{"id":"11356ff1-6143-48a8-8c36-a16f6cfcc5a8","arxiv_id":"2507.02792","paper_version":5,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"RichControl decouples condition-feature injection timesteps from the denoising process, improving training-free spatial control for text-to-image diffusion models.","lead":"This paper presents a training-free framework that improves how text-to-image diffusion models follow structural guides like edge maps and poses. The method works by injecting structure features at a different denoising timestep than usual, then refining the image with restarts and enriched prompts.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"FLUX extension uses synchronous g(t)=t, contradicting the claimed generality of the decoupled g(t)=600 schedule.","rationale":"The reader correctly flags the KL/self-similarity proxy in Sec. 3 as a weak link, and the lack of error bars and held-out validation. I agree those are real concerns. But the single most load-bearing issue is the paper's own supplementary report that the proposed decoupled schedule is abandoned on FLUX in favor of synchronous injection. This directly undercuts the headline contribution's scope: the framework is advertised as architecture-agnostic, yet the one architecture beyond UNet shown in the paper performs better with the baseline schedule. This is not a matter of outside consensus; it is an internal contradiction between the main text's emphasis on g(t)=600 and the appendix's admitted preference for g(t)=t on DiT. The conditional recommendation stands, but the condition should explicitly require either (a) a quantitative schedule ablation on FLUX demonstrating when decoupling helps, or (b) a revised claim scoped to UNet/SDXL. My disagreement with the reader's weakest_assumption is partial: the proxy analysis is a plausible threat, but the FLUX inconsistency is a more concrete and falsifiable defect in the central argument.","tokens_in":24130,"tokens_out":5203,"duration_ms":60123,"concrete_test":"Run the SRI schedule ablation on FLUX.1-dev following the Fig. 6 protocol (constant schedules C=200, C=400, C=600, C=800 and synchronous g(t)=t), with RR and ARP fixed, over the dataset from Supp. E.2 or a held-out subset. Report Self-sim, CLIP, LPIPS, DreamSim, ImageReward, and HPSv2. If the constant schedule does not match or beat the synchronous schedule, the central claim must be scoped to UNet-based SDXL and the abstract's architecture-generality claim should be revised.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The paper's core contribution is the decoupled injection schedule g(t)=600 (Sec. 4.1), presented as the principled resolution of the structure/domain trade-off identified in Sec. 3. The abstract and Fig. 7(b) claim this framework 'generalizes across architectures' to DiT backbones such as FLUX. However, Supp. Sec. E.1 states: 'We found that a synchronous schedule g(t)=t performs slightly better than the constant schedule g(t)=C, so we adopt the former in our experiments.' Thus the FLUX results—the primary evidence for DiT generality—were produced with the baseline synchronous schedule, not with the proposed decoupled schedule. This is an internal inconsistency: the central mechanism is abandoned exactly in the setting used to claim plug-and-play generalization. Either the decoupling insight is architecture-specific to UNet/SDXL, in which case the 'new axis for tuning' does not transfer to DiT, or the FLUX comparison provides no evidence for the schedule contribution. Additionally, the SDXL ablation (Fig. 6) and the chosen constants (tau=400, C=600) are tuned on the same evaluation dataset, so the schedule is partly a fitted hyperparameter rather than a model-derived principle. The central claim's scope is therefore unsupported as stated.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes RichControl, a training-free framework for conditional text-to-image generation. Its central idea is to decouple the condition-feature injection timestep from the denoising process via a schedule g(t), and, after an ablation over constant, linear, and nonlinear schedules, to recommend a constant schedule g(t)=600 on SDXL. The framework combines this structure-rich injection with a restart refinement schedule and an appearance-rich prompting stage, and reports improved structural alignment, prompt alignment, and user preference against training-free and some training-based baselines. It also demonstrates plug-in use with FreeControl and an extension to the DiT-based FLUX backbone.","tokens_in":24543,"tokens_out":7624,"duration_ms":84568,"significance":"If the central claims are supported, the paper opens a practical new tuning axis for training-free controllable generation, with a simple constant schedule that is easy to implement and cache, and it provides a useful analysis of feature-domain mismatch. The component design is modular, the qualitative coverage spans many condition types, and the plug-in demonstration with FreeControl is a strength. However, the load-bearing schedule conclusion is currently validated only on a small, self-constructed dataset, and the FLUX extension abandons the proposed schedule, so the scope of the contribution is narrower than the abstract states.","major_comments":[{"comment":"The FLUX experiments do not test the decoupled schedule. Supp. E.1 states that for FLUX \"a synchronous schedule g(t)=t performs slightly better than the constant schedule g(t)=C, so we adopt the former,\" while the main text and Fig. 7(b) claim the framework \"generalizes across architectures.