{"id":"4ea4b497-fc2e-4160-9fad-dfe795b8b563","arxiv_id":"2607.06553","paper_version":2,"verdict":"ACCEPT","confidence":"HIGH","novelty_score":7.0,"correctness_risk":"low","formal_verification":"none","parameter_count":3,"one_line_summary":"A frozen DiT plus task LoRA and a ~33K-parameter token-local linear head reads out pixel-space dense fields and sets SOTA on matting, KITTI depth, and referring segmentation while running up to 2.48\times faster than edit-plus-decode baselines.","lead":"Text-to-image DiTs can solve dense prediction by reading task values straight from their patch tokens instead of generating image-like targets through a VAE. The simpler readout is faster and often more accurate, showing generative priors help perception without the generative output path.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"No significant objection identified","rationale":"The strongest claim is empirical and interface-focused: generative pretraining supplies an RGB-native spatial lattice that can be read out directly once lightly adapted, without inheriting the target-side generative decoder. Sec. 3.3 and the diagnostic ablations in Table 5 directly test the critical premise that the adapted token field already carries the needed spatial structure. The paper does not overclaim uniqueness of generative pretraining (explicitly notes the contribution is the output interface) and reports both wins and competitive results honestly. The reader's weakest_assumption is the right soft spot, yet it is already stress-tested inside the paper; no stronger load-bearing concern (e.g., hidden spatial mixing, evaluation protocol mismatch, or circular use of the VAE) emerges from the manuscript. Therefore the ACCEPT / high-confidence verdict stands without adjustment.","tokens_in":14920,"tokens_out":534,"duration_ms":5901,"concrete_test":"Reproduce the Table 5 '4-stage conv head (425K)' and 'Edit paradigm' rows on the same FLUX-Klein 4B checkpoint, LoRA rank, and 512^{2} resolution for NYU normals and P3M-500-P/NP; if either exceeds ReChannel-4B accuracy by a clear margin or the latency gap shrinks below ~1.5\times, the sufficiency of the token-local linear interface would be weakened.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that a frozen DiT plus task LoRA plus a shared token-local linear head (no spatial mixing, ~33K params) is a sufficient and superior output interface for dense prediction relative to target-side VAE/edit decoding. The reader's weakest assumption correctly flags the generality of pure locality, but Table 5 already isolates that axis under a matched 4B backbone: a 13\times larger 4-stage conv head, full fine-tuning, latent targets, VAE-frozen pixel supervision, and the edit paradigm are all equal or worse on normals and matting, while head-only and random-init collapse. Main results further show the same head form works for continuous geometry, high-frequency alpha, multi-channel pose heatmaps, and language-conditioned masks. No internal inconsistency, circularity, or untested premise that would overturn the reported SOTA and 2.48\times speed claims is present; remaining limits (backbone family, resolution, non-pixel-aligned targets) are ordinary empirical scope, not load-bearing flaws.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The paper argues that dense prediction should not inherit the target-side VAE/edit interface of text-to-image models. A pretrained DiT already organizes RGB inputs on a patch\to token\to patch lattice, so each token can be read out as a fixed p\times p\times K_t task-native pixel patch rather than RGB appearance. ReChannel keeps the VAE encoder for the DiT input distribution, freezes the backbone, adapts with task LoRA, and uses a shared token-local linear head (~33K parameters, no spatial mixing). On FLUX-Klein (4B/9B) this recipe is evaluated on six dense tasks and a dozen-plus benchmarks, claiming new SOTA on trimap-free matting, KITTI depth, and referring segmentation, competitive results on normals/saliency/pose, and 2.48\times lower latency than a matched edit-plus-latent-decode baseline under the same 4B backbone.","tokens_in":15231,"tokens_out":684,"duration_ms":6454,"significance":"If the results hold, the work cleanly separates generative pretraining priors from the generative output interface and shows that a minimal token-local linear readout is sufficient for heterogeneous dense targets. The matched 4B ablations (Table 5) isolate LoRA necessity, pretraining necessity, head capacity, and generative-interface alternatives, while main tables report multi-benchmark gains and direct latency. This is a practical, falsifiable contribution for the growing line of generative dense predictors and is supported by public-benchmark protocols and a public code link. Strengths include the multi-task scope, the controlled interface ablations, and