Glance: Accelerating Diffusion Models with 1 Sample¶

Conference: ECCV2026
Paper: ECCV Paper
Code: https://zhuobaidong.github.io/Glance/
Area: Image Generation
Keywords: diffusion models, phase-aware acceleration, low-rank adaptation, flow matching, one-sample training

TL;DR¶

Glance freezes a pretrained image generator and trains two LoRA adapters on one image-text sample for early semantics and late details, using nonuniform step reduction for the paper's claimed fivefold acceleration, but without preserving text rendering losslessly.

Background & Motivation¶

Diffusion and flow matching models repeatedly invoke a large denoising network, so generation cost depends on both parameter count and repeated evaluations along a noise trajectory. FLUX.1 and Qwen-Image already contain extensive visual knowledge, yet directly reducing sampling from 50 steps to a handful can disrupt their generation trajectories. Existing few-step distillation trains students through trajectory, consistency, or distribution matching, with additional data and compute costs that can greatly exceed ordinary customization fine-tuning. Repeating expensive distillation after every backbone customization can partly offset inference savings with additional development costs.

Glance starts from nonuniformity within denoising: early high-noise steps determine subjects, layout, and global semantics, whereas later low-noise steps focus more on textures and details. Uniform skipping places the same compression pressure on both phases, overlooking how early structural mistakes can propagate through subsequent steps. The paper therefore retains the original generator instead of designing a smaller network and assigns more of its limited sampling budget to semantically sensitive stages. Here, slow and fast describe progress along the original trajectory, not networks with different sizes or hardware execution speeds.

One-sample training does not mean learning open-domain generation from one image: the frozen pretrained parameters already contain that knowledge. The added parameters adapt the model to a sparser timestep schedule rather than relearning object categories, composition, and textual semantics. This interpretation makes extreme data efficiency a reasonable experimental question, but does not itself establish a theoretical guarantee of one-sample generalization. Core Idea: use phase-specific low-rank updates to correct few-step denoising, advancing conservatively early and taking larger strides later to reuse pretrained generation capabilities with very little training data.

Method¶

Overall Architecture¶

Training takes one image-text sample, a pretrained generator, and denoising timesteps; inference still takes a new prompt and initial noise. Glance applies phase-aware step allocation, adapts Slow-LoRA and Fast-LoRA to their respective noise regions, and performs few-step inference using the same phase division. The adapters share a frozen backbone; they are neither two complete generators connected in sequence nor two experts evaluated together at every step. The training image supplies flow-matching supervision and is not required as an additional conditioning input at test time.

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
    Input["Training: one image-text sample<br/>frozen backbone and timesteps"] --> Phase["Phase-aware step allocation"]
    Phase --> Experts["Dual-expert low-rank adaptation"]
    Experts --> Supervision["Phase-specific flow-matching supervision"]
    Supervision -->|Two trained adapters| Sampling["Inference: Slow-LoRA first<br/>then Fast-LoRA"]
    Prompt["New prompt and initial noise"] --> Sampling
    Sampling --> Output["Generated image"]

The supervision stage in the diagram runs only during training; inference does not feed generated results back into a training loss. The current phase selects the expert, without a separately trained router that scores image content. Inference savings therefore mainly come from fewer network evaluations, while low-rank adaptation lowers training costs and recovers quality after step reduction.

Key Designs¶

1. Phase-aware step allocation: preserve denser sampling for early semantic formation

The paper treats the high-noise region as semantic formation and the low-noise region as detail refinement, using signal-to-noise ratio (SNR) to motivate their transition boundary. This division is meant to align timestep allocation with different generation roles rather than mechanically splitting all timestep indices into equal intervals. Slow-LoRA operates in the former region with smaller strides to stabilize global structure; Fast-LoRA operates in the latter with larger strides that skip some repeated refinement. The design does not establish that all late steps are useless: under a limited evaluation budget, it prioritizes early decisions whose errors can propagate.

