{"id":"526d19de-e447-4e44-9f1d-677666bc78aa","arxiv_id":"2504.12395","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"InstantCharacter is a diffusion-transformer adapter and three-stage training recipe that generates text-editable images of open-domain characters while preserving identity.","lead":"InstantCharacter adds a character-personalization adapter to a large diffusion transformer image model: one photo of a character plus a text instruction produces new images of that character in different poses, scenes, or styles. The paper reports high-fidelity and identity-consistent results on a 10-million-image training set, with qualitative comparisons against several existing systems.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'consistently performs the best' claim rests solely on qualitative figures; the paper promises quantitative measurements in Fig. 6, but Fig. 6 is captioned 'More qualitative results' and contains no metrics.","rationale":"The reader's conditional verdict already calls for quantitative evaluation, ablations, and dataset release. My stress test confirms that need and identifies a sharper, more load-bearing problem: the manuscript's own text asserts that quantitative support exists, but the cited Fig. 6 is qualitative, so the central comparative claim is unsupported in the current version. The reader's weakest assumption about stage decomposition is valid but secondary; even if the stages are orthogonal, the paper gives no measurement of the final system. I keep the verdict UNCHANGED because a conditional accept is the appropriate disposition for a systems paper whose architecture is plausible but whose central performance claim is not yet evidenced. An outright reject would be too strong, since the missing metrics could be supplied with code and dataset release, and no internal inconsistency in the design is apparent. However, if the promised quantitative evaluation never materializes, the paper should not be accepted on the strength of the current qualitative comparisons.","tokens_in":5781,"tokens_out":4069,"duration_ms":42016,"concrete_test":"Run the released code and dataset on a fixed benchmark of at least 50 unseen characters with 200 hand-written prompts; compute identity consistency (e.g., mean cosine similarity of DINOv2 or face embeddings between reference and outputs) and text alignment (e.g., CLIP score), and compare these numbers against OminiControl, EasyControl, ACE+, and UNO on identical prompts. Also run a stage ablation: train the full three-stage schedule, skip stage two, and reverse stages one and two; if the final identity and editability metrics change in a material direction, the claimed clean separation in Section 3.2 is false. If code or data are not released, the quantitative table promised by the text should be supplied and its collection protocol specified.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim of the paper is that a single frozen-FLUX adapter with stacked encoders and the three-stage schedule achieves open-domain identity preservation together with text editability. Section 4 claims InstantCharacter 'consistently performs the best' against OminiControl, EasyControl, ACE+, UNO, and GPT-4o, but the only evidence offered is a small set of qualitative grids (Figs. 3-5). The sentence 'These qualitative advantages are further supported by quantitative measurements shown in Fig. 6' is not backed up: Fig. 6 is captioned 'More qualitative results of InstantCharacter' and contains no numeric evaluation. No identity-similarity metric, no text-alignment score, no user study, and no statistical test are reported anywhere. In addition, Section 3.2 states that the three training stages decouple character consistency, textual controllability, and image fidelity and that stage two 'efficiently eliminates the copy-paste effect,' yet no ablation supports this decomposition. If the stages interact, or if the qualitative examples are cherry-picked, the reported simultaneous identity, editability, and generalization would not reproduce.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"InstantCharacter proposes a scalable adapter for the frozen FLUX.1-dev diffusion transformer, aimed at open-domain character personalization. The adapter combines SigLIP and DINOv2 features, adds dual-stream low-level and region-level fusion through intermediate transformer encoders, and projects the result into the denoising space with a timestep-aware Q-former. Training uses a claimed 10-million-image dataset split into unpaired self-reconstruction data and paired text-image editing data, organized into a three-stage schedule at 512 resolution and then high resolution. The paper reports qualitative comparisons against OminiControl, EasyControl, ACE+, UNO, and GPT-4o, and claims that InstantCharacter 'consistently performs the best' in fidelity, identity preservation, and text controllability.","tokens_in":6126,"tokens_out":2122,"duration_ms":23396,"significance":"If the claims are correct, the paper would make a useful contribution: it demonstrates an adapter architecture that scales with modern DiT backbones, introduces a large-scale dual-data training recipe for identity and editability, and releases code. The proposed design choices (SigLIP+DINOv2 fusion, patch-level region features, three-stage training) are plausible and worth reporting. However, the evidence provided is exclusively qualitative. The comparison grids in Figures 3-5 are selected examples without quantitative metrics, user studies, or ablations, so the central claims of state-of-the-art performance and of the three-stage decomposition are not yet supported. The paper's value is therefore conditional on adding rigorous evaluation.","major_comments":[{"comment":"The sentence 'These qualitative advantages are further supported by quantitative measurements shown in Fig. 6' is contradicted by the manuscript itself: Figure 6 is captioned 'More qualitative results of InstantCharacter' and contains no numeric measurements, error bars, or statistical tests. The claim that InstantCharacter 'consistently performs the best' is therefore supported only by a small set of qualitative grids. Please add actual quantitative evaluation (e.g., identity similarity metrics, text-image alignment scores, user preference studies) or remove the unsupported sentence.","section":"§4, first paragraph"},{"comment":"The paper asserts that the three stages decouple character consistency, textual controllability, and image fidelity, and that the paired training stage 'efficiently eliminates the copy-paste effect,' but no ablation or metric is provided to show that the stages compose without interference or that the ordering matters. If the staged objectives interact, the reported simultaneous identity fidelity, editability, and generalization may not reproduce. Please include stage-wise ablations and a sensitivity analysis of the training schedule.","section":"§3.2, three-stage training"},{"comment":"The evaluation set is described only as 'a set of open-domain character images not present in the training data,' with no information about its size, selection criteria, or overlap checking procedure. Because the training dataset is not released, this claim cannot be verified. Please specify the evaluation set construction, the number of prompts and images, the baselines' configuration details, and the exact comparison protocol so that the qualitative comparison is reproducible.","section":"§4, evaluation protocol"},{"comment":"The comparison against GPT-4o, a closed and non-reproducible system, and the absence of any quantitative comparison raise fairness concerns. The paper does not state the random seeds, prompt templates, or post-processing for any method. The 'consistently performs the best' claim needs a defined protocol, ideally with multiple independent raters or a scoring rubric, to move beyond selected examples.","section":"§4, comparison fairness"}],"minor_comments":[{"comment":"The baseline name is spelled 'OminiControl' in the text but 'Ominicontrol' in reference [20]; please standardize the spelling.","section":"§1 and references"},{"comment":"The base model is referred to as both 'FLUX1.0-dev' in the introduction and 'FLUX.1' elsewhere; please use a consistent name.","section":"§2"},{"comment":"The dataset size is described as '10-million-level'; please state the exact number of images in the unpaired and paired subsets and describe the data curation and filtering process.","section":"§3.2"},{"comment":"The adapter's capacity, number of parameters, and training hyperparameters (learning rate, batch size, optimizer, number of iterations per stage) are not reported; these details are needed to assess the scalability claim.","section":"§3.1"},{"comment":"The caption says 'progressively' but likely should read 'progressive' or 'progressively refines'; please correct the typo.","section":"Figure 2"}],"recommendation":"major_revision","confidential_remarks":"The main issue is evidentiary rather than technical: the paper's headline claims are qualitative, while the text promises quantitative support that does not exist. This is fixable with additional experiments, but as submitted the central performance claim is not supported. I would also encourage the authors to consider whether the dataset will be released or at least described in more detail, since the reproducibility of the comparison depends on it."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. First, the architecture is a reasonable systems contribution: replacing CLIP with SigLIP+DINOv2, fusing low-level and patch-level features through a dual-stream transformer, then projecting with a timestep-aware Q-former into FLUX. That is a real and coherent engineering choice, and the 10M-scale paired/unpaired dataset plus three-stage schedule is a serious attempt to separate identity and editability. If the code and data actually ship, this becomes a useful open-domain adapter.\n\nSecond, the evidence for the central claim is absent. Section 4 says InstantCharacter 'consistently performs the best' against OminiControl, EasyControl, ACE+, UNO, and GPT-4o, but the only comparison is a handful of qualitative grids. The sentence pointing to 'quantitative measurements shown in Fig. 6' is a misstatement: Fig. 6 is captioned 'More qualitative results' and contains no metrics. There is no identity-similarity score, no text-alignment metric, no user study, no statistical test. For a paper claiming to 'set a new benchmark,' that is a gap, not a style preference.