\" The comparison with Qwen-Image-Edit and FLUX Kontext in Fig. 7(b) therefore provides no evidence for the paper's central mechanism on DiT backbones; it shows only that the pipeline with synchronous injection, restart refinement, and appearance-rich prompting can be run on FLUX. This is an internal inconsistency between the abstract's generality claim and the implementation used for the FLUX results. Please add a FLUX ablation over g(t) (including g(t)=C) or explicitly limit the generalization claim to the UNet/SDXL setting.","section":"Supp. E.1, §5.4, Fig. 7(b)"},{"comment":"The central schedule parameters are selected and reported on the same evaluation data. C=600 and tau=400 are fixed by an ablation (Fig. 6, Fig. 16, Supp. E.6) over the same dataset used for the final quantitative comparisons in Tab. 4 and the user study, and the same applies to N=3 for restart refinement. The text says all experiments were repeated three times, yet no standard deviations, confidence intervals, or significance tests are reported anywhere in the tables. The difference in Self-Sim between Ours (0.096) and Ctrl-X (0.104) in Tab. 4, for instance, is impossible to interpret without variance and could easily be noise on roughly 130 condition-prompt pairs. The paper should either hold out a validation set for hyperparameter selection or report error bars and per-condition breakdowns, and it should avoid describing the empirical optimum as a \"principled\" model-derived result.","section":"§5.3, §5.1, Tab. 2, Tab. 4, Supp. E.6"},{"comment":"The motivational analysis is a proxy whose link to generation quality is not established. Eq. (7) measures KL divergence between Gaussian KDEs fitted to the first PCA component of diffusion features; both the PCA projection and the 1D density estimate are substantial reductions of the diffusion feature space, and no evidence is given that this scalar tracks structure preservation or visual quality under injection. Similarly, the self-similarity curve in Fig. 2 is computed on DINO keys of images, not on the injected condition features. To make the claim that the decoupled schedule is \"guided by\" this analysis, please show a quantitative correspondence between the proxy metrics and the downstream Self-sim/ImageReward/HPSv2 values across the schedules in Fig. 6, or present the analysis as purely illustrative.","section":"§3, Supp. C.1"}],"minor_comments":[{"comment":"The schedule labels A-I are not defined in the caption or in the surrounding text; please add a table or legend specifying the functional form of each schedule.","section":"Fig. 6"},{"comment":"The metric called \"Condition LPIPS\" is described in the main text as appearance quality, but Supp. E.5 defines it as perceptual deviation from the condition image, with higher being better; please make this direction explicit in the main text.","section":"§5.1, Supp. E.5"},{"comment":"The FLUX implementation differs from the main SRI formulation by injecting only self-attention query matrices and by using different layers and timestep thresholds; please state explicitly how this relates to the feature and attention substitution described in §4.1.","section":"Supp. E.1"},{"comment":"The user study reports a single preference rate of 56.25% without confidence intervals or per-case agreement; please report the number of votes per method and a variance estimate.","section":"Tab. 1, Supp. E.3"},{"comment":"The evaluation dataset is self-constructed and small, but the paper does not state whether it will be released; please add an availability statement, since the quantitative claims rest entirely on this dataset.","section":"Supp. E.2"}],"recommendation":"major_revision","confidential_remarks":"The main blocker is the FLUX inconsistency: the paper claims DiT generality for the decoupled schedule while the FLUX experiments use a synchronous schedule. If the authors can provide a FLUX schedule ablation or revise the generalization claim, the paper is a plausible accept after major revision. I would also ask the editor to weigh whether the small self-constructed dataset is adequate for the \"state-of-the-art\" claim without variance or a held-out validation split."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the decoupled injection schedule is a real, useful improvement for training-free spatial control on SDXL, and the paper deserves peer review. But the claimed generalization to FLUX is contradicted by their own supplementary: they used the old synchronous schedule for all FLUX experiments.\n\nThe paper does several things well. The systematic investigation of schedule shapes (linear, quadratic, log, trig) is new and useful, and the finding that schedules with values around medium timesteps work well regardless of functional form is a genuinely empirical insight. The constant schedule g(t)=600 is simple, easy to cache, and appears to improve both structure preservation and visual quality over synchronous injection. Restart refinement and appearance-rich prompting are sensible plug-in additions, and the paper tests across seven condition types with a broad set of baselines, including a plug-in to FreeControl. The writing is clear and the qualitative results look strong.