Generation gets all the attention, but a finished piece usually needs post work: remove a watermark the model invented, fix a frame rate that does not match the project, upscale before delivery. This release covers all three, and the ordering matters.
The three stages
| Stage | What it fixes |
|---|---|
| Watermark handling | Model-invented text or marks the prompt did not ask for |
| Frame interpolation (GIMM) | Frame rate too low for the delivery target |
| Upscale (FlashVSR) | Resolution too low |
Why the order is what it is
The instinctive order — upscale first, then fix — is wrong for two reasons.
Upscaling amplifies defects. A watermark removed after upscale has to be removed at 4K, from a much larger image, with more context to invent. Removing it at source resolution is cheaper and more accurate.
Interpolation before upscale creates frames the upscaler must handle. Interpolated frames are synthesised, and synthesised frames upscale worse than real ones. Upscaling first, then interpolating, keeps the interpolation working on high-quality input.
So: watermark removal → upscale → interpolation. The release's own ordering follows this, and it is the ordering to preserve.
Watermark removal on generated video
Model-invented watermarks are a specific problem: the model learned that video has text in the corner, so it puts text in the corner. The text is not a real watermark, it is a hallucinated one.
The fix is not masking — that leaves a rectangle. It is regeneration of the region, using surrounding frames as the reference for what should be there.
This is different from removing a real watermark from real footage, where you are trying to reconstruct what was behind it. For a hallucinated mark, the correct content is simply whatever the model would have produced without it, and neighbouring frames are a good guide.
Frame interpolation to a delivery rate
GIMM (as used here) synthesises intermediate frames. Two properties matter:
Motion coherence. Interpolation works on consistent motion. A clip with jump cuts or occlusion transitions will produce warping at those points.
Rate choice. Interpolating from 24 to 60 fps gives smooth slow motion potential. Interpolating from 24 to 30 gives a subtle improvement. Interpolating 24 to 120 gives you more frames than most content needs and more warping than you want.
Upscaling for delivery
FlashVSR is a video super-resolution model. The distinction from image upscalers matters: video upscaling has temporal information to work with, so it can hold detail consistent across frames in a way a per-frame image upscaler cannot.
What it does not do: fix motion blur, recover detail that was never generated, or improve composition. Upscaling makes a good clip deliverable; it does not make a bad clip good.
Where this pipeline breaks
- Watermark removal needs frames without the mark. If the model put text in every frame at the same position, there is no clean reference — you are reconstructing from spatial context only, which is weaker.
- Interpolation warps at transitions. Occlusion and cuts are where it fails.
- Three passes compound compression. Each stage re-encodes; use high-bitrate intermediates.
- No model version pins in the description — FlashVSR and GIMM have multiple releases, and results differ between them.
Practical workflow
- Generate at whatever resolution is practical
- Remove hallucinated text at that resolution
- Upscale to delivery resolution
- Interpolate to delivery rate
- Encode once at the end
Encode once. Every intermediate should be high-bitrate or lossless, because the final encode is the only one that gets optimised for distribution.
Files
https://huggingface.co/AI-KSK/ksk-video-176-comfyui-workflows
Notes
- Workflow files only. Model weights are not included; each keeps its own licence.
- Tutorial exports, not a product.
- Independent AI-KSK release. Not affiliated with ByteDance, FlashVSR or RunningHub.
Optional API platform - 88API (affiliate link)
Explore 88API for access to multiple AI model APIs. Available models, routes and prices are listed on the platform; Alipay top-ups are supported.
https://88api.ai/sign-up?aff=usx3
Disclosure: I may earn a commission from eligible purchases after you register through this referral link.
Top comments (0)