{"id":"a47c914d-beb3-466b-a32b-545681a0009f","arxiv_id":"2501.12910","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"PreciseCam enables precise camera control (roll, pitch, vFoV, distortion) in text-to-image generation by conditioning SDXL with Perspective Field maps and a new dataset of 57,380 images.","lead":"The paper trains a ControlNet adapter for Stable Diffusion XL that conditions image generation on per-pixel maps derived from four camera parameters: roll, pitch, vertical field of view, and distortion. This gives users slider-based camera control over generated images without 3D scene geometry, and the authors release a 57,380-image dataset with ground-truth camera settings.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central claim of 'precise' camera control rests on qualitative examples only; no quantitative measurement verifies that generated images match the specified roll, pitch, vFoV, and ξ. This is the load-bearing gap.","rationale":"The reader identified the absence of quantitative camera-accuracy evaluation as the weakest assumption. I agree. I read the paper in good faith: the method is clearly described, the PF-US representation is mathematically sound, the dataset construction is reasonable, and the qualitative results are suggestive. However, the central word 'precise' is a quantitative claim, and the paper supplies no quantitative evidence for it. CLIP/BLIP scores only show that adding control does not degrade prompt adherence; they say nothing about whether the camera geometry is correct. The paper's own limitation section concedes failures at extreme roll and under prompt-camera conflicts, which further indicates that precision is conditional. The dataset-transfer premise is the underlying cause: training on 360° crops may teach the model a mapping from PF-US to image appearance, but whether that mapping is geometrically exact for out-of-distribution content is unverified. This is not an internal inconsistency; it is an unsupported central claim. A quantitative camera-calibration evaluation would settle it. Since the reader's CONDITIONAL verdict already asks for exactly this, I recommend no change. The concrete test above is a feasible way to move from conditional to accept/reject.","tokens_in":15502,"tokens_out":4340,"duration_ms":50809,"concrete_test":"Generate a grid of 100+ images with PreciseCam over parameter combinations (roll ∈ {-60,-30,0,30,60}, pitch ∈ {-60,-30,0,30,60}, vFoV ∈ {20,50,80,120}, ξ ∈ {0.1,0.5,0.9}) using diverse photorealistic and artistic prompts. Run a single-image camera calibration method (e.g., GeoCalib [41] or the Perspective-Fields estimator [24], after validating it on synthetic renders with known cameras) on each generated image. Report median absolute error and correlation between specified and estimated roll, pitch, vFoV, and ξ. If median roll error exceeds ~5°, pitch error exceeds ~5°, or vFoV relative error exceeds ~10%, the 'precise' claim is overstated; if errors are small, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that PreciseCam provides 'precise camera control' over roll, pitch, vFoV, and ξ (Abstract; Sec. 4). The only evidence is qualitative image series (Figs. 5-7) and prompt-adherence scores (CLIP/BLIP, Fig. 8), which do not measure camera geometry. Since generated images have no ground-truth camera parameters, the paper never tests whether a requested roll=30° actually yields a 30° roll, whether vFoV=20° vs 50° changes the field of view by the specified factor, or whether ξ=0.1 vs 0.9 produces the correct distortion profile. The PF-US map is a deterministic function of the four parameters, so the training signal is self-consistent; but the model could learn to approximate the qualitative look of a parameter (e.g., 'wide-angle feel') without achieving the quantitative values. That distinction is exactly what 'precise' claims. The dataset-transfer premise (Sec. 3.3) — that crops from 360° panoramas teach the model to produce geometrically correct views for arbitrary prompts, including artistic content absent from training — is plausible but unverified. The acknowledged limitations (Sec. 5: extreme roll, prompt-camera conflicts) show precision is not uniform. Without a camera-accuracy metric, the headline claim is underdetermined.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"PreciseCam introduces a ControlNet-based conditioning module for SDXL that takes four camera parameters — roll, pitch, vertical field of view (vFoV), and distortion ξ — and encodes them as Perspective Field maps under the Unified Spherical camera model (PF-US). The maps are computed deterministically from the parameters via Eq. (3) and fed into a ControlNet. The authors contribute a dataset of 57,380 images with ground-truth camera parameters, obtained by sampling and cropping 360° panoramas, with BLIP-2 generated prompts. The paper presents qualitative results varying each parameter, comparisons with prompt-engineered SDXL and Adobe Firefly, CLIP/BLIP prompt-adherence scores, an ablation of residual injection locations, and proof-of-concept applications to background generation, video control, and multi-ControlNet setups. The central claim is that PreciseCam provides precise, continuous camera control while preserving prompt adherence.","tokens_in":15713,"tokens_out":5851,"duration_ms":54997,"significance":"The contribution is timely and the formulation is clean: using PF-US as a lightweight camera conditioning representation avoids 3D reconstruction and multi-view data. The dataset is a valuable resource, and the code and model are promised