{"id":"e2a52d81-ed18-42bd-ae9f-d352a43e9370","arxiv_id":"2411.17769","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"Scaling the denoising noise by one knob, omega, controls granularity in diffusion outputs globally, per region, or per timestep.","lead":"This paper shows that multiplying the predicted noise by a single parameter, omega, at each diffusion denoising step shifts generated images and videos between smooth and highly detailed. Generalists should care because the trick needs no retraining, works across popular image and video models, and can be limited to selected regions or stages of generation.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claimed granularity control rests on an untested behavioral assumption: that off-manifold residual noise left by omega<1 is interpreted by the network as structure rather than removed as noise; the SNR derivation in Eq. 5 does not establish this.","rationale":"The reader's weakest assumption identifies exactly the same load-bearing point: the effect depends on the pretrained network's response to off-manifold latents, which is asserted rather than derived. My analysis adds precision by noting that Eq. (5) is only a one-step on-manifold calculation and that after the first modified step the latent is off-manifold, so the subsequent denoising behavior is the actual mechanism. This is not an internal inconsistency in the paper; it is a generalization risk. The method is nevertheless credible for the tested models: it is simple, reproducible, supported by frequency-domain measurements, and backed by a user study. The lack of error bars and the absence of a failure analysis across architectures justify the conditional verdict but do not warrant rejection. Therefore the reader's CONDITIONAL verdict remains unchanged.","tokens_in":12007,"tokens_out":6016,"duration_ms":56013,"concrete_test":"Run a controlled experiment on a model outside the paper's set (e.g., a small DDPM trained on LSUN or a latent DiT) with fixed seeds and several prompts. Generate with omega = 0.7, 1.0, and 1.3, and measure three quantities per setting: per-step residual noise variance relative to the forward schedule, final high-frequency energy, and a no-reference artifact score such as DISTS. If HFE moves opposite to Table 2 or artifact scores jump before HFE changes, the behavioral assumption fails for that model. Additionally, run a residual-injection control: at each step, replace Omegance with the standard DDIM update plus Gaussian noise with variance chosen to match c(omega)^2 - (1 - alpha_{t-1}). If this control reproduces the granularity effect, the mechanism is purely added noise; if it does not, the network-specific 'justification' response is confirmed and should be characterized.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Equation (5) is a one-step SNR computation that assumes the input latent z_t is on the forward noising manifold. After a modified step, z'_{t-1} has noise coefficient c(omega) different from sqrt(1 - alpha_{t-1}), so the latent is off-manifold. The next denoising step therefore receives a latent whose SNR does not match its timestep. The central claim depends on what the network does with this mismatch. Section 3.2 asserts that for omega<1 the model 'justifies' the residual noise by generating more intricate structures, and for omega>1 interprets reduced residual noise as a cue to simplify. This is an empirical property of specific pretrained models, not a consequence of the SNR algebra. If a network instead treats the residual as ordinary noise to be removed, omega<1 will yield little change or smoothing, and omega>1 will yield overshoot or artifacts. The HFE measurements in Table 2 and the user study support the assumption on the tested SDXL/FLUX setups, but no independent derivation, error bars, or failure analysis establishes when the assumption holds. Because the paper claims the method is architecture-agnostic and applicable to any progressive denoising process, this missing condition is load-bearing.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces Omegance, a single scalar parameter omega that multiplies the predicted noise term in each reverse-diffusion step. The central claim is that omega<1 retains additional high-frequency residual noise, which the pretrained model interprets as a cue to generate more intricate structures and richer textures, while omega>1 removes more noise and produces smoother, simpler outputs. The same parameter is extended to spatial masks and time-dependent schedules, and experiments are reported across SDXL, FLUX, ControlNet, SDEdit, ReNoise, inpainting, and video diffusion models. Quantitative evidence includes FID, Inception Score, CLIP score, Q-Align, PickScore, SSIM, high-frequency energy measurements, and a user study, along with code released publicly.","tokens_in":12252,"tokens_out":6028,"duration_ms":57098,"significance":"If the central claim holds, Omegance is a practically valuable training-free and architecture-agnostic control for diffusion-based synthesis, with negligible computational cost and simple integration into existing samplers. The paper has concrete strengths: it releases code, evaluates on 1,000 prompts from DiffusionDB, includes multiple base models and tasks, and includes an explicit limitation statement. The claimed mechanism, however, is an empirical property of pretrained networks rather than a consequence of the SNR algebra, and the quantitative tables do not report variance or multiple-seed statistics. The mask and schedule extensions are