DiffusionBrowser: Interactive Video Preview

• DiffusionBrowser is a model-agnostic, lightweight decoder framework that extracts semantically meaningful previews from intermediate video diffusion features.
• It decodes multiple scene intrinsics such as base color, depth, and normals, enabling users to inspect and control the denoising process early on.
• The framework achieves over 4× real-time preview speed and supports interactive controls like stochastic renoising and modal steering for enhanced video editing.

DiffusionBrowser is a model-agnostic, lightweight decoder framework for interactive previewing of video diffusion generation while denoising is still in progress. Rather than waiting for the full reverse process to terminate, it attaches a learned decoder to intermediate features of a frozen video diffusion transformer and produces preview videos from arbitrary denoising timesteps or transformer blocks. The preview space is multi-modal: alongside RGB, the system decodes scene intrinsics including base color or albedo, depth, surface normals, roughness, and metallicity. The framework is presented as a preview, probing, and intermediate-control layer for video diffusion models, with reported preview generation in less than 1 second for a 4-second video, i.e. more than $4\times$ real-time speed (Hong et al., 15 Dec 2025).

1. Problem setting and intended function

DiffusionBrowser is motivated by two practical limitations of video diffusion models: they are slow, and they remain opaque while denoising. A user often commits substantial compute before discovering whether a sample is worth continuing, while the denoising trajectory itself exposes little about the emerging scene, object layout, or motion. DiffusionBrowser addresses this by decoding semantically meaningful previews from intermediate internal features rather than waiting for the final clean sample. The resulting workflow is explicitly interactive: previews can be inspected during denoising, unpromising samples can be terminated early, and partially denoised states can be reused for branching or steering (Hong et al., 15 Dec 2025).

A central design decision is the use of scene intrinsics as preview targets. The paper argues that these representations provide a compromise between human interpretability and early emergence during denoising. Depth and normals expose geometry and spatial layout; base color exposes object identity and color without lighting artifacts; RGB remains the final appearance channel but is not necessarily the earliest semantically stable one. This suggests that DiffusionBrowser is aimed less at post hoc visualization than at exposing what the denoiser “already knows” before photorealistic synthesis has converged (Hong et al., 15 Dec 2025).

2. Intermediate-feature representation and preview space

The framework assumes a transformer-based video diffusion model with $N_t$ denoising steps and $N_b$ transformer blocks. At denoising timestep $t$ and block $b$, the relevant internal representation is the intermediate feature tensor $\mathbf{f}_{t,b}$. DiffusionBrowser trains a decoder that maps $\mathbf{f}_{t,b}$ into a preview output

$$\mathbf{y}_{t,b} \in \{\mathbf{b}, \mathbf{d}, \mathbf{n}, \mathbf{r}, \mathbf{m}, \mathbf{c}\},$$

where $\mathbf{b}$ denotes base color or albedo, $\mathbf{d}$ depth, $N_t$0 surface normals, $N_t$1 roughness, $N_t$2 metallicity, and $N_t$3 RGB color (Hong et al., 15 Dec 2025).

The paper situates this decoder within standard continuous and flow-based diffusion notation. It states the continuous noising process as

$N_t$4

and the flow-matching or ODE perspective as

$N_t$5

For DiffusionBrowser, however, the operational object is not the state equation itself but the feature tensor $N_t$6, which is treated as a probeable representation from which structured previews can be decoded at arbitrary points of the denoising trajectory. The system is therefore “preview anywhere” by construction: timestep selection and block selection are both first-class controls (Hong et al., 15 Dec 2025).

3. Multi-branch decoder and supervision design

The paper first considers a naive single decoder that predicts all modalities jointly, then identifies a superposition problem at noisy timesteps. The intuition is that the conditional distribution $N_t$7 can be highly multimodal early in denoising. The paper writes the forward corruption as

$N_t$8

and notes that mean-like predictors are driven toward

$N_t$9

which can traverse low-density in-between states and produce blur or hallucinated mixtures of plausible futures. A toy moving-dot experiment is used to illustrate duplicated dots, missing dots, and mode-averaging artifacts under early or few-step prediction (Hong et al., 15 Dec 2025).

To mitigate this, DiffusionBrowser introduces a multi-branch decoder. Instead of a single head, it uses $N_b$0 independent decoder branches, each with four 3D convolutional layers followed by two upscaling 3D convolutional layers. If $N_b$1 is branch $N_b$2, then each branch predicts

$N_b$3

and the ensemble prediction is

$N_b$4

Training combines branch-wise losses with an ensemble loss on the averaged prediction; the reported ensemble-loss weight is $N_b$5. The intended effect is that different branches specialize to different plausible modes while the ensemble remains aligned with the target preview (Hong et al., 15 Dec 2025).

Supervision is constructed from a synthetic dataset generated with Wan 2.1 as the frozen backbone. The paper reports 1,000 videos from 40 scene categories with 25 prompts each. DiffusionRenderer is used to obtain RGB and scene intrinsic channels as pseudo-ground truth. Intermediate features are cached from the video diffusion model, then the decoder is trained on those cached features. Preview outputs are generated at roughly $N_b$6 resolution, RGB and pseudo-ground truth are downsampled by linear interpolation, and every fourth frame is temporally subsampled to match the feature temporal size. This makes the trained module explicitly an external decoder rather than a retrained generator (Hong et al., 15 Dec 2025).

