamics transfers to the 4-step DMD inference regime. Table 3 shows the merged model improves over DMD-only, which is positive but not isolating evidence: there is no sequential baseline (RRT trained directly on the DMD student), no 4-step RRT training baseline, and no control with a random LoRA of the same rank. If the RRT delta is not transferable, the decoupled training claim is not established, and the 27.2 FPS 'unbounded' claim rests on an engineered additive heuristic rather than on the proposed decoupling.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Avatar-Forever, a framework for long-horizon audio-driven avatar generation built on a 22B video foundation model. It decouples training into an efficiency branch (full-parameter DMD distillation to a four-step generator) and a robustness branch (Recovery-oriented Rollout Training, RRT, a LoRA adapter trained by perturbing early history, rolling out K chunks without gradient, and applying flow-matching supervision only to the following target chunk). At deployment the two deltas are added to the base weights (Eq. 6), and ForeverCache caches historical chunk features after the first denoising step and forwards only current tokens in later steps. Training data are synthesized by the same base model from filtered dialogue prompts. Experiments on EMTD, HDTF, and TalkVid compare 5s and 30s clips against four prior systems using an LLM judge, automatic metrics, and a 20-person user study, reporting 27.2 FPS at 768x512 on an H100.","tokens_in":20365,"tokens_out":6784,"duration_ms":60377,"significance":"The significance is potentially high if the central claim holds: few-step efficiency and long-horizon robustness can be trained independently and composed additively, and history features can be cached chunk-wise without quality loss. The manuscript has several genuine strengths: the RRT objective is a simple, well-motivated standard flow-matching loss under autoregressive rollout; the paper publishes the evaluation prompt in an appendix; the decoupling hypothesis is tested with a relevant three-way ablation (Table 3); and the external evaluation spans three datasets. Falsifiable claims are clearly stated (Eq. 6, 27.2 FPS, 23% throughput gain). However, the load-bearing conditions—transferability of the RRT delta across denoising schedules, validity of self-synthetic training data for real-world robustness, and the exactness of ForeverCache—are not yet established, so I cannot recommend acceptance without revision.","major_comments":[{"comment":"The central claim that the RRT LoRA delta, trained against the 30-step base-model rollout, transfers to the four-step DMD generator is not isolated by any experiment. Table 3 shows that adding RRT to DMD improves over DMD-only, but it does not rule out that the improvement comes from a generic LoRA on the DMD student, nor does it test whether a correction calibrated under 30-step rollout errors is valid under the 4-step inference used at deployment. Please add a sequential control (RRT trained on the DMD student), a 4-step rollout training control, and a same-rank random-LoRA control; without these, Eq. (6) is an untested composition rule rather than evidence for decoupled parallel training.","section":"Section 3.2, Eq. (6); Section 5.1"},{"comment":"The training corpus is synthesized by the same LTX-2.3 model that Avatar-Forever adapts, so RRT is trained on the base model's own output distribution; failure modes that occur only with real speakers, natural lighting, camera noise, or recording artifacts are absent from training. The external evaluation uses 40-sample splits per dataset, the principal perceptual metric is an author-designed LLM judge without reported agreement with human ratings or run-to-run variance, Table 1 reports no confidence intervals, and the extended-duration claim rests on a single 11-minute qualitative example. Please report error bars and per-sample variance, validate the LLM judge against the human study (or report inter-rater agreement), and include a longer real-video evaluation or a targeted domain-shift test before claiming 'unbounded' real-world avatar generation.","section":"Section 4 and Section 5.1"},{"comment":"ForeverCache's correctness rests on the assumption that feature representations of clean historical chunks computed at the first denoising step remain valid at later noise levels, but Eq. (8) and Eq. (9) simply define two execution modes and do not establish that the cached attention values approximate the full-window forward pass. The only evidence is the global gap between 'Ours' and 'Ours w/ FC' in Table 1 and qualitative figures; no per-step velocity error or a cached-vs-recomputed ablation on identical chunks is reported. A direct measurement of the approximation error is needed to support the claim that the cache preserves the denoising interface and removes redundant computation without quality loss.","section":"Section 3.3, Eqs. (8)-(9)"}],"minor_comments":[{"comment":"The phrase 'complicated foring strategies' appears to be a typo for 'complicated forcing strategies'.","section":"Section 2.2"},{"comment":"The label '在此处键入公式' in Figure 1 is an untranslated placeholder and should be removed or replaced with an English caption.","section":"Figure 1"},{"comment":"The abstract's 'unbounded' claim is stronger than the evidence, which covers 5s and 30s splits plus one 11-minute qualitative example; please qualify the wording accordingly.","section":"Abstract and Section 5.2"},{"comment":"The DMD reference is inconsistent: Section 3.1 cites [44] while Section 5.1 cites [43]; please use the correct reference for the DMD variant actually adopted.","section":"Section 3.1 and Section 5.1"},{"comment":"The OmniAvatar latency entry '850.00/>1h' uses inconsistent units and formatting; please define the notation and ensure all latency entries are comparable.","section":"Table 1"},{"comment":"In Eq. (3), the notation sg(G_theta(...)) is used for a single denoising step, but G_theta elsewhere denotes the full generator; please clarify the sampler-step notation.","section":"Section 3.2, Eq. (3)"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is in scope and the central idea is attractive. My main concern is that the headline claims (Eq. 6 composition and cache-preserved quality) are supported only by aggregated comparisons without the isolating ablations the claims require; I would like the editor to prioritize those experiments if the paper is revised."