Bagel-NHR-Edit: Efficient NHR Image Editing

• Bagel-NHR-Edit is a parameter-efficient, open-source model for non-human-region image editing that leverages automated triplet mining and LoRA fine-tuning.
• It modifies the generation expert within the BAGEL framework to enhance edit fidelity and achieve rapid, near-real-time inference through Hyper-Bagel acceleration.
• The extensive NHR-Edit dataset combined with validator-based fine-tuning delivers state-of-the-art consistency and perceptual accuracy across diverse editing tasks.

Bagel-NHR-Edit is a parameter-efficient, open-source adaptation of the BAGEL multimodal model, specifically optimized for non-human-region (NHR) image editing through large-scale, automated instruction-following triplet mining. It unifies advances in data synthesis, model architecture, validation, and acceleration for high-fidelity object-level editing at considerable computational efficiency and state-of-the-art faithfulness.

1. Model Architecture and Fine-Tuning Paradigm

Bagel-NHR-Edit is based on the original 14B-parameter BAGEL transformer, which features an Mixture-of-Transformer-Experts scaffold: a dedicated "understanding" expert processes textual and multimodal embeddings (essential for reasoning/recognition tasks), while a "generation" expert, sharing contextualized self-attention representations, produces edited images. For Bagel-NHR-Edit, only the generation expert is modified: all original parameters are frozen and LoRA adapters (rank 16, α=16, dropout=0.05) are inserted in the attention and feedforward blocks (Kuprashevich et al., 18 Jul 2025).

Supervised fine-tuning is performed on a triplet corpus—original image $I_0$, instruction $p_e$, edited image $I_e$—using a diffusion-based decoder. The loss optimized is: $$\mathcal{L}_\mathrm{SFT} = -\mathbb{E}_{(I_0,p_e,I_e)\sim\mathcal{D}} \left[ \log p_\theta(I_e \mid I_0, p_e) \right]$$ No auxiliary style or edit-specific losses are introduced; edit fidelity is entirely data-driven. At inference, LoRA weights are merged into BAGEL, producing an end-to-end editing model with enhanced faithfulness and perceptual coherence (Kuprashevich et al., 18 Jul 2025).

2. Autonomous Triplet Mining: NHR-Edit Dataset Construction

The core asset underlying Bagel-NHR-Edit is the NHR-Edit dataset—358,463 high-fidelity triplets acquired via a fully automated, human-out-of-the-loop pipeline. This pipeline comprises:

• Prompt Engineering: High-diversity text-to-image (T2I) prompts $p_\mathrm{t2i}$ and related I2I editing instructions $\{p_e\}_k$ are generated via OpenAI o3 (Kuprashevich et al., 18 Jul 2025).
• Candidate Synthesis: Multiple ($N,M$) original–edit pairs are sampled using third-party diffusion models, with initial filtering on caption-text plausibility.
• Two-Stage Validation: Coarse screening eliminates visually implausible or irrelevant outputs (Qwen-VL-72B). A fine-grained Gemini-2.0-Flash model scores "instruction adherence" ($s_\mathrm{adh}$) and "aesthetics" ($s_\mathrm{aes}$); the combined validator score $s$ is their geometric mean:

$p_e$0

Best-edited candidates pass only if $p_e$1.

• Data Augmentation: Semantic inversion doubles each triplet by auto-generating the reverse instruction, while compositional bootstrapping synthesizes multi-stage edit chains by mixing compatible edit pairs (Kuprashevich et al., 18 Jul 2025).

After a compositional expansion step and consistency re-filtering, this yields a $p_e$22.2$p_e$3 enlarged collection. NHR-Edit covers both photorealistic and synthetic scenes, a wide spectrum of aspect ratios (1:6 to 6:1), and diverse visual styles (anime, oil, glitch, caricature, etc.). Instruction complexity ranges from single-object edits to multi-part spatial, semantic, or stylistic changes.

3. Quantitative Benchmarks and Editing Outcomes

Bagel-NHR-Edit's performance is evaluated on two leading benchmarks, following each source's VLM-based automated protocol:

Model	ImgEdit-Bench Overall	GEdit-Bench SQ	GEdit-Bench PQ	GEdit-Bench O
BAGEL	3.30	7.98	6.57	6.92
Bagel-NHR-Edit	3.39	8.07	6.88	7.12

Where "Overall" is the composite rating; "SQ" and "PQ" are semantic consistency and perceptual quality (0–10). Bagel-NHR-Edit yields a +2.7% absolute gain in ImgEdit-Bench Overall and a +0.19 composite gain in GEdit-Bench over vanilla BAGEL. Task-sliced scores further show improved faithfulness on add (3.98→4.19), replace (3.50→3.77), remove (3.04→3.18), and style edits (4.22→4.30) (Kuprashevich et al., 18 Jul 2025).

A crucial attribute is the absence of additional handcrafted losses or human-in-the-loop validation, with gains entirely attributable to the scale, diversity, and precision of the synthetic triplets.