\n\nThe three-stage schedule is another soft spot. The paper states the stages decouple consistency, editability, and fidelity and that stage two 'eliminates the copy-paste effect,' but no ablation shows the stages compose as claimed. It is possible they do; the claim is plausible. But 'plausible' is not support.\n\nThe novelty framing needs correction. Calling this 'the first DiT-based framework specifically optimized for character customization' is hard to defend when the paper itself cites OminiControl, EasyControl, and ACE++ — all FLUX/DiT-based adapters — and the related work includes 'Personalize anything for free with diffusion transformer.' The specific encoder stack and training regime are new, but the category is not.\n\nThis deserves a serious referee because the architecture is meaningful and the failure mode — qualitative-only evaluation — is fixable. I would accept it for peer review but ask for major revision: quantitative comparison on identity and text alignment, ablation of the three-stage design, and a corrected claim about what is actually new. If the repo genuinely releases weights and data, that would strengthen it.","headline":"A coherent DiT-based character adapter with a plausible training scheme, but the paper's strongest claims rest entirely on qualitative grids and a missing quantitative figure.","tokens_in":6534,"tokens_out":2046,"would_cite":false,"duration_ms":21509,"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":"InstantCharacter claims that a frozen diffusion transformer, fitted with a stacked encoder adapter and a three-stage training schedule, can personalize any unseen character from a single image while keeping full text control.","keywords":["character personalization","diffusion transformer","subject-driven generation","identity preservation","text-to-image generation","adapter","character consistency","text controllability"],"falsifier":"Train the adapter with the paired text-editing stage removed, using only unpaired self-reconstruction, then run text-edit prompts on unseen characters: if outputs still follow the prompt well or still show the copy-paste effect, the claimed separation between identity and editability fails. A second check is to measure identity similarity and prompt-following scores on identical prompts across InstantCharacter and the FLUX-based baselines; if the claimed consistent best performance does not survive a quantitative head-to-head, the qualitative central claim is falsified.","tokens_in":5617,"feed_emoji":"🎨","tokens_out":6144,"duration_ms":60909,"temperature":0.7,"pith_summary":"The paper proposes InstantCharacter, a framework that personalizes characters by training only a scalable adapter on top of a frozen diffusion transformer, with no per-subject fine-tuning. It claims this achieves open-domain personalization: an unseen character from one reference image can be placed in new poses, actions, and styles while preserving identity and following text prompts. The paper argues that previous adapter designs built for U-Net models fail on billion-parameter transformers, and that its full-transformer encoder stack is what makes identity injection and text editability coexist. It also introduces a 10-million-image dataset split into paired and unpaired subsets, with a three-stage training plan that separates identity learning, text editability, and image fidelity into distinct phases.","feed_headline":"One adapter personalizes any character on a frozen transformer","feed_subtitle":"Stacked encoders and 10 million training images keep identity while following text edits on unseen characters.","key_machinery":"The load-bearing object is the scalable adapter: a stack of transformer encoders that progressively refine character representations before a timestep-aware Q-former projects them into the frozen DiT's denoising space through learnable cross-attention layers. It combines two general vision encoders, SigLIP for fine-grained texture and DINOv2 for robustness to background interference, with a dual-stream feature fusion that preserves low-level details from shallow layers and region-level cues from image patches. The other mechanism is the training setup: a 10-million-image dataset split into paired and unpaired subsets plus a three-stage schedule that the paper says separates identity consistency, textual editability, and image fidelity.","core_discovery":"The core claim is that a frozen, billion-parameter diffusion transformer can be turned into an open-domain character personalizer by training a lightweight, stackable transformer adapter, provided the adapter is designed to feed rich multi-scale visual features into the DiT latent space. The paper reports that the adapter, combining SigLIP and DINOv2 features with low-level and region-level feature streams, preserves identity details that CLIP-based adapters lose, while a timestep-aware Q-former projection keeps character features compatible with the base model's text-conditioned denoising. A second claimed discovery is that paired multi-view data teaches textual editability and unpaired data teaches identity, so a three-stage curriculum lets both properties emerge simultaneously without the copy-paste effect or loss