\n\nThe soft spots are real, though not fatal for the SDXL story. The FLUX extension is the big one. Supp. E.1 says they found the synchronous schedule g(t)=t performs slightly better on FLUX and adopted it. So the FLUX comparison, which is the evidence for the 'generalizes across architectures' claim, does not use the proposed mechanism. Either the decoupling insight is UNet-specific, in which case the statement is overbroad, or the FLUX result is just a demonstration of the overall framework, in which case it doesn't support the schedule contribution. The paper should state this explicitly and, ideally, run the constant schedule on FLUX.\n\nSecond, the constants C=600 and tau=400 are selected by ablating on the same evaluation dataset used for the final comparisons. The ablation itself is the evidence, so there is a circularity burden. A held-out validation or at least error bars across seeds would help. The paper says experiments were repeated three times, but no variance is reported anywhere. The user study is a single forced-choice vote with no variance. These are fixable with modest effort.\n\nThe PCA/KDE analysis in Sec. 3 is a proxy, but it is used as motivation, not as proof; the main evidence is the ablations and comparisons. I would not sink the paper on that.\n\nWho is this for? Anyone building training-free control or extending FreeControl-type pipelines. The SDXL result is likely reproducible, and the schedule design space is worth knowing. I'd send it to peer review, with the expectation of a major revision: tighten the FLUX claim, add held-out validation and variance reporting, and clarify that the constant schedule is the contribution while the framework, not the schedule, is what transfers.","headline":"Decoupled injection schedule is a real win on SDXL, but the FLUX generality claim is undercut by their own supplementary note.","tokens_in":24917,"tokens_out":3029,"would_cite":true,"duration_ms":33891,"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":"Spatial control improves when condition features skip ahead to timestep 600.","keywords":["text-to-image generation","training-free spatial control","feature injection","injection schedule","diffusion models","structure preservation","appearance quality","restart refinement"],"falsifier":"A systematic grid over constant schedules $g(t)=C$ on the same SDXL setup that shows synchronous injection $g(t)=t$ matching or beating $C=600$ on both structural and appearance metrics, or a full-dimensional feature analysis that reverses the KL/self-similarity trade-off, would falsify the central claim.","tokens_in":23948,"feed_emoji":"🎨","tokens_out":4991,"duration_ms":53123,"temperature":0.7,"pith_summary":"This paper argues that training-free spatial control of text-to-image diffusion models fails when the condition image's features are injected at the same denoising step as the output image. The authors identify a temporal trade-off: as diffusion progresses, condition features become structurally richer but increasingly out-of-distribution relative to natural image features. The central proposal is a decoupled injection schedule, sampling condition features from a fixed middle timestep ($g(t)=600$) rather than synchronously, which balances structure preservation and appearance quality. The paper adds a restart refinement schedule and an appearance-rich prompting step to suppress artifacts, and reports the strongest results among the compared methods across seven condition types.","feed_headline":"Injecting condition features at a fixed middle step wins","feed_subtitle":"Decoupling injection from denoising balances structure and appearance without retraining.","key_machinery":"The load-bearing object is the injection schedule $g(t)$ used in the substitution where the generation branch's feature $f_{l,t}$ becomes $f_{l,g(t)}^{struct}$ and its attention map $A_{l,t}$ becomes $A_{l,g(t)}^{struct}$, decoupling when condition features are extracted from the denoising process. A restart refinement schedule then iteratively adds noise and re-denoises within a fixed interval to wash out artifacts, and appearance-rich prompting uses a multimodal language model to rewrite the prompt so the appearance image matches the condition image's semantics.","core_discovery":"On the paper's own terms, the central discovery is that the sampling schedule of condition features is a previously unexplored control variable, and the correct setting is asynchronous: inject features and attention maps extracted at a medium timestep, not the current denoising timestep. This is supported by measurements showing that structural self-similarity between condition and natural features improves as noise decreases while KL divergence between their distributions widens, so no single synchronized timestep can simultaneously maximize structure and domain alignment. The paper claims that any schedule whose values cluster around medium timesteps performs well regardless of functional form, and that the constant schedule $g(t)=600$ with caching is simple, effective, and computationally efficient.","pith_inferences":["The optimal constant $C$ is likely model- and schedule-dependent; a natural extension is to select it automatically per backbone or per condition modality instead of fixing it at 600.","If the middle-stage dominance reflects a general property of diffusion feature distributions, similar asynchronous schedules could improve other attention-based editing or style-transfer methods that currently inject synchronously.","A testable extension would be to replace the PCA and Gaussian-density proxy with full-dimensional feature statistics to see whether the same trade-off curve is recovered.","Because the prompt-rewriting step relies on a multimodal language model, its behavior on rare or abstract conditions may inherit that model's