publicly. The ablation identifying bottleneck-only residual injection is a useful practical finding. However, the headline claim of 'precise' camera control rests entirely on qualitative image inspection and prompt-adherence metrics; no measurement establishes that generated images match the specified camera parameters. If the authors add a quantitative camera-accuracy evaluation, the paper would be strong.","major_comments":[{"comment":"The central claim of 'precise camera control' is not supported by any quantitative measurement of camera geometry. Figures 5–7 are qualitative, and the CLIP/BLIP scores in Fig. 8 measure prompt adherence only; they do not test whether a requested roll=30° actually produces a 30° roll, whether vFoV=20° vs 50° changes the field of view by the specified amount, or whether ξ=0.1 vs 0.9 yields the correct distortion profile. Because the PF-US map is a deterministic function of the four parameters, the training signal is self-consistent, but the model could learn a qualitative approximation of each parameter without achieving the numeric value, which is exactly what 'precise' claims. Please add a quantitative camera-accuracy metric, for example by running a single-image calibrator such as Jin et al. [24] or GeoCalib [41] on generated images, or by generating images from 3D scenes with known camera parameters and measuring the error.","section":"Sec. 4, Figs. 5–8"},{"comment":"The comparison to prompt-engineered SDXL and Adobe Firefly does not quantify camera-control accuracy. Fig. 8 reports only CLIP and BLIP scores, which establish that prompt adherence is preserved but cannot distinguish whether the desired camera view was achieved. To support the Abstract claim that PreciseCam 'surpasses traditional prompt engineering approaches,' the paper should report a camera-parameter error metric or a forced-choice human study across methods, so that camera-control success is measured directly.","section":"Sec. 4, Comparisons and Fig. 7"},{"comment":"The conversion from (roll, pitch, vFoV, ξ) to PF-US is not fully specified. Eq. (3) uses focal length f and distortion ξ, but the paper never states how the vertical field of view is converted into f, nor what image height is assumed, nor how ξ modifies that relationship in the Unified Spherical model. Without this formula, the dataset generation and the conditioning maps are not reproducible, and the mapping from user-facing camera parameters to the learning signal is ambiguous. Please provide the explicit vFoV-to-f relation for the US model, including the image height used.","section":"Sec. 3.1, Eq. (3); Sec. 3.3"},{"comment":"The dataset-transfer premise — that crops from 360° panoramas teach the model to produce geometrically correct views for arbitrary prompts, including artistic content largely absent from training — is plausible but unverified, and the limitations in Sec. 5 show that precision is not uniform (extreme roll, prompt-camera conflicts). The requested quantitative evaluation should be stratified by prompt style (photographic vs artistic) and by parameter range, so that the reader can see where 'precise' holds and where the acknowledged failure modes emerge.","section":"Sec. 5 and Sec. 3.3"}],"minor_comments":[{"comment":"In the parameter sampling description, 'two ξ values (low ∈ [15, 60) and high ∈ [60, 140))' appears to be a typo: these ranges are in degrees and describe vFoV, while ξ lies in (0,1). Please correct.","section":"Supplementary Sec. B"},{"comment":"The caption states 'Top row: Variation in pitch' but Sec. 4 and the row labels indicate that the top row varies roll; please align the caption with the figure and text.","section":"Fig. 5 caption"},{"comment":"The sentence 'Our model, PreciseCam, archives precise camera view control' should read 'achieves.'","section":"Sec. 1"},{"comment":"The phrase 'perfectioning novel tools' should be 'perfecting.'","section":"Sec. 5"},{"comment":"The 'Camera conditioning adherence' subsection refers the reader to the supplementary material for examples but provides no quantitative measure of stability across noise; adding a small statistic, such as variation of estimated camera parameters across noise seeds, would strengthen this claim.","section":"Sec. 4.1"}],"recommendation":"major_revision","confidential_remarks":"This is a well-executed system paper with a valuable dataset. The reader's conditional verdict is appropriate: the central concern — absence of a quantitative camera-accuracy evaluation — is the main barrier to acceptance. I see no circularity or authorship concerns. I would encourage the editor to treat the requested evaluation as a required revision rather than optional."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core idea is neat: encode roll, pitch, vFoV, and distortion into PF-US maps and feed them to ControlNet. That combination is new, and it avoids the baggage of 3D representations. The 57k-image dataset built from 360° crops with ground-truth camera parameters is a real contribution—I can see myself using it as a benchmark or training source. The ablation showing that bottleneck-only residual injection works best is genuinely informative, and the applications (background matching, video frame anchoring, multi-ControlNet) are sensible proof-of-concept extensions. The paper is also honest about its limitations, which I appreciate.