of prompt control. This, the paper argues, is what lets it beat UNet-era adapters and optimization-based personalization on unseen characters without per-subject fine-tuning.","pith_inferences":["If the three-stage separation is real, an ablation that trains only the unpaired stage should show strong identity but weak text following, while omitting the high-resolution stage should degrade texture; the paper does not report such ablations, so the composition claim is testable but unverified.","The same adapter design should transfer to non-character subjects such as products, animals, or multi-character scenes where identity must survive viewpoint and occlusion, but the paper demonstrates only single-character portraits and game assets.","The qualitative finding that style LoRAs can be combined without breaking identity suggests a practical identity-plus-style pipeline, yet the paper's evidence for that combination is visual only and lacks a quantitative consistency measure."],"forward_implications":["A single frozen foundation transformer can be repurposed for identity-preserving generation by training only the adapter, so scaling to larger base models does not require retraining the generator.","Open-domain characters never seen during training can be personalized in one forward pass, without per-subject fine-tuning or test-time optimization.","Text edits such as new poses, actions, and backgrounds remain controllable while identity is preserved, the property that optimization-based methods sacrifice.","The paired/unpaired dataset split and three-stage schedule offer a recipe for separating identity, editability, and fidelity in other subject-customization tasks."],"supporting_citations":[{"why":"Supplies the scalable rectified-flow transformer architecture that the adapter must interact with.","marker":"[2]"},{"why":"Provides the 12-billion-parameter FLUX model used as the frozen base generator.","marker":"[8]"},{"why":"Defines the cross-attention subject-adapter design and timestep-aware Q-former that the projection head is built on.","marker":"[24]"},{"why":"Supplies the SigLIP vision encoder used for fine-grained character features.","marker":"[25]"},{"why":"Supplies the DINOv2 encoder used to make character features robust to background interference.","marker":"[14]"},{"why":"Provides CLIP, the alternative image encoder the paper argues loses texture detail and is therefore replaced.","marker":"[16]"},{"why":"Defines the per-subject fine-tuning baseline whose text-controllability cost motivates the adapter approach.","marker":"[18]"},{"why":"Represents prior UNet-era stacked-embedding personalization that the paper claims does not scale to DiTs.","marker":"[10]"},{"why":"One of the FLUX-based customization baselines the paper compares against in qualitative evaluations.","marker":"[20]"}],"fun_headline_variants":["Frozen DiT + one adapter personalizes any character","Scalable adapter, frozen DiT: any character, no fine-tuning","10M-image training makes one adapter preserve identity and edits","Single adapter, frozen transformer: any character, text-editable","One adapter, 10M images: open-domain character customization"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The framework's success depends on the assumption that the three-stage training schedule separates identity, editability, and fidelity so cleanly that they reinforce rather than interfere, but no ablation is reported to show the stages compose without interference.","fun_headline_variants_meta":{"raw":{"variants":["Frozen DiT + one adapter personalizes any character","Scalable adapter, frozen DiT: any character, no fine-tuning","10M-image training makes one adapter preserve identity and edits","Single adapter, frozen transformer: any character, text-editable","One adapter, 10M images: open-domain character customization"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000776,"raw_usage":{"total_tokens":3431,"prompt_tokens":944,"completion_tokens":2487,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":560,"completion_tokens_details":{"reasoning_tokens":2401}},"tokens_in":560,"tokens_out":2487,"duration_ms":20064,"temperature":1.0,"reasoning_tokens":2401,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T12:32:23.510039+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the adapter with the paired text-editing stage removed, using only unpaired self-reconstruction, then run text-edit prompts on unseen characters: if outputs still follow the prompt well or still show the copy-paste effect, the claimed separation between identity and editability fails. A second check is to measure identity similarity and prompt-following scores on identical prompts across InstantCharacter and the FLUX-based baselines; if the claimed consistent best performance does not survive a quantitative head-to-head, the qualitative central claim is falsified.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the 12-billion-parameter FLUX model used as the frozen base generator."},{"cited_title":"In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition","cited_arxiv_id":null,"evidence_quote":"Represents prior UNet-era stacked-embedding personalization that the paper claims does not scale to DiTs."}],"review_version":1}