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Worth a serious referee, but the load-bearing claim is not yet proven. The paper's core move is to train efficiency (four-step DMD distillation) and long-horizon robustness (RRT LoRA adapter) in parallel and then add the deltas at deployment: θ* = θ0 + Δθ_DMD + Δθ_RRT. That composition sounds clean, but the evidence for it is incomplete.\n\nWhat is genuinely new: RRT supervises only after K autoregressive rollout steps from a perturbed early history, so the model learns recovery under its own error dynamics rather than from one-step corruption; ForeverCache reuses per-block historical features across denoising steps, cutting redundant history computation by about 23%; and the parallel-branch framing is a real alternative to the sequential forcing and distillation pipelines in the cited self-forcing and hybrid-forcing work. Table 3 supports the minimal claim that both branches matter: DMD-only drifts, FM-only looks poor, and DMD+RRT is best on every reported metric. External evaluation on EMTD, HDTF, and TalkVid plus a 20-person user study is real grounding, and the 27.2 FPS throughput at 768x512 on one H100 is a concrete engineering result.\n\nThe soft spots are in proportion. The stress-test note is on target: Eq. (6) is an untested additive-merge assumption. RRT rollout uses the default 30-step denoising schedule, while the efficiency branch produces a four-step DMD student. The paper never runs RRT directly on the student, never trains RRT at four steps, and never checks a random LoRA of the same rank as a control. Table 3 shows the merged model beats DMD-only, but that does not isolate whether the robustness delta transfers across the distillation gap. Without that control, the decoupled-training claim is a plausible hypothesis, not an established result.\n\nThe evaluation itself is thinner than the claims. Forty samples per split, no error bars; the LLM judge is designed by the authors; twenty raters is small. Training data are synthesized by the same LTX base model being adapted, so the self-training loop is disclosed but not stress-tested for real-speaker failure modes. No code or checkpoints accompany the preprint. These are fixable issues, not dealbreakers. ForeverCache's feature-reuse approximation is also not analyzed, but for an inference-only mechanism that is a minor concern. The related-work coverage is adequate; the overlap with the authors' own Hybrid Forcing work is acknowledged, though the precise delta would be sharper with the missing baselines.\n\nI would send this to referees. Ask for the sequential baseline, a 4-step RRT training baseline, a random-LoRA control, error bars, and at least one real-data transfer check before accepting. The core ideas deserve attention, but the central claim needs the missing experiments.","headline":"Worth a serious referee, but the load-bearing decoupling claim rests on an additive merge that the paper does not actually isolate.","tokens_in":20880,"tokens_out":2966,"would_cite":true,"duration_ms":26312,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Decoupled training yields real-time infinite avatars at 27.2 FPS.","keywords":["audio-driven avatar generation","streaming video generation","distribution matching distillation","flow matching","autoregressive rollout training","feature caching","synthetic data pipeline","real-time diffusion inference"],"falsifier":"Run the trained system for several minutes on real conversational videos from a held-out dataset with camera noise, motion blur, and unseen lighting. If identity or lip-sync drift accumulates over time at the same rate as the distillation-only model, then the synthetic training distribution does not cover the failure modes that matter, and the central claim fails.","tokens_in":19803,"feed_emoji":"🎥","tokens_out":6597,"duration_ms":57116,"temperature":0.7,"pith_summary":"This paper argues that efficient few-step generation and long-horizon robustness are two independent capabilities of a streaming avatar generator, and that they should be trained in parallel rather than in a single sequential distillation pipeline. It builds a 22-billion-parameter video foundation model into a four-step generator by distillation in one branch, and trains a lightweight low-rank adapter in another branch to recover from errors that accumulate over autoregressive rollout. At inference the two updates are added to the original weights, producing one generator that is both fast and stable. A chunk-wise history feature cache removes redundant recomputation of fixed context during streaming. If the claims hold, avatar systems can be made real-time and unbounded in duration without the fragile, stage-coupled training pipelines currently used.","feed_headline":"Two parallel branches stabilize infinite avatars at 27 FPS","feed_subtitle":"A 22B video model gains four-step speed and long-horizon robustness separately, then merges them for 768x512 streaming.","key_machinery":"The load-bearing mechanism is the additive composition of two separately trained parameter updates, implemented through Distribution Matching Distillation for the efficiency branch and Recovery-oriented Rollout Training with a masked flow-matching objective for the robustness branch. RRT is the conceptual engine: it simulates the inference-time failure mode by corrupting the earliest history chunk, letting the model generate several future chunks under its own