\n\nBut the central claim of 'precise' camera control is not actually demonstrated. The evaluation consists of qualitative image grids and CLIP/BLIP prompt-adherence scores, neither of which measures whether a requested 30° roll produces a 30° roll, whether vFoV 20° vs 50° changes the geometry by the specified factor, or whether ξ=0.1 vs 0.9 gives the right distortion profile. The stress-test note is right: the conditioning map is a deterministic function of the four parameters, so training is self-consistent, but the model could easily be approximating a qualitative look rather than hitting the numeric values. This is exactly the distinction the word 'precise' promises. The dataset-transfer premise—that crops from 360° panoramas teach the model to produce geometrically correct views for arbitrary artistic prompts—is plausible but also unverified.\n\nWhat's missing is straightforward to supply: run an off-the-shelf camera calibration method (Jin et al.'s PF estimator, GeoCalib, or even a simple horizon/vanishing-point detector) on generated images and report roll, pitch, vFoV, and ξ errors. The authors cite Jin et al. for PF estimation and even reject it for their training data, so using it as an eval metric would be consistent. Without that, the paper is a good system paper with an unproven headline claim.\n\nFor a serious referee, I'd send it out—the idea and dataset deserve engagement—but I'd expect a major revision adding quantitative camera-accuracy evaluation. As it stands, the contribution is real but the 'precise' claim overreaches the evidence.","headline":"Clean conditioning recipe and a useful dataset, but the paper never measures the camera accuracy it claims, so 'precise' remains an assertion.","tokens_in":645,"tokens_out":670,"would_cite":true,"duration_ms":23824,"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":"Four camera parameters, encoded as per-pixel maps, give text-to-image models precise camera control.","keywords":["camera control","text-to-image generation","diffusion models","ControlNet","Perspective Field","Unified Spherical camera model","panorama dataset","lens distortion"],"falsifier":"Generate a set of images with known parameter quartets, then run an off-the-shelf single-image camera calibrator (for example, a Perspective-Field estimator) on each output and plot estimated versus requested roll, pitch, vertical field of view, and $\\xi$; if the estimates do not track the requested values, or the error is no better than prompt-engineered SDXL, the precision claim is falsified.","tokens_in":15287,"feed_emoji":"🎥","tokens_out":4413,"duration_ms":42043,"temperature":0.7,"pith_summary":"The paper claims that text-to-image diffusion models can be given precise camera control by conditioning them on a per-pixel geometric map derived from just four user-set parameters: roll, pitch, vertical field of view, and lens distortion. The authors build this map, called a PF-US map, by combining the Perspective Field representation with the Unified Spherical camera model, and train a ControlNet adapter for SDXL on a new dataset of 57,380 images cropped from 360-degree panoramas with ground-truth camera parameters. If the claim holds, artists and designers could specify camera language numerically, such as a Dutch angle, a worm's-eye view, or a fisheye lens, instead of coaxing the model through prompt engineering, while keeping the content fully described by the text prompt. The paper also shows the conditioning surviving artistic styles, noise variation, and prompt variation, plus applications to object-compositing backgrounds and frame-by-frame video camera control.","feed_headline":"Four sliders put precise camera control into AI images","feed_subtitle":"Roll, pitch, field of view, and lens distortion become exact generation settings instead of prompt guesswork.","key_machinery":"The PF-US map is a two-channel image in which each pixel stores the up-vector (opposed to gravity) and the latitude angle of its viewing ray, computed through the Unified Spherical camera projection function $P(X)$, whose distortion parameter $\\xi$ interpolates from pinhole to omnidirectional optics. These maps convert the four abstract camera parameters into local per-pixel appearance cues, so a ControlNet can learn the mapping without any 3D scene representation or multi-view data. At inference, only the ControlNet middle-block residuals are injected into SDXL, which the paper finds gives the best trade-off between camera adherence and image quality.","core_discovery":"PreciseCam's central claim is that four camera parameters, roll, pitch, vertical field of view, and distortion $\\xi$, fully determine the geometric appearance of a generated image once they are encoded, pixel by pixel, as up-vectors and latitude angles (the Perspective Field) under the Unified Spherical camera model. A ControlNet trained on these PF-US maps can steer Stable Diffusion XL to produce images whose camera geometry matches the specified settings, while prompt adherence, measured by CLIP and BLIP scores, stays comparable to the base model. The paper demonstrates systematic variation of each parameter with other settings fixed, showing consistent adherence across photographic and artistic outputs, and argues this exceeds what prompt engineering or preset style tags can achieve.","pith_inferences":["Because the conditioning is a purely 2D geometric map, the same PF-US interface could be plugged into other UNet diffusion backbones, or applied to image editing and inpainting, without retraining the base model; this transfer is plausible but not tested in the paper.","The paper's precision claim is qualitative; a natural extension is a quantitative benchmark that estimates roll, pitch, vertical field of view, and $\\xi$ from generated images and reports error