README.md

metadata

title: HighQualityVideoGeneration
emoji: 🎬
colorFrom: blue
colorTo: purple
sdk: gradio
sdk_version: 6.0.1
app_file: app.py
pinned: false
short_description: Generate a short video from an image and a text prompt

Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference

What this is

Image-to-video generation from thornmaze/WAMU_v3_WAN2.2_I2V_LIGHTNING via WanImageToVideoPipeline, running on Hugging Face ZeroGPU.

Post-processing

Two optional steps can run after generation:

step	what	runs on
Frame interpolation ("Video Fluidity")	RIFE v4.26 (thornmaze/RIFE), half precision. 2x/4x/8x the native 16 fps.	GPU, inside the generation call
Upscaling ("Upscale 4×")	4xLSDIRCompact (SRVGGNetCompact, num_conv=16), fp16 on GPU (fp32 CPU fallback), tiled to bound peak memory. Fixed 4x.	GPU, inside the generation call

Upscaling runs after interpolation, inside the same @spaces.GPU call as generation, so it bids for ZeroGPU worker time like the rest of the pipeline (moved off the Space's shared CPU — see issue #11 — to avoid CPU contention across concurrent visitors).

RIFE support code (model/warplayer.py, model/loss.py, model/pytorch_msssim/) is vendored from hzwer/Practical-RIFE (MIT, see LICENSES/RIFE-LICENSE) — required as sibling-import targets for the train_log/RIFE_HDv3.py module downloaded at runtime from thornmaze/RIFE.

The upscaling code (postprocess/upscale/) is vendored/adapted from xinntao/Real-ESRGAN (BSD-3-Clause License, see LICENSES/REAL-ESRGAN-LICENSE). The weights are Phips/4xLSDIRCompact (CC BY 4.0, see LICENSES/4xLSDIRCompact-LICENSE), not Real-ESRGAN's own — see the comment at the top of postprocess/upscale/upscale.py for why.

Debug logging (optional)

Each inference call can optionally be logged (prompt, seed, generation/interpolation/upscale settings, input image, pre-upscale output video, timing, success/error) to a private Hugging Face Hub dataset repo. Disabled by default — nothing is logged until both secrets below are set. Logging is based on legitimate interest (GDPR Art. 6(1)(f)), not consent — see the notice and Privacy Policy shown in the app for details, including how to request access to or deletion of your data. Configure independently per Space under Settings → Variables and secrets:

variable	type	notes
LOG_HF_TOKEN	Secret	fine-grained token, write-only on the target dataset repo. Do not reuse the deploy HF_TOKEN.
LOG_DATASET_REPO	Secret	target dataset repo id; auto-created (private) on first log.
LOG_STORAGE_CAP_GB	Variable	total-storage retention cap; oldest entries pruned first once exceeded. Defaults to 10GB if unset — dev should set 50, prod 450.
LOG_BATCH_INTERVAL	Variable	seconds between batched log commits. Default 60.

Logs are committed asynchronously in batches and never add latency to a generation request. The logged video is the pre-upscale result (post-interpolation, before 4x super-resolution); metadata references the image/video files by path rather than embedding them, so both stay browsable/playable in the Hub dataset viewer.

Status

Baseline (non-AOT) pipeline. AOT-compiled inference (faster, but requires a compatible precompiled package for WAMU_v3) is tracked separately — see the repo's open issues.

LoRA loading is out of scope for this version (see SRS FR-8/C-4).