Figure 3 provides a concrete 50-to-10-step configuration: select every 2 steps within the first 10 timesteps, producing 5 slow-stage sampling points. Uniformly select 5 fast-stage points from the remaining 40 timesteps, then execute the first group before the second at inference. Although 10 sampling evaluations remain, their density is higher in the first region than the second, unlike uniform sampling across the entire trajectory. However, Section 3.2 says both that SNR decreases during denoising and that the process moves from low to high SNR; the interpolation endpoints in Section 3.1 also conflict with the later time-direction description. This note therefore retains the readable phase principle and illustrated allocation rather than turning phrases such as half the initial SNR into an executable threshold algorithm.

2. Dual-expert low-rank adaptation: share knowledge while separating phase corrections

A single set of LoRA parameters covering semantic formation and texture repair can receive competing updates from different noise regions. Glance uses two independent pairs of low-rank matrices and adds the update associated with the current phase to the frozen weights. For an adapted weight matrix, the structural relationship is:

\[ \Theta_i'=\Theta+B_iA_i,\qquad i\in\{1,2\}. \]

Here \(\Theta\) denotes the original weights, and \(A_i\) and \(B_i\) are the trainable matrices of expert \(i\); the intermediate rank is much smaller than the original input and output dimensions. The experts do not store two complete backbones; phase-specific capacity primarily resides in the two sets of incremental parameters. This also clarifies student terminology: the student is a few-step generator consisting of a backbone plus adapters, not an independent network with dramatically fewer parameters. The paper does not provide an explicit rank configuration sufficient to calculate the complete trainable parameter count, so common defaults should not be assumed.

Section 4.1 explicitly extends adapter placement beyond the standard attention query, key, value, and output projections. Adapters also cover auxiliary projection layers and modality-specific MLPs in both visual and textual branches to preserve cross-modal dependencies under few-step updates. These placements belong to the same dual-expert low-rank adaptation design rather than introducing a third expert type. The trade-off is that inference still runs the large FLUX.1 or Qwen-Image backbone; few trainable parameters do not imply a small resident model memory footprint.

3. Phase-specific flow-matching supervision: learn denoising directions rather than memorize the training image

Training constructs intermediate noisy latents from the image, noise, timestep, and text condition, and asks the active expert to predict the velocity field at that state. The paper describes a mean-squared error between predicted velocity and a target flow vector, with an optional timestep-dependent weight \(w(t)\). Each expert receives supervision only within its assigned phase: the frozen backbone supplies existing representations, while low-rank parameters absorb phase-specific adaptation updates. The objective does not require another discriminator or introduce an adversarial loss between the outputs of the two experts.

The extracted flow-matching loss is missing some operators, and the construction procedure for the target flow vector is not fully specified in the prose. This note therefore explains the explicitly described velocity-regression mechanism without presenting a repaired expression as the author's exactly reproducible equation. In particular, the word distillation alone does not establish an additional teacher-trajectory cache, per-step teacher regression, or an unstated distribution-matching loss. The authors attribute data efficiency to direct supervision of denoising directions through flow matching, but this explanation is not independently established by a theoretical proof.

A Worked Example¶

Consider the 10-step configuration in Figure 3: first train the two adapters using the selected image-text sample, then generate from an unseen prompt. The new prompt passes through the backbone's existing text processing, and initial noise enters the denoising trajectory as the image latent. The first 5 retained sampling points use Slow-LoRA to maintain relatively fine progress through the region corresponding to the original first 10 steps. The process then switches to Fast-LoRA and performs only another 5 sampling evaluations across the span of the original final 40 steps. The original generation pipeline produces the image; it does not retrieve or copy patches from the training image. An insufficient late refinement budget can blur small character strokes, consistent with the paper's text-rendering failures, though this is not a deterministic prediction for every prompt.

