Jasna: Selective Video Restoration — When AI Only Processes 5% of the Frame
A Counterintuitive Engineering Choice
Traditional video restoration processes entire frames: every frame is fed into a super-resolution model, and the GPU reconstructs the whole image pixel by pixel. This is slow — a 10-minute 1080p 30fps video has 18,000 frames at ~2 million pixels each, meaning roughly 3.6 billion pixels of inference.
Kruk2/jasna makes a counterintuitive choice: only process the regions that need restoration. If only 5% of a frame needs fixing (e.g., a mosaic artifact block), Jasna runs inference on just that 5% and leaves the other 95% untouched. Compute drops from 3.6 billion pixels to 180 million — a 20x speedup.
This sounds simple but is engineering-intensive. The pipeline involves:
1. Detecting which regions need restoration (segment detection) 2. Cropping those regions 3. Running model inference on the small patches (which must handle arbitrary input sizes) 4. Compositing the results back into the original frame with seamless blending
Each step involves precision/performance trade-offs: coarse detection misses edges, small crops lose context. Jasna packages this pipeline into a usable desktop tool — free, open source, GPU-accelerated. Not an academic demo, but a practical product.
Two Models, Two Philosophies
BasicVSR++ (fast mode): The classic CVPR 2022 model with open weights, designed specifically for video super-resolution. Its core innovation is temporal propagation — rather than processing frames independently, it uses information from preceding and following frames via a bidirectional recurrent structure, with forward and backward passes fused at the current frame.
Temporal propagation matters because adjacent video frames contain massive redundancy: a region occluded in one frame may be clean in the previous or next frame. BasicVSR++ uses optical flow to align neighboring frames and "moves" clean information into degraded regions. This dramatically outperforms single-frame models like Real-ESRGAN, which can only infer from within one frame. The cost is speed and VRAM — but Jasna's selective strategy mitigates this by running multi-frame inference only on small patches.
LTX (high-quality mode): A newer, supporter-paid model using generative restoration — a diffusion model that generates missing content rather than only extracting it, producing better results at much higher compute cost.
The dual-model setup reflects a philosophy: fast mode for preview and iteration, high-quality mode for final output. Users run BasicVSR++ quickly to tune parameters, then render with LTX — more practical than a single model, since "good enough" restoration is ultimately a subjective human judgment.
Real-Time Preview: No Full-File Rendering Required
Traditional tools (FFmpeg, HandBrake) follow an input → process → output workflow, forcing users to wait for the entire file — possibly 30 minutes for a 10-minute video — before seeing results.
Jasna uses streaming processing + live playback: frames play as soon as they're processed, and users can pause, adjust parameters, and re-process at any time. The technical challenge is that video encoding typically requires "lookahead" (future frames as reference). Jasna solves this with a frame buffer: process N frames first, start playback, and keep processing subsequent frames into the buffer during playback.
This shrinks the iteration cycle from ~30 minutes to ~30 seconds.
Segment Editor: Humans in the Loop
Jasna's Segment Editor lets users manually select which regions to restore. This is deeper than manual annotation — it's AI detection + human confirmation.
Fully automatic detection can misjudge: a blurry region might be intentional depth-of-field, a compression artifact, or genuine damage. Automatic detection can't distinguish these. The Segment Editor lets users add missed regions, remove false positives, and adjust boundaries — an "AI proposes, human decides" collaboration that's more reliable than full automation and more efficient than full manual work.
GPU Optimization: VRAM Is the Bottleneck
Restoration models are VRAM-hungry — BasicVSR++ on 1080p video allocates large multi-frame feature maps. Jasna's selective processing creates a virtuous cycle:
- Small-patch inference → smaller intermediate feature maps → lower VRAM usage
- Lower VRAM usage → larger batch sizes or longer temporal windows → better restoration quality
Per the README, Jasna supports NVIDIA and AMD GPUs via both CUDA and DirectML backends — meaning Windows users with AMD cards can use it, unlike many AI video tools that are CUDA-only.
Beyond the Tech
Jasna's architecture — selective processing, dual-model strategy, real-time preview, human confirmation — forms a complete engineering paradigm for AI-assisted video restoration. Its core idea isn't "use a stronger model" but "use smarter engineering to make existing models efficient."
While AI video discussion focuses on bigger models, longer temporal windows, and higher resolutions, Jasna shows another path: don't process what doesn't need processing. The principle generalizes: not all data needs LLM inference, not all images need VLM analysis. The key is identifying which 5% truly needs AI, handling the remaining 95% with rules and caching.
This is engineering pragmatism correcting AI fundamentalism: not every problem needs the heaviest model.
---
*Project: Kruk2/jasna* *Model paper: BasicVSR++ (arXiv:2104.13371)*