4. System Implementation, Accessibility, and Inference

Bagel-NHR-Edit is distributed as open-source LoRA adapters (available at https://riko0.github.io/No-Humans-Required/ and Hugging Face). The recommended inference pipeline uses the Hugging Face diffusers API:

$I_e$2

Hyperparameters include LoRA dropout=0.05, standard guidance scales, 25–50 denoising steps, and fixed-seed reproducibility settings. The data mining script is parameterized for $p_e$4 seeds per base image, $p_e$5 edit attempts, and strict validator cutoffs $p_e$6 (Kuprashevich et al., 18 Jul 2025).

5. Acceleration through Hyper-Bagel: 1-NFE Real-Time Editing

Hyper-Bagel introduces a suite of architectural and training innovations enabling substantial speedups for Bagel-NHR-Edit without compromising edit quality (Lu et al., 23 Sep 2025). These include:

• Speculative Decoding: A small "draft" model proposes $p_e$7 next tokens, batch-validated by the base model to accept maximal-matching prefixes, achieving $p_e$82.16$p_e$9 token throughput.
• Multi-Stage Diffusion Distillation: Original 100-NFE denoising is compressed to a 6-NFE "lossless" model via staged consistency and adversarial distillation, followed by a further reduction to a 1-NFE model using adversarial diffusion pre-training and reward feedback learning (HPSv3-based). The distilled 1-NFE variant achieves $I_e$022$I_e$1 faster editing, enabling near-instantaneous inference.
• Quantitative Results: On GEdit-Bench, Hyper-Bagel (6-NFE) matches or slightly exceeds the original BAGEL baseline (e.g., Overall 6.612 vs. 6.602); the 1-NFE variant, while lower in fine perceptual quality (Overall 5.975), preserves semantic correctness at real-time speeds: | Model | GEdit-Bench Overall (EN) | |---------------------------|--------------------------| | BAGEL (132-NFE) | 6.602 | | Hyper-Bagel (6-NFE) | 6.612 | | Hyper-Bagel (1-NFE) | 5.975 |

Acceleration stages are fully described in (Lu et al., 23 Sep 2025), with detailed hyperparameters and implementation notes.

6. Integration with Reasoning-Centric and Animal Knowledge Workflows

Bagel-NHR-Edit is positioned as a generic NHR editing engine, but its construction and benchmarking situate it within broader ecosystems:

• Unified Reasoning-Based Editing: By synthesizing instruction data with multi-step and nested logic (e.g. shape, color, count, location tasks), Bagel-NHR-Edit aligns with reasoning benchmarks such as UniREditBench, but focuses on object- and region-level edits rather than full symbolic chain-of-thought traces (Han et al., 3 Nov 2025). This suggests a plausible pathway for further improvements: augmenting data pipelines with programmatic or game-rule-driven CoT supervision, as in UniREdit-Bagel.
• BAGEL Animal Expertise Evaluation: The BAGEL benchmark targets species-level knowledge and is designed for continuous accuracy-tracking during NHR knowledge editing in LLMs (Shen et al., 17 Apr 2026). By analogy, Bagel-NHR-Edit’s parameter-efficient update strategies (e.g., LoRA) and triplet-based supervision could serve as a template for region-specific factual correction in parametric animal-knowledge models, with closed-book MCQ evaluation on taxonomy, bioacoustics, and ecological relations.
• Uni-Edit for Generalized Fine-Tuning: Uni-Edit demonstrates that conditional editing (with reasoning-intensive data and nested-instruction logic) lifts understanding, generation, and editing metrics jointly in UMMs such as BAGEL. In Bagel-NHR-Edit, the robust selection of object-centric instructions and the use of segmentation- or mask-aware pipelines directly implement these best practices for NHR contexts (Zheng et al., 20 May 2026).

7. Future Directions and Extensions

Recent results indicate that Bagel-NHR-Edit's modular and data-driven pipeline is extensible to domain-specific, logic-intensive, and high-velocity applications:

• Mask-aware and hierarchical region editing, as suggested by Uni-Edit, remain active research pathways for further sharpening performance in NHR use cases.
• Integration with dual-reference (image+text) evaluation, as in UniREditBench, may improve alignment between perceptual fidelity and rule-consistency in edits.
• Continuous validation on domain-specific benchmarks such as BAGEL can ensure targeted factual edits do not introduce regressions across knowledge categories.

A plausible implication is that the automated, validator-centric triplet mining underlying Bagel-NHR-Edit can be generalized to any closed-region or object-centric edit regime requiring minimum human annotation, thereby scaling instruction-following capabilities for both research and production-grade editing engines.

---

References:

• "NoHumansRequired: Autonomous High-Quality Image Editing Triplet Mining" (Kuprashevich et al., 18 Jul 2025)
• "Hyper-Bagel: A Unified Acceleration Framework for Multimodal Understanding and Generation" (Lu et al., 23 Sep 2025)
• "Uni-Edit: Intelligent Editing Is A General Task For Unified Model Tuning" (Zheng et al., 20 May 2026)
• "UniREditBench: A Unified Reasoning-based Image Editing Benchmark" (Han et al., 3 Nov 2025)
• "BAGEL: Benchmarking Animal Knowledge Expertise in LLMs" (Shen et al., 17 Apr 2026)