IAMCCS FlashVSR Panel Batch Prep

FlashVSR is the open video upscaler that actually got adopted - it arrived in late 2025 and the community treated it like a gift, because it upscales video at diffusion quality without a data center. But it has a real-world quirk: it wants a decent number of frames to chew on. If you're feeding it a panel-based Ideogram workflow - a shot that's six still frames because that's what a storyboard sheet produced - it may not have enough temporal context to do its thing.

That's the problem FlashVSRPanelBatchPrep solves. It takes your small IMAGE batch, figures out how many times it needs to repeat frames to hit a minimum count, and hands FlashVSR a full-length clip. Pair it with its mirror node IAMCCS_FlashVSRPanelBatchRestore, and the round trip is invisible to your final edit.

How it works

The mechanism is delightfully simple, and the source is small enough to trust. It reads images (the batch), counts them (original_count), and computes:

repeat_factor = ceil(min_frames / original_count)

Then it duplicates frames with torch.repeat_interleave - every frame is repeated repeat_factor times, so a 6-frame shot with min_frames = 24 becomes a 24-frame clip where each frame appears 4 times in a row. Outputs:

• flashvsr_frames - the padded IMAGE batch, ready for the upscaler.
• original_count and repeat_factor - the exact numbers you'll feed back into the Restore node afterwards.
• report - a string with the shape math.

Note min_frames is clamped to at least 21 in the code, so you can't accidentally set a target below what the upscaler tolerates.

Inputs that matter

• images - your panel/sequence batch.
• min_frames - the only real knob. Default 24 is sensible; raise it if your FlashVSR path wants more temporal context.

Install

Part of IAMCCS-nodes:

cd ComfyUI/custom_nodes
git clone https://github.com/IAMCCS/IAMCCS-nodes.git

Or ComfyUI Manager → "IAMCCS" → install → restart. FlashVSR itself is a separate install (e.g. ComfyUI-FlashVSR_Ultra_Fast) - this node just prepares its input.

Gotchas

Repeated frames are not new motion. The upscaler sees 24 frames but they're really 6 distinct images repeated, so don't expect temporal smoothing that wasn't there. If you want smoother results, interpolate between the real frames before upscaling. Also keep the original_count and repeat_factor outputs wired into the Restore node - losing them means you can't cleanly get your original shot count back. And keep in mind FlashVSR's own split: tiled mode is far slower but sips VRAM; the non-tiled path is fast and hungry.