the explicit efficiency comparison under a fixed backbone.","major_comments":[],"minor_comments":[{"comment":"Fig. 2 participation-ratio values are labeled “illustrative”; a short appendix note on how PR is computed (token covariance, layer, after/before LoRA) would make the diagnostic fully reproducible.","section":null},{"comment":"Sec. 3.3 / Eq. (2): state the concrete p and token dimension used for FLUX-Klein so the ~33K parameter count can be verified without external lookup.","section":null},{"comment":"Table 1: ReChannel-9B trails Edit2Perc on NYU absRel (0.051 vs 0.044); a one-sentence discussion of this split would balance the SOTA claims on KITTI/ScanNet.","section":null},{"comment":"Sec. 4.1: briefly list LoRA rank, learning rate, and epochs (or point to a config) so the matched-4B ablations can be re-run exactly.","section":null},{"comment":"Conclusion already notes scope limits (backbone family, non-pixel-aligned targets); a short remark on resolution scaling of the pure token-local head would further clarify generality.","section":null}],"recommendation":"accept","confidential_remarks":"The central claim is empirically well supported and the interface ablations are clean. I see no load-bearing technical flaw that would require major revision; remaining issues are presentation and ordinary empirical scope. Fit for a strong CV venue is good."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The useful takeaway is simple: generative dense predictors have been carrying unnecessary target-side VAE/edit machinery. Once a frozen DiT is lightly adapted with task LoRA, a shared token-local linear head (~33K params, no spatial mixing) that maps each token to its p\times p\times K_t pixel patch is enough, and often better and faster.\n\nWhat is new is the clean separation of the RGB-native field organizer from the generative output interface, plus the empirical demonstration that pure locality works across six heterogeneous tasks. They keep the input VAE for distribution match, drop the target decoder entirely, and supervise in pixel space. Tables 1–4 show new SOTA on trimap-free matting (P3M and zero-shot AIM-500), KITTI depth, and the full RefCOCO suite, with competitive normals, saliency, and pose. The matched 4B ablations in Table 5 are the real backbone of the claim: head-only and random-init collapse, a 13\times larger conv head and full fine-tuning do not beat the thin linear readout, and latent / VAE-frozen / edit paths are both less accurate and up to 2.48\times slower. Latency numbers match the design (no decoder at all). Citations cover the right generative-for-perception line (Marigold, Lotus, GenPercept, Edit2Perc) without padding.\n\nSoft spots are ordinary scope, not load-bearing holes. Everything is on FLUX-Klein; we do not yet know how the lattice behaves on other DiT families or at higher resolution. The “token field already carries the structure” premise is tested only on the reported tasks and resolutions—if it fails for some non-pixel-aligned or multi-scale targets, the minimal head would underperform—but Table 5 already isolates that axis and the multi-task results make the claim credible rather than hand-wavy. No circularity, no invented math, free parameters are just the usual LoRA rank and the tiny head.\n\nThis is for people who actually adapt T2I models for depth, matting, or referring segmentation and for anyone writing the next “generative priors for perception” paper. It deserves a serious referee. I would engage with it and expect it to influence practice.","headline":"Clean, well-supported interface paper: drop the target-side VAE and read dense fields linearly from a DiT lattice; SOTA and speed gains look real under matched ablations.","tokens_in":15787,"tokens_out":559,"would_cite":true,"duration_ms":6616,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"Dense prediction can reuse a text-to-image model by reading task fields from its token lattice, not by generating image-like targets.","keywords":["dense prediction","text-to-image models","DiT","ReChannel","pixel-space readout","generative priors","LoRA","matting"],"falsifier":"On a matched backbone and resolution, if a larger spatial decoder or a target-side VAE or edit path clearly beat the token-local linear head on high-frequency matting or multi-channel pose heatmaps, the claim that the adapted token field already carries the output would fail.","tokens_in":15853,"feed_emoji":"🖼️","tokens_out":1034,"duration_ms":21799,"temperature":0.7,"pith_summary":"Large text-to-image models already turn an RGB image into a patch-aligned grid of tokens. The authors argue that dense prediction does not need the rest of the generative pipeline: you do not have to encode depth, mattes, masks, or heatmaps into an RGB-trained latent space and decode them back as if they were new images. Their method, ReChannel, keeps