- Rating
- 3
- Confidence
- 4
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
SwiftI2V tackles a practical problem: generating high-resolution image-to-video results without the cost of running one model over the full 2K sequence. It uses a low-resolution stage to generate a motion reference, then a high-resolution stage conditioned on that reference and the input image. Its CSG strategy generates short temporal segments with a bounded context window.
Critical strengths
- High-resolution generation has become increasingly important in the field of video generation. Higher resolution requires more computation and memory, so efficiency and visual quality are difficult to improve together. The authors address both in a single system: the first stage handles motion at low resolution, and the second generates high-resolution detail using both the input image and the motion reference. This is a reasonable design for efficient generation, and the transition training is a practical attempt to account for artifacts passed between the stages.
- CSG has a clear practical purpose. Processing a bounded temporal window gives the second stage a way to control memory use as video length grows. The bidirectional interaction within the window is a plausible choice for maintaining alignment with the input and nearby frames. The ablation against a causal version, along with the longer-video scaling results, helps show what this design contributes.
- The paper includes benchmark and human ratings, component ablations, analysis of segment boundaries, scaling to longer videos, and a run on an RTX 4090.
Critical weaknesses
- The related-work discussion omits several relevant HR video generation papers: UltraGen and T3-Video, which explore native high-resolution generation with hierarchical attention; PixelWizard, which uses a compact anchor to guide 2K/4K detail synthesis; and Ultra Flash, which targets real-time streaming generation at 2K.
- More information about dataset sizes and filtering, baseline inference settings, and what is included in the timing measurements would also help. There is a small resolution discrepancy to clarify: the default output is listed as 2560 × 1408, while the VAE reconstruction experiment uses 2560 × 1440.
- The human study reports 150 ratings per method and dimension, but not the number of raters, how samples were assigned and randomized, or how much the ratings varied. Those details would help readers judge how strong the human-study result is.
Questions for authors
- How are text prompts used in each stage, and do all baselines receive the same prompts and input images?
- Can you provide more detail on the human-study protocol and uncertainty?
- How does SwiftI2V differ from other related high-resolution anchor-guided or streaming methods?
3 comments
I appreciate the authors’ response and will maintain my positive rating.
Thank you for your positive assessment and constructive feedback, which helped improve the revised paper. We appreciate your support.
We thank the reviewer for the positive and detailed feedback.
W1 & Q3. Related work and differences. We have added and discussed these four papers.
SwiftI2V differs in two respects. (i) Every fixed-size window contains the user's HR input image as an appearance anchor and follows the global Stage I motion reference; none of the above methods keeps a persistent HR image condition (e.g., Ultra Flash conditions on text and LR latents). (ii) History blocks interact bidirectionally with the current segment at every denoising step, but only the current segment is updated, so per-step tokens are bounded and output is incremental; in contrast, the history in FlashVSR and Ultra Flash is read-only, and UltraGen, T3-Video and PixelWizard process the whole video at every step.
W2. Implementation details. Data. Stage I: a 360P LoRA trained for 20K steps on about 10K OpenViD-HD clips. Stage II: full DiT fine-tuning, first for one epoch on 1080P videos from the same pool, then for 10K steps on a subset of UltraVideo (about 40K 2K clips); the first 7K steps use real videos only, and the last 3K steps mix in about 2K stage-transition synthesized samples at a ratio of 7:3 (synthesized:downsampled). Both datasets are already quality-curated, and we only filter by resolution (at least the training resolution) and duration (at least 81 frames). Training uses 8 H800 GPUs with a per-GPU batch size of 2 for Stage I and 1 for Stage II.
Baselines. All baselines use official code, weights and default sampling settings, with outputs at 2560×1408 and the same input images and prompts; the VSR baselines (DiffVSR, Stream-DiffVSR, FlashVSR-v1.1) perform 4× super-resolution on the same Stage I 360P output. Sampling steps are 50 for CineScale and DiffVSR, 40+3 for LTX-2 (half-resolution generation, then latent upsampling and refinement), 4 for Stream-DiffVSR and 1 for FlashVSR. CineScale runs out of memory on a single H800 at 2K with 81 frames, so Tab. 3 reports its minimum runnable configuration (4 GPUs).
Timing. All methods in Tab. 3 follow the same protocol: after one full warm-up run (which also loads the models), each timed run reloads the models and covers the whole process up to the final 2K video (both stages for SwiftI2V; Stage I plus SR for the VSR pipelines; the full pipeline for CineScale and LTX-2). GPU time = #GPUs × time.
Resolution. The model requires sizes in multiples of 16, so the 360P Stage I uses 640×352 and Stage II outputs 4× that size, i.e., 2560×1408. The VAE experiment bypasses the two-stage pipeline, so it uses the standard 2K resolution of 2560×1440 to measure VAE reconstruction of 2K details.
W3 & Q2. Human study. Fifteen randomly recruited volunteers (a few researchers, mostly non-experts) rated independently. We randomly sampled 20 cases from scenes where all methods have outputs, and each volunteer was randomly assigned 10 cases, yielding 150 ratings per method and dimension. In each case, the six videos are anonymized and randomly ordered, and raters score each from 1 to 3 on four dimensions against the HR input image (per-level criteria in App. B.4). Mean±std of each method:
Q1. Prompts. All methods use the official VBench-I2V prompts and the same input images. Both stages of SwiftI2V use the same prompt: Stage I takes the downsampled input image and the prompt; Stage II encodes the prompt with the frozen text encoder for the DiT. CineScale and LTX-2 take the same input image and prompt; the VSR baselines take the same Stage I output, which is generated from the same input image and prompt.
Revisions. Related work: the four papers above; Sec. 3.2: the comparison above; App. B.1–B.3: data, prompts, resolution, baseline settings; App. C.8: VAE resolution; Tab. 3 caption and App. B.3: timing protocol; App. B.4: user-study protocol and rating criteria.