4. Quantitative behavior and evaluation

At 10% denoising, the main quantitative comparison is against four baselines: an $N_b$7-prediction baseline, Video Depth Anything, DiffusionRenderer applied to the RGB preview, and a linear probing decoder. Table 1 reports that DiffusionBrowser achieves RGB PSNR $N_b$8, base color $N_b$9, depth $t$0, normal $t$1, metallicity $t$2, and roughness $t$3, with runtime $t$4 s. The corresponding runtimes reported for the baselines are $t$5 s for the $t$6-prediction baseline, $t$7 s for Video Depth Anything, $t$8 s for DiffusionRenderer, and $t$9 s for the linear probe. In the same table, the linear probe is competitive in runtime but lower on most preview metrics, while the $b$0-prediction baseline is much slower and weaker for early previewing (Hong et al., 15 Dec 2025).

The supplementary metrics cover PSNR, MSE, $b$1, and LPIPS from 4% to 20% of the denoising schedule. The reported pattern is that DiffusionBrowser outperforms $b$2-prediction early for RGB, while $b$3-prediction becomes competitive or better around 16% denoising. For intrinsic channels, the proposed decoder consistently outperforms linear decoding and far exceeds the $b$4 baseline, which cannot provide those modalities directly. The user-facing significance is reinforced by a study with 35 participants, each evaluating 10 examples, yielding 350 responses per question. Preferences for DiffusionBrowser over the $b$5 baseline were 74.6% for content predictability, 72.9% for visual fidelity, and 76.9% for scene clarity (Hong et al., 15 Dec 2025).

5. Interactive control and probing of denoising dynamics

DiffusionBrowser is not limited to preview display. It adds two intermediate-control mechanisms. The first is stochasticity reinjection or stochastic renoising. Given a clean latent estimate $b$6, noise is reintroduced at a chosen preview timestep $b$7 via

$b$8

This preserves coarse structure while allowing multiple sibling continuations with different later details. The paper describes the resulting workflow as tree-like: a partially denoised state becomes a branch point rather than a one-way trajectory (Hong et al., 15 Dec 2025).

The second mechanism is modal steering. With the decoder frozen, intermediate features are optimized toward a target preview: $b$9 The supplementary specifies small gradient-based feature-space edits using the Jacobian of $\mathbf{f}_{t,b}$0. Demonstrated targets include base-color steering through K-means++ color clustering, depth-edge steering using a Sobel operator, and normal steering by flipping the $\mathbf{f}_{t,b}$1-axis of normals. The paper characterizes these as proof-of-concept controls rather than full video editing; it also reports failure cases in which inserted structures dissolve during later denoising, which it attributes to limited 3D understanding of the base model, limited decoder capacity, out-of-distribution steering targets, and the simplicity of the steering procedure (Hong et al., 15 Dec 2025).

The same decoder also functions as an interpretability probe. Linear and nonlinear probing indicate that intrinsic predictive power saturates early, around the 5th–15th of 50 timesteps and around the 10th–20th of 30 blocks. Depth and normals are decodable earlier than RGB, while RGB quality improves more monotonically with denoising progress and layer depth. Across blocks, lower layers contain coarse geometry and color distributions, mid-level layers are most predictive for intrinsics, and final layers are more specialized for appearance refinement. The paper interprets this as evidence that video diffusion features encode physically meaningful scene attributes and that denoising follows a coarse-to-fine organization in which layout, geometry, motion direction, and coarse object structure emerge before texture, lighting, and fine appearance (Hong et al., 15 Dec 2025).

6. Related meanings of “DiffusionBrowser” and stated limitations

Despite its name, DiffusionBrowser is not presented as a browser-native educational interface or a web analytics dashboard. This distinguishes it from “Diffusion Explorer,” which runs completely on the front end, uses TensorFlow.js for training and sampling, uses D3.js for animated vector graphics, and lets users train and inspect 2D diffusion models directly in the browser (Helbling et al., 1 Jul 2025). It also differs from DiVA, a scalable web interface for information diffusion on large networks with a JavaScript/HTML5 frontend, a Python/Flask backend, and synchronized comparison of graph diffusion runs (Sahnan et al., 2021). This suggests that “browser” in DiffusionBrowser names a preview-and-control layer for video denoising rather than a literal browser-based visualization platform.

The paper states several limitations. Its current scope is limited to scene intrinsics; interaction with text conditioning is not deeply studied. Steering can fail, especially when edits are geometrically extreme or out of distribution. The decoder is shallow and the previews are relatively low resolution. Pseudo-ground-truth intrinsics come from DiffusionRenderer and may contain errors, although the paper notes that the decoder can sometimes produce cleaner geometry than the pseudo-ground truth. The steering interface remains preliminary rather than a polished editing system. Future work is described in terms of stronger decoder architectures, clearer mode separation, higher-resolution outputs, richer intrinsic modalities, and tighter integration with text conditioning and more sophisticated control targets (Hong et al., 15 Dec 2025).