against the requested values, which would reveal which parameter ranges are actually reliable.","Yaw is deliberately excluded because a single image has no left-right reference; a testable consequence is that conditioning with an explicit scene reference, such as a text-specified landmark direction, might enable full six-degree-of-freedom control from the same map representation.","The panorama-cropping dataset construction could be reused to generate camera-labeled data for other conditioning tasks, such as depth or normal estimation under extreme fields of view, where real paired data is scarce."],"forward_implications":["A user can specify roll, pitch, vertical field of view, and $\\xi$ with sliders and get generated images that visually follow those settings, with content decided by the text prompt.","The same camera conditioning holds across different noise seeds and small prompt variations, meaning the control is independent of content.","Artistic styles such as paintings, ink wash, and lithograph follow the camera conditioning even though the training crops come from photographic panoramas.","The method can generate object-matched backgrounds for compositing and can condition each frame of a video, or anchor the initial frame for relative camera-control video models.","PreciseCam composes with other ControlNets, so pose, depth, or edge control can be combined with camera control in one generation."],"supporting_citations":[{"why":"Supplies the Perspective Field representation of per-pixel up-vectors and latitude angles that PreciseCam adopts and adapts.","marker":"[24]"},{"why":"Defines the Unified Spherical camera model used to encode roll, pitch, vertical field of view, and distortion $\\xi$ into the PF maps.","marker":"[5]"},{"why":"Provides the ControlNet conditioning architecture that PreciseCam trains to inject camera guidance into the frozen SDXL model.","marker":"[52]"},{"why":"The SDXL base text-to-image model whose encoder and middle-block layers are duplicated for the ControlNet setup.","marker":"[36]"},{"why":"BLIP-2 generates the descriptive text prompt for each dataset image, decoupling prompt learning from camera-view learning during training.","marker":"[29]"},{"why":"Supports the inference-time choice of injecting only middle-block residuals, cited by the paper for keeping conditioning adherence without degrading image quality.","marker":"[6]"}],"fun_headline_variants":["Four numbers now control camera in AI images","PreciseCam: exact lens angles without prompt guessing","Four camera parameters now generate any angle","Dial in roll, pitch, zoom, and distortion for AI art","PreciseCam: four sliders beat prompt engineering"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The premise is that patches cropped from 360-degree panoramas, with their corresponding PF-US maps, are representative training material for what a real or artistic image looks like under a given camera setting, so that a model trained on them will reproduce the geometry in generated images; the paper never measures camera accuracy on the outputs.","fun_headline_variants_meta":{"raw":{"variants":["Four numbers now control camera in AI images","PreciseCam: exact lens angles without prompt guessing","Four camera parameters now generate any angle","Dial in roll, pitch, zoom, and distortion for AI art","PreciseCam: four sliders beat prompt engineering"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000658,"raw_usage":{"total_tokens":2951,"prompt_tokens":825,"completion_tokens":2126,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":441,"completion_tokens_details":{"reasoning_tokens":2064}},"tokens_in":441,"tokens_out":2126,"duration_ms":13814,"temperature":1.0,"reasoning_tokens":2064,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T16:37:57.706770+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Generate a set of images with known parameter quartets, then run an off-the-shelf single-image camera calibrator (for example, a Perspective-Field estimator) on each output and plot estimated versus requested roll, pitch, vertical field of view, and $\\xi$; if the estimates do not track the requested values, or the error is no better than prompt-engineered SDXL, the precision claim is falsified.","supporting_citations":[{"cited_title":"Perspective fields for single image camera calibration","cited_arxiv_id":null,"evidence_quote":"Supplies the Perspective Field representation of per-pixel up-vectors and latitude angles that PreciseCam adopts and adapts."},{"cited_title":"A unifying geometric representation for cen- tral projection systems","cited_arxiv_id":null,"evidence_quote":"Defines the Unified Spherical camera model used to encode roll, pitch, vertical field of view, and distortion $\\xi$ into the PF maps."},{"cited_title":"Adding conditional control to text-to-image diffusion models","cited_arxiv_id":null,"evidence_quote":"Provides the ControlNet conditioning architecture that PreciseCam trains to inject camera guidance into the frozen SDXL model."},{"cited_title":"Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models","cited_arxiv_id":null,"evidence_quote":"BLIP-2 generates the descriptive text prompt for each dataset image, decoupling prompt learning from camera-view learning during training."},{"cited_title":"Loosec- ontrol: Lifting controlnet for generalized depth conditioning","cited_arxiv_id":null,"evidence_quote":"Supports the inference-time choice of injecting only middle-block residuals, cited by the paper for keeping conditioning adherence without degrading image quality."}],"review_version":1}