plausible but would benefit from stronger regional and temporal consistency measurements.","major_comments":[{"comment":"The load-bearing mechanism is asserted rather than tested. Equation (5) computes the one-step SNR under the assumption that the input latent is on the forward noising manifold. After a modified step with omega != 1, the noise coefficient of the latent is no longer sqrt(1-alpha_{t-1}), so subsequent latents are off-manifold and the SNR identity does not directly apply. The key statement in Sec. 3.2 that the model 'justifies' residual noise by generating more intricate structures is an empirical assumption about how a pretrained network responds to off-manifold inputs, not a derivation from the SNR formula. If a network instead removes the residual as ordinary noise, omega<1 would produce little change or smoothing rather than richer detail. Please add a quantitative test of this assumption, for example by measuring the distribution of predicted noise and residual latents before and after modified steps, by comparing against explicit injection of Gaussian noise of comparable magnitude, and by reporting at least one failure case across the tested models.","section":"Sec. 3.2, Eq. (5)"},{"comment":"The quantitative comparisons lack error bars, confidence intervals, or multiple-seed statistics. Table 1 reports FID, IS, CLIP, Q-Align, and PickScore on 1,000 prompts, but several differences are small relative to likely sampling noise, and Table 2 reports SSIM and high-frequency energy for single runs. Without multiple seeds or bootstrap intervals, the claims that Omegance 'outperforms' structure-modification and scheduler-based methods, or that HFE changes align with schedule design, are not firmly supported. Please report mean and standard deviation over at least three seeds, or use paired bootstrap confidence intervals.","section":"Tables 1 and 2"},{"comment":"The locality claim for omega masks is stronger than the evidence provided. The text states that the omega mask 'leverages the locality of the denoising process, ensuring that adjustments to omega in one region do not affect the SNR or visual properties of neighboring areas.' However, latent diffusion backbones commonly have global receptive fields through self-attention or large convolutions, so the assertion is not obvious. The ControlNet results in Fig. 7 are qualitative only. Please add a quantitative region-consistency evaluation, for example measuring SSIM or LPIPS in the unmodified regions with and without masking, or temper the claim to state that cross-region effects are empirically small in the tested configurations.","section":"Sec. 3.2.1"}],"minor_comments":[{"comment":"The column header reads 'HDE (Changes)' while the caption and text refer to High-Frequency Energy (HFE); please align the notation.","section":"Table 2"},{"comment":"The caption introduces S1(t) and S2(t) but does not define them; please add explicit definitions or point to the schedule curves shown in Fig. 6.","section":"Fig. 4"},{"comment":"Equation (5) is typeset in a way that is difficult to parse; a simplified closed form or a step-by-step derivation in the main text would help readers verify the monotonicity argument.","section":"Eq. (5)"},{"comment":"There are minor typos and undefined abbreviations, for example 'denosing' in Sec. 3.2.1 and 'SD3' used without expansion; please proofread and expand all model names.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the scope of the venue and the proposed control is simple enough to be useful. The main risk is that the causal mechanism is presented as a consequence of the SNR algebra when it actually depends on an empirical property of the pretrained network; the requested experiments should be feasible within the manuscript's scope. The absence of error bars in the key tables is the other substantive concern."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this paper is a useful, training-free control for diffusion granularity that probably works as advertised on the models tested. The soft spots are real but don't sink it. Send it to review.\n\nThe genuinely new thing is treating the predicted-noise scale as a single user-facing dial, with spatial masks and temporal schedules on top. Scaling predicted noise is an old primitive, but nobody has packaged it as a granularity control with this breadth. The paper earns credit for testing across SDXL, FLUX, ControlNet, SDEdit, ReNoise, inpainting, and video models, and for releasing code. The DDIM SNR derivation in Eq. 5 is a nice sanity check, and the HFE measurements and the 1,000-prompt table support the direction of the effect. The user study is small but directionally consistent.\n\nThe main soft spot is exactly what the stress-test note flags: the causal story is 'the model justifies residual noise by generating more structure' (Sec. 3.2), which is an empirical assertion, not a consequence of the SNR algebra. Eq. 5 assumes the latent is on the forward noising manifold; after one modified step, z' is off-manifold, and the method's success depends on how the network handles that mismatch. The paper shows it works for the tested models, and the HFE numbers are consistent, but it does not characterize when it would fail. For a method claimed to be architecture-agnostic, that missing condition is a moderate gap, not a fatal one.