predictions, and then supervising recovery only at the final target chunk, so the adapter learns to undo accumulated drift rather than one-step noise. ForeverCache is the inference-time mechanism: it performs a full-window forward pass at the first denoising step, caches per-block history features, and reuses them while denoising only the current chunk, removing redundant computation without changing the learned weights.","core_discovery":"The central claim is the composition identity $\\theta^* = \\theta_0 + \\Delta\\theta_{\\mathrm{DMD}} + \\Delta\\theta_{\\mathrm{RRT}}$: starting from the pretrained base weights $\\theta_0$, a full-parameter distillation branch learns the few-step update $\\Delta\\theta_{\\mathrm{DMD}}$, and a separate low-rank adaptation branch learns the long-horizon update $\\Delta\\theta_{\\mathrm{RRT}}$; the two updates can be trained independently and then merged additively. The robustness branch uses Recovery-oriented Rollout Training (RRT), which degrades an early history chunk, rolls out several future chunks without gradients, and applies a standard flow-matching loss only to the next ground-truth chunk after the degradation has propagated. The efficiency branch never sees autoregressive rollout, and the robustness branch never distills. Together with the ForeverCache inference-time cache, the final generator is claimed to support unbounded audio-driven avatar generation at 768x512 resolution with end-to-end throughput of 27.2 FPS on a single H100 GPU, while maintaining identity consistency, motion coherence, and visual fidelity.","pith_inferences":["Beyond the paper's claims, the same decoupling principle suggests that any streaming video task with autoregressive drift could train a robustness adapter in parallel with a distillation branch and merge them additively.","Because the adapter is trained on videos produced by the same base model, its robustness may be bounded by error modes the base model can produce; mixing in real degraded footage or adversarially perturbed synthetic frames is a testable extension the paper does not explore.","ForeverCache's insight, caching stable context features across denoising steps, could extend to other fixed conditioning streams such as audio features or reference images, potentially cutting latency further."],"forward_implications":["Few-step efficiency and long-horizon robustness can be optimized by separate objectives and composed additively, so training pipelines become simpler to diagnose and scale.","Corrupted-history training alone is insufficient: RRT's benefit comes from supervising recovery after model-induced errors have propagated through several rollout chunks, with $K=4$ giving the most stable results.","ForeverCache raises streaming throughput by roughly 23% on short videos and 45% on 30-second generation while preserving most perceptual gains, with no retraining.","A fully synthetic data pipeline from the same base model can supply long-horizon avatar training data without large-scale curation of real long videos.","The combined system supports continuous generation beyond 11 minutes without visible drift, with 27.2 FPS end-to-end throughput at 768x512 on a single H100."],"supporting_citations":[{"why":"Supplies the 22-billion-parameter joint audio-visual foundation model that both adaptation branches start from.","marker":"[6]"},{"why":"Defines the standard flow-matching objective that RRT applies to the final target chunk.","marker":"[35]"},{"why":"Provides the distribution matching distillation method used by the efficiency branch to obtain a four-step generator.","marker":"[44]"},{"why":"Introduces corrupted-history training that RRT extends by supervising recovery after multi-chunk autoregressive rollout.","marker":"[21]"},{"why":"Supplies the low-rank adaptation mechanism used for the lightweight long-horizon adapter.","marker":"[60]"}],"fun_headline_variants":["Decoupled training yields infinite avatars at 27 FPS","Parallel branches: speed and stability for endless avatars","Train speed and robustness apart, then merge for infinite avatars","Decoupled parallel training makes infinite avatars real-time","Infinite avatars at 27 FPS: train branches in parallel"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that videos synthesized by the same base video model, after automatic filtering, form a valid training distribution for long-horizon robustness, so the adapter trained on them transfers to real human videos with real lighting, noise, and recording conditions.","fun_headline_variants_meta":{"raw":{"variants":["Decoupled training yields infinite avatars at 27 FPS","Parallel branches: speed and stability for endless avatars","Train speed and robustness apart, then merge for infinite avatars","Decoupled parallel training makes infinite avatars real-time","Infinite avatars at 27 FPS: train branches in parallel"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000682,"raw_usage":{"total_tokens":3153,"prompt_tokens":1055,"completion_tokens":2098,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":671,"completion_tokens_details":{"reasoning_tokens":2014}},"tokens_in":671,"tokens_out":2098,"duration_ms":12734,"temperature":1.0,"reasoning_tokens":2014,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T00:16:15.475944+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the trained system for several minutes on real conversational videos from a held-out dataset with camera noise, motion blur, and unseen lighting. If identity or lip-sync drift accumulates over time at the same rate as the distillation-only model, then the synthetic training distribution does not cover the failure modes that matter, and the central claim fails.","supporting_citations":[{"cited_title":"Lora: Low-rank adaptation of large language models.Iclr, 1(2):3, 2022","cited_arxiv_id":null,"evidence_quote":"Supplies the low-rank adaptation mechanism used for the lightweight long-horizon adapter."}],"review_version":1}