Loss & Training¶

Tables 1 and 2 consistently use the Fox image-text sample illustrated in Figure 6, and backbone parameters remain frozen throughout training. Experiments cover FLUX.1-12B and Qwen-Image-20B, with FlowMatchEulerDiscreteScheduler used for the teacher baseline. Table 2 lists 0.6 GPU hours for the 8-step model and 0.8 GPU hours for the 10-step model; the abstract claims training within one hour on a single V100. Figure 1 instead labels its horizontal axis in A100 GPU hours, while Table 2 does not identify hardware per entry, so these figures should not be treated as a rigorously hardware-normalized cost comparison. One sample can still supply supervision at multiple noise states and timesteps; it does not mean only one parameter update. The cached text does not sufficiently specify learning rate, optimizer, training epochs, or the exact phase threshold, preventing recovery of a complete training recipe from this account alone.

Key Experimental Results¶

Main Results¶

Evaluation covers COCO-10k, HPSv2, OneIG-Bench, GenEval, DPG-Bench, and LongText-Bench, with the main generation resolution stated as \(1024^2\). The following subset of original Table 2 on page 6 reports OneIG-Bench scores, all higher-is-better; Data denotes the number of training samples. NFE counts network function evaluations; Qwen's \(\times2\) notation is retained, and its 10-step setting must not be treated as the same budget as another method's 4 NFE.

Model / Config	Data	NFE	Alignment	Text	Diversity	Style	Reasoning
FLUX.1 dev	Not applicable	50	0.790	0.556	0.238	0.370	0.257
Glance (FLUX)	1	10	0.788	0.328	0.204	0.358	0.231
Qwen-Image	Not applicable	50×2	0.880	0.888	0.194	0.427	0.306
Qwen-Image, direct step reduction	Not applicable	10×2	0.802	0.693	0.156	0.410	0.290
Qwen-Image-Lightning	0.4M	4	0.885	0.923	0.116	0.417	0.311
Single LoRA, uniform timesteps	1	10×2	0.621	0.332	0.097	0.298	0.193
Glance (Qwen)	1	10×2	0.868	0.734	0.160	0.421	0.303

Qwen Alignment recovers from 0.802 with direct step reduction to 0.868, close to the teacher's 0.880; Text still declines from the teacher's 0.888 to 0.734. FLUX Text falls from 0.556 to 0.328, so near-teacher performance cannot be generalized to preservation of every capability. Table 1 on page 6 also reports Qwen COCO FID changing from 34.1 for the teacher to 37.8 for Glance, compared with 39.1 for direct 10-step sampling; lower FID is better. On HPSv2 prompts in that table, HPSv2.1 changes from the teacher's 0.309 to 0.308, illustrating that similar preference scores do not rule out losses elsewhere. Reducing 50 steps to 10 gives a fivefold step-count reduction, but the text lacks sufficient matched-hardware end-to-end latency statistics to unconditionally equate this with measured fivefold throughput.

Ablation Study¶

Original Table 3 on page 12 uses Qwen-Image and OneIG-Bench to compare two experts, adaptation in only one phase, and single-expert configurations. Base means the corresponding phase does not use a trained phase adapter; this table is reported separately rather than merged with identically named configurations in Table 5.

Config	Alignment	Text	Diversity	Style	Reasoning
Slow3–Fast5	0.849	0.614	0.152	0.396	0.284
Slow3 + Base5	0.805	0.567	0.123	0.368	0.255
Base3 + Fast5	0.747	0.521	0.125	0.372	0.243
Single (identical)	0.702	0.453	0.110	0.342	0.218
Single (uniform)	0.621	0.332	0.097	0.298	0.193

Removing early adaptation gives Alignment 0.747, whereas removing late adaptation gives 0.805, supporting greater sensitivity to the early phase. At identical timesteps, two experts score 0.849 versus 0.702 for one expert, suggesting that reallocating sampling points is not the only source of gains. Original Table 5 on page 12 separately varies slow and fast step counts while holding the training set at one image-text sample.