\n\nTwo smaller issues. The quantitative tables (Tab. 1, 2) have no error bars or multiple-seed stats; differences between settings like omega=6.0 and omega=-6.0 may be within noise for all we know. And the paper's own limitation sentence admits Omegance does not inherently improve quality, which is fine, but the 'fixing artifacts' examples in Fig. 5 are anecdotal. These are worth addressing in revision but don't undermine the central claim.\n\nNet: I would bring this to reading group and cite it as a baseline for inference-time controls. Despite the hand-waving mechanism, the empirical demonstration is broad enough to deserve serious peer review. The authors should be asked to add error bars, failure cases, and a more careful generalization claim.","headline":"A simple inference-time noise-scaling knob that plausibly controls granularity across many diffusion models; the empirical case is solid, the mechanism is hand-waved, but it's worth referee time.","tokens_in":12761,"tokens_out":2077,"would_cite":true,"duration_ms":19278,"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":"A single scalar, omega, multiplying the predicted noise in each reverse-diffusion step controls output granularity across image and video diffusion models without retraining.","keywords":["diffusion models","granularity control","noise prediction scaling","text-to-image synthesis","text-to-video synthesis","detail enhancement","omega mask","omega schedule"],"falsifier":"Run a fixed prompt and base model with $\\omega<1$ and measure high-frequency energy, artifact rate, and text-alignment across a sweep of $\\omega$ values; if some model produces visible noise or content drift instead of richer detail, or if high-frequency energy does not rise as $\\omega$ decreases, the claimed control fails for that model.","tokens_in":11805,"feed_emoji":"🎚️","tokens_out":9617,"duration_ms":78690,"temperature":0.7,"pith_summary":"The paper claims that a single scalar $\\omega$, multiplying the predicted-noise term in every reverse-diffusion step, controls the granularity of generated images and video without retraining or architectural changes. Setting $\\omega<1$ removes less noise, leaving residual high-frequency content that the model responds to by generating more intricate textures and more complex scenes; setting $\\omega>1$ removes more noise and yields smoother, simpler outputs. The same scalar can be applied globally, as a spatial mask for region-specific control, or as a time-varying schedule to affect layout versus fine detail. This matters because existing detail-control tools usually require retraining, guidance, or architecture-specific tricks, whereas Omegance is a single knob that works across the schedulers and models tested.","feed_headline":"One scalar tunes detail in diffusion images and video","feed_subtitle":"Below 1 adds texture and complexity; above 1 smooths and simplifies, with no retraining and almost no extra cost.","key_machinery":"The central object is the scalar $\\omega$ inserted into the reverse-diffusion update $z'_{t-1} = \\delta_t z_t + \\zeta_t \\epsilon_\\theta(z_t,t)\\,\\omega$, with an unbounded-input rescaling used in practice. The load-bearing identity is the modified SNR expression derived in Equation (5) for the DDIM scheduler (a deterministic diffusion sampling rule): any deviation of $\\omega$ from 1 moves the sampling SNR away from the forward SNR, either leaving residual high-frequency content in the latent ($\\omega<1$) or stripping it out ($\\omega>1$). This identity turns a one-number change into a quantitative statement about detail, and it is what lets the method operate across different schedulers without retraining.","core_discovery":"The central discovery is that the denoising network's noise prediction $\\epsilon_\\theta(z_t,t)$ can be rescaled by $\\omega$ without breaking generation, and that this rescaling shifts the effective signal-to-noise ratio of each reverse step. The paper derives a modified SNR schedule (Equation 5) for the DDIM scheduler: $\\omega<1$ makes the sampling SNR lower than the forward-process SNR, so the latent $z'_{t-1}$ retains extra high-frequency information; $\\omega>1$ pushes the SNR higher and suppresses high frequencies. The authors state that the model 'justifies' residual noise by generating more intricate structures, which is the empirical mechanism behind richer detail. The same operation extends to Euler and flow-matching schedulers, to spatial $\\omega$ masks, and to temporal $\\omega$ schedules, and experiments across image and video models show consistent granularity shifts.","pith_inferences":["Editorial inference: because the stated mechanism is residual high-frequency content, the same omega scaling should transfer to any iterative denoiser with an explicit noise or velocity prediction, such as consistency models or discrete diffusion; testing there would map the true boundary of the claim.","Editorial inference: Equation (5) could be inverted to design omega schedules that steer the sampling SNR along any target curve, turning the hand-picked EXP and COS schedules in the paper into a principled scheduling problem.","Editorial inference: the mask results suggest a composable editing tool for existing images: an omega mask derived from segmentation could re-balance detail between subject and background during real-image editing, which the paper demonstrates for generation but not for post-hoc editing."],"forward_implications":["Users can dial global granularity continuously: $\\omega<1$ gives busier scenes and richer textures, while $\\omega>1$ gives cleaner, smoother images.","Spatial omega masks let a character be sharpened while its background is simplified, or vice versa, with untouched regions left nearly unchanged.","Temporal