the input encoder, freezes the transformer, adapts it with lightweight task LoRA, and maps each token to its local pixel patch with a shared linear head of roughly 33K parameters and no spatial mixing. Across six dense tasks and more than a dozen benchmarks this minimal readout sets new state of the art on trimap-free matting, KITTI depth, and referring segmentation, stays competitive elsewhere, and in a matched setting is both more accurate and 2.48 times faster than an edit-plus-latent-decode counterpart. A reader who cares about perception backbones gets a concrete claim: generative pretraining is useful for the field it organizes, not for the rendering interface it was trained to use.","feed_headline":"Skip the decoder: read dense fields from image tokens","feed_subtitle":"A tiny linear head on a frozen text-to-image model beats generative dense predictors and runs 2.48x faster.","key_machinery":"ReChannel: after task LoRA adapts the frozen DiT, a shared token-local linear head maps each spatial token to its fixed p by p by K_t pixel patch, reinterpreting the token's channels from RGB appearance to task-native fields with no target-side VAE and no inter-token mixing in the head.","core_discovery":"Dense perception can benefit from generative pretraining without inheriting its output interface. A pretrained DiT already organizes RGB inputs through a patch-to-token-to-patch lattice on the image plane, so each token indexes a fixed output patch whose channels can carry task-native quantities instead of RGB appearance. ReChannel therefore keeps the VAE encoder for the DiT's input distribution, drops the target-side decoder, adapts the frozen DiT with task LoRA, and reads each token to its p by p by K_t pixel-space patch through a shared token-local linear head.","pith_inferences":["If adapted token lattices already carry spatial structure, many decoder-heavy dense pipelines may be oversized once a strong generative prior is in place.","The same rechanneling pattern should transfer to other patch-token generative backbones, not only the family evaluated here.","Non-image-like but still pixel-aligned fields such as optical flow or multi-frame disparity stacks are natural next targets for the same linear readout form.","Discriminative foundation encoders might also admit similarly minimal readouts if light adaptation collapses their token fields into compact task subspaces."],"forward_implications":["Dense targets need not be forced through an RGB-trained VAE encode-decode path to reuse a text-to-image prior.","One frozen T2I backbone plus task LoRA and a shared linear readout can cover continuous geometry, alpha mattes, language-conditioned masks, and multi-channel heatmaps.","Removing the target-side generative path can improve both accuracy and latency (up to 2.48 times faster) versus edit-plus-latent-decode.","Extra head capacity or full fine-tuning need not beat a token-local linear readout once the token field is task-adapted.","New state-of-the-art results on trimap-free matting, KITTI depth, and referring segmentation are reachable with this interface."],"fun_headline_variants":["Skip the decoder: map DiT tokens to dense pixel fields","Read task-native maps from frozen image tokens","Token lattice to dense fields without RGB decode","Tiny linear head turns DiT tokens into dense predictors","Drop target decoder: dense readout from image tokens"],"cache_read_input_tokens":128,"weakest_assumption_plain":"Once the transformer is lightly adapted for a task, its token grid already holds the spatial structure needed, so a purely local linear map with no mixing between tokens is enough for every dense target tried.","fun_headline_variants_meta":{"raw":{"variants":["Skip the decoder: map DiT tokens to dense pixel fields","Read task-native maps from frozen image tokens","Token lattice to dense fields without RGB decode","Tiny linear head turns DiT tokens into dense predictors","Drop target decoder: dense readout from image tokens"]},"model":"grok-4.5","effort":"low","cost_usd":0.00307,"raw_usage":{"total_tokens":1201,"prompt_tokens":938,"num_sources_used":0,"completion_tokens":57,"cost_in_usd_ticks":30700000,"prompt_tokens_details":{"text_tokens":938,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":206,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":938,"tokens_out":57,"duration_ms":3280,"temperature":1.0,"reasoning_tokens":206,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-11T00:01:13.758546+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"On a matched backbone and resolution, if a larger spatial decoder or a target-side VAE or edit path clearly beat the token-local linear head on high-frequency matting or multi-channel pose heatmaps, the claim that the adapted token field already carries the output would fail.","supporting_citations":[],"review_version":2}