biases; a version using only visual cues would isolate the contribution of the schedule itself."],"forward_implications":["Injection-based training-free control gains a new tuning axis: the injection timestep can be chosen independently of the denoising timestep.","Because condition features are extracted once and cached, the decoupled schedule can be cheaper than synchronous injection while improving quality.","The framework can be dropped into existing methods such as FreeControl and into transformer-based backbones such as FLUX without retraining.","Multiple spatial conditions can be composed in a single generation, which the paper demonstrates with prompts and several condition images.","The paper claims its training-free method surpasses training-based controllable generation in structural fidelity on challenging conditions, including on transformer architectures."],"supporting_citations":[{"why":"Supplies the spatially-aware appearance transfer mechanism that RichControl builds on and serves as the primary training-free baseline.","marker":"[49]"},{"why":"Provides a training-free spatial control baseline that the framework plugs into and compares against.","marker":"[55]"},{"why":"Provides the plug-and-play feature injection baseline whose synchronous sampling schedule is analyzed and critiqued.","marker":"[82]"},{"why":"DDIM inversion and sampling are used to obtain noisy condition latents and to run the denoising process.","marker":"[74]"},{"why":"Inspires the restart refinement schedule that iteratively forward-and-backward denoises to remove artifacts.","marker":"[101]"},{"why":"Supplies the DINO self-similarity distance used to measure structural alignment in the analysis and evaluation.","marker":"[81]"},{"why":"Supplies the multimodal language model used for appearance-rich prompting.","marker":"[1]"},{"why":"Supplies the one-step attention extraction approach used for structure-rich injection on the FLUX backbone.","marker":"[48]"}],"fun_headline_variants":["Asynchronous condition injection fixes structure and appearance","Constant mid-step injection beats synchronized schedules","Decouple feature injection from denoising for better control","Training-free control: inject features at a fixed middle step","Simple schedule: same injection step for all diffusion steps"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The central analysis assumes that projecting diffusion features onto one PCA component and measuring KL divergence with Gaussian density estimates faithfully represents the domain gap that actually hurts generation quality; if that proxy is misleading, the schedule conclusions may not transfer.","fun_headline_variants_meta":{"raw":{"variants":["Asynchronous condition injection fixes structure and appearance","Constant mid-step injection beats synchronized schedules","Decouple feature injection from denoising for better control","Training-free control: inject features at a fixed middle step","Simple schedule: same injection step for all diffusion steps"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000442,"raw_usage":{"total_tokens":2236,"prompt_tokens":942,"completion_tokens":1294,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":558,"completion_tokens_details":{"reasoning_tokens":1221}},"tokens_in":558,"tokens_out":1294,"duration_ms":10180,"temperature":1.0,"reasoning_tokens":1221,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T20:19:31.444042+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A systematic grid over constant schedules $g(t)=C$ on the same SDXL setup that shows synchronous injection $g(t)=t$ matching or beating $C=600$ on both structural and appearance metrics, or a full-dimensional feature analysis that reverses the KL/self-similarity trade-off, would falsify the central claim.","supporting_citations":[{"cited_title":"Ctrl-x: Controlling structure and appear- ance for text-to-image generation without guidance","cited_arxiv_id":null,"evidence_quote":"Supplies the spatially-aware appearance transfer mechanism that RichControl builds on and serves as the primary training-free baseline."},{"cited_title":"Freecontrol: Training-free spatial control of any text-to-image diffusion model with any condition","cited_arxiv_id":null,"evidence_quote":"Provides a training-free spatial control baseline that the framework plugs into and compares against."},{"cited_title":"Plug-and-play diffusion features for text-driven image-to- image translation","cited_arxiv_id":null,"evidence_quote":"Provides the plug-and-play feature injection baseline whose synchronous sampling schedule is analyzed and critiqued."},{"cited_title":"De- noising diffusion implicit models","cited_arxiv_id":null,"evidence_quote":"DDIM inversion and sampling are used to obtain noisy condition latents and to run the denoising process."},{"cited_title":"Restart sampling for improving generative processes","cited_arxiv_id":null,"evidence_quote":"Inspires the restart refinement schedule that iteratively forward-and-backward denoises to remove artifacts."},{"cited_title":"Splicing ViT features for semantic appearance transfer","cited_arxiv_id":null,"evidence_quote":"Supplies the DINO self-similarity distance used to measure structural alignment in the analysis and evaluation."},{"cited_title":"Freecontrol: Efficient, training-free structural con- trol via one-step attention extraction","cited_arxiv_id":null,"evidence_quote":"Supplies the one-step attention extraction approach used for structure-rich injection on the FLUX backbone."}],"review_version":1}