toobusy FlashVSR VRAM & Resolution Preset
Pick your VRAM budget and your output size, get every knob pre-set
- width
- height
- scale
- chunk_frames
- chunk_overlap
- offload
- aggressive_offload
- tile_preset
FlashVSR is powerful and, left to its own devices, a VRAM management project. The toobusy FlashVSR VRAM & Resolution Preset node is the answer to that: two dropdowns that set everything. Pick your GPU class and your target output resolution, and it hands you the width, height, scale, chunk settings, offload flags, and tile preset that the rest of the FlashVSR chain expects - all derived from the author's own validated configurations.
How it works
The node encodes two tables. The VRAM profile table maps GPU classes to an offload strategy and a decoder tile preset:
12GB safe- aggressive offload + safe tile preset (max compatibility, slowest).16GB balanced- standard offload + balanced tiles.24GB+ fast- GPU resident (no offload) + fast tiles (fastest on big cards).32GB aggressive- aggressive offload + balanced tiles, tuned so high-resolution work fits without melting.
The resolution table lists sixteen output presets - landscape and portrait for 2K, 3K, 4K/UHD and QHD-aligned sizes. The sharp edge here: the preset names show the final output resolution, but FlashVSR upscales 2x, so the sampler actually works at half that. The node handles the bookkeeping - it outputs the pre-upscale width/height for the sampler and the effective scale, so you don't have to divide by two in your head. High-res presets are rounded so internal 128-multiple crops match the stated output.
The outputs and the chain
Eight outputs, and they wire up in an obvious order:
ToobusyFlashVSRSettings → ToobusyFlashVSRLoader (offload, aggressive_offload)
→ ToobusyFlashVSRSampler (width, height, scale, chunk_frames, chunk_overlap)
→ ToobusyFlashVSRDecoder (tile_preset)
width/height- pre-upscale sampling size (half of the named output).scale- always 2 for these presets.chunk_frames(21) /chunk_overlap(8) - the long-video chunking that keeps a 5-minute clip from OOMing.offload/aggressive_offload- for the Loader.tile_preset- a STRING for the Decoder's tile mode.
Install and gotchas
Pack install is the standard Manager "toobusy" or git clone + restart. This node itself has no heavy deps - the FlashVSR chain's requirements (and the dreaded block-sparse-attention wheel) live in requirements_flashvsr.txt, installed once for the whole chain.
The important habit: this node is a preset, not a guarantee. It was validated on a 24GB RTX 3090 at 1024x576 input, 21 chunk frames, 8 overlap, full tiled VAE, BSA - those are the numbers you're getting when you pick "24GB+ fast" + "2048x1152". If you're on a 12GB card and pick a 4K preset, the presets will keep you alive but the decode will be slow, because "12GB safe" is paired with the safest tile preset by design. When in doubt, start at a smaller output preset and climb - the node makes the knob changes trivial, which is exactly why it exists.
Inputs (2)
| Name | Type | Default | Description |
|---|---|---|---|
| vram_profile | COMBO | 24GB+ fast | 4 options: 12GB safe, 16GB balanced, 24GB+ fast, 32GB aggressive |
| output_resolution | COMBO | 2048x1152 landscape | 16 options: 2048x1152 landscape, 1152x2048 portrait, 1792x1024 landscape, 1024x1792 portrait, 1536x896 landscape, 896x1536 portrait, +10 |
Outputs (8)
| Name | Type | Description |
|---|---|---|
| width | INT | — |
| height | INT | — |
| scale | INT | — |
| chunk_frames | INT | — |
| chunk_overlap | INT | — |
| offload | BOOLEAN | — |
| aggressive_offload | BOOLEAN | — |
| tile_preset | STRING | — |