Timestep Config	Alignment	Text	Diversity	Style	Reasoning
Slow3 + Fast5	0.863	0.692	0.162	0.414	0.286
Slow5 + Fast5	0.868	0.734	0.160	0.421	0.303
Slow5 + Fast10	0.874	0.813	0.175	0.422	0.305

Tables 3 and 5 disagree for Slow3–Fast5: Alignment/Text are 0.849/0.614 in the former and 0.863/0.692 in the latter, without a clear explanation of the difference. Table 3's Single (uniform) scores also exactly match Table 2's uniform-LoRA row at 10×2 NFE, whereas Section 4.3 describes an eight-step comparison; budget correspondence needs verification.

Key Findings¶

Increasing Fast5 to Fast10 in Table 5 raises Text from 0.734 to 0.813; the late phase is not entirely redundant, and detail fidelity still requires a sampling budget.
Table 4 on page 12 increases training data from 1 to 100 samples: Alignment changes from 0.868 to 0.876 and Text from 0.734 to 0.753, with limited and nonmonotonic gains across metrics.
That data experiment fixes training epochs, so increasing sample count also changes the number of updates; it is not a strictly compute-matched data-scaling comparison.
Table 6 on page 14 gives Fox and Bookstore Alignment scores of 0.868 and 0.797, but Bookstore Text is 0.751 versus Fox's 0.734; Fox is not best on every metric.
Failure when training on Gaussian noise is a qualitative observation in Section 4.4 without a numerical table row, so no benchmark score should be inferred for it.

Highlights & Insights¶

One-sample training matters as sampling adaptation that reuses pretrained knowledge, not as generation capability emerging from almost no data. This framing is more explanatory than the data count alone.
Timestep allocation and parameter specialization are complementary controls. The identical-timestep single-expert comparison helps distinguish their contributions.
Late-detail failures identify a boundary for the fast phase. Character strokes remain important content and cannot be assumed skippable without loss simply because they are refined late.

Limitations & Future Work¶

The authors explicitly show failures on dense small text; page 14 reports LongText-Bench gaps of 14.7 and 13.0 points behind the Qwen and FLUX teachers, respectively.
Sample selection still matters, with Bookstore substantially below Fox on some metrics; the limited sample comparisons do not establish equal effectiveness for arbitrary natural images.
SNR direction, time conventions, correspondence between table configurations, and GPU cost accounting remain unclear and should be verified rather than completed from standard assumptions.
Section 4.2 mentions image editing in Appendix A, but the available cache ends with the references, so its editing methods and results cannot be confirmed here.
A suggestion from this note is to allocate more late-stage steps to text-dense prompts and evaluate data scaling and phase allocation under matched compute with multiple random seeds; these are not completed experiments in the paper.

Compared with Timestep Master: both use time-dependent LoRA, while Glance emphasizes nonuniform semantic-to-detail step reduction and extreme low-data adaptation; its contribution is not simply the first use of multiple LoRAs.
Compared with consistency or distribution-matching distillation: Glance retains the complete backbone and trains phase-specific updates, prioritizing low training cost; other methods' lower NFE and stronger results on some metrics remain practically meaningful.
Compared with direct step reduction: the matched Qwen backbone results show partial quality recovery through adaptation, demonstrating that reducing evaluations and learning to generate in fewer steps are distinct operations.
Transferable insight: phase-dependent compute allocation is worth exploring in other generative tasks, but this text-to-image evidence does not establish effectiveness for video or arbitrary customized models.

Rating¶

Novelty: 4/5. Phase-aware dual LoRA and one-sample results form an appealing combination, although timestep experts and low-rank adaptation have precedents.
Experimental Thoroughness: 3/5. Multiple backbones, benchmarks, and ablations provide useful coverage, but inconsistent configurations and cost accounting limit conclusion strength.
Writing Quality: 3/5. The central idea is intuitive, but conflicting SNR statements, missing details, and overly strong lossless-quality claims require careful reading.
Value: 4/5. A practical direction for inexpensive few-step adaptation in generation settings that can tolerate losses in textual detail.