omega schedules separate layout from texture: adjusting $\\omega$ in roughly the first 10 of 50 denoising steps affects scene complexity, while later adjustments affect fine detail.","The method transfers across DDIM, Euler, and flow-matching schedulers and across text-to-image, image-to-image, inpainting, editing, and text-to-video models without retraining and with negligible extra compute.","The same knob can incidentally fix artifacts on weaker models ($\\omega>1$ suppresses malformed fingers) and restore realism on over-smoothed models ($\\omega<1$ re-adds texture in FLUX)."],"supporting_citations":[{"why":"It supplies the DDIM reverse-step formulation and the signal-to-noise definition from which the paper derives its modified SNR schedule in Equation (5).","marker":"[39]"},{"why":"It establishes the forward noise-addition process and the alpha_t schedule that the omega-scaling analysis operates on.","marker":"[15]"},{"why":"It provides the progressive denoising-dynamics description, with layout emerging early and fine detail late, that motivates the omega schedules.","marker":"[40]"},{"why":"It serves as the primary text-to-image base model for the global, mask, and schedule experiments and for the quantitative comparisons.","marker":"[30]"},{"why":"It supplies the flow-matching scheduler used for SD3 and FLUX, for which the paper introduces a mean-preserving omega variant.","marker":"[8]"},{"why":"It provides the ControlNet testbed used to demonstrate omega masks driven by pose, depth, and canny control signals.","marker":"[48]"},{"why":"It acts as the image-to-image editing pipeline that Omegance modifies to add granularity control.","marker":"[26]"},{"why":"It is the main training-free quality-enhancement baseline compared against, replaced here by a single scalar.","marker":"[38]"},{"why":"It supplies the rescaled noise-schedule baseline that Omegance is compared against in the quantitative tables.","marker":"[24]"},{"why":"It supplies the cosine noise-schedule baseline that Omegance is compared against in the quantitative tables.","marker":"[28]"}],"fun_headline_variants":["One scalar controls detail in diffusion images and video","Single parameter adjusts granularity without retraining","Tune diffusion detail with one omega, no retraining","One knob: rescale noise to control diffusion detail","Diffusion detail dial: one scalar, zero retraining"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a pretrained diffusion network responds to a latent carrying extra residual high-frequency noise by generating more complex structures and textures, rather than by producing artifacts or shifting content; this behavior is asserted for the tested models, not derived.","fun_headline_variants_meta":{"raw":{"variants":["One scalar controls detail in diffusion images and video","Single parameter adjusts granularity without retraining","Tune diffusion detail with one omega, no retraining","One knob: rescale noise to control diffusion detail","Diffusion detail dial: one scalar, zero retraining"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000221,"raw_usage":{"total_tokens":1415,"prompt_tokens":872,"completion_tokens":543,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":488,"completion_tokens_details":{"reasoning_tokens":468}},"tokens_in":488,"tokens_out":543,"duration_ms":5156,"temperature":1.0,"reasoning_tokens":468,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T12:22:40.171232+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a fixed prompt and base model with $\\omega<1$ and measure high-frequency energy, artifact rate, and text-alignment across a sweep of $\\omega$ values; if some model produces visible noise or content drift instead of richer detail, or if high-frequency energy does not rise as $\\omega$ decreases, the claimed control fails for that model.","supporting_citations":[{"cited_title":"Denois- ing diffusion implicit models","cited_arxiv_id":null,"evidence_quote":"It supplies the DDIM reverse-step formulation and the signal-to-noise definition from which the paper derives its modified SNR schedule in Equation (5)."},{"cited_title":"SDXL: Improving latent diffusion mod- els for high-resolution image synthesis","cited_arxiv_id":null,"evidence_quote":"It serves as the primary text-to-image base model for the global, mask, and schedule experiments and for the quantitative comparisons."},{"cited_title":"Scaling rec- tified flow transformers for high-resolution image synthesis","cited_arxiv_id":null,"evidence_quote":"It supplies the flow-matching scheduler used for SD3 and FLUX, for which the paper introduces a mean-preserving omega variant."},{"cited_title":"SDEdit: Guided image synthesis and editing with stochastic differential equa- tions","cited_arxiv_id":null,"evidence_quote":"It acts as the image-to-image editing pipeline that Omegance modifies to add granularity control."},{"cited_title":"FreeU: Free lunch in diffusion U-Net","cited_arxiv_id":null,"evidence_quote":"It is the main training-free quality-enhancement baseline compared against, replaced here by a single scalar."},{"cited_title":"Common diffusion noise schedules and sample steps are flawed","cited_arxiv_id":null,"evidence_quote":"It supplies the rescaled noise-schedule baseline that Omegance is compared against in the quantitative tables."},{"cited_title":"Improved denoising diffusion probabilistic models","cited_arxiv_id":null,"evidence_quote":"It supplies the cosine noise-schedule baseline that Omegance is compared against in the quantitative tables."}],"review_version":1}