ComfyUI Node
RelayVSR • X2 / X4 Settings
A ComfyUI node in RelayVSR with 21 inputs and 1 output.
RelayVSR • X2 / X4 Settings
- loop_experiment
- settings
◄variantX4 original (recommended, also for 2x)►
◄lightx2v_rootauto►
◄model_dirauto►
◄python_executableauto►
◄execution_modesingle GPU►
◄gpu_ids0►
◄output_scale2►
◄compile_modelfalse►
◄warmupfalse►
◄clip_len16►
◄queue_size4►
◄codeclibx264►
◄crf17►
◄preserve_audiotrue►
◄unload_comfy_modelstrue►
◄qualityfaithful: back-projection + residual 0.85►
◄back_projection1.00►
◄residual_gain0.85►
◄downscalebicubic►
◄vramauto (pick by GPU memory)►
CategoryRelayVSR
Inputs (21)
| Name | Type | Default | Description |
|---|---|---|---|
| variant | COMBO | X4 original (recommended, also for 2x) | 2 options: X4 original (recommended, also for 2x), X2 native |
| lightx2v_root | STRING | auto | Path to ModelTC/LightX2V checkout, or auto |
| model_dir | STRING | auto | Folder containing BOTH variant checkpoints |
| python_executable | STRING | auto | LightX2V venv Python; auto picks its .venv/bin/python |
| execution_mode | COMBO | single GPU | 3 options: single GPU, Wan block offload, two GPUs (SP) |
| gpu_ids | STRING | 0 | Single GPU 0 or, for SP, 0,1. Physical GPU indexes (CUDA_VISIBLE_DEVICES). |
| output_scale | COMBO | 2 | 2 options: 2, 4 |
| compile_model | BOOLEAN | false | — |
| warmup | BOOLEAN | false | — |
| clip_len | INT | 164–128 | — |
| queue_size | INT | 41–64 | — |
| codec | COMBO | libx264 | 2 options: libx264, libx265 |
| crf | INT | 170–51 | — |
| preserve_audio | BOOLEAN | true | — |
| unload_comfy_models | BOOLEAN | true | — |
| quality | COMBO | faithful: back-projection + residual 0.85 | Named presets set the three knobs below (and, for the Flash ones, a Flash loop on its last 4 blocks at half strength). 'custom' uses the knobs as set. Measured with the X4 model at 2x output; 'smooth' was measured with clip_len 32. |
| back_projection | FLOAT | 1.000–1 | custom only. Shrinks each output frame back to the input size and adds this much of the difference back, so shapes, thickness and brightness match the input while the added texture stays. 0 = off. |
| residual_gain | FLOAT | 0.850–3 | custom only. How much of the detail RelayVSR adds on top of a plain bicubic upscale: below 1 is more faithful and calmer, above 1 crisper with more invented texture. 1 = stock. |
| downscale | COMBO | bicubic | custom only. Filter for the X4 model's 4x result when the output is 2x. bilinear is calmer and softer; bicubic is stock. |
| vram | COMBO | auto (pick by GPU memory) | Memory/speed tier (single GPU). 24 GB+: int8 weights + int8 KV cache, ~23 % faster than bf16 with offload. 16 GB: W6A8 weights + int4 KV, fits a 16 GB card at 1280x704 in. Uses the matching wan_<mode>.safetensors next to flash.safetensors when present, else quantizes wan.safetensors on load. Drift vs full precision is invisible (44-53 dB). With Wan block offload only the KV cache is quantized; two GPUs stay full precision. |
| loop_experimentopt | RELAYVSR_LOOP_SETTINGS | — |
Outputs (1)
| Name | Type | Description |
|---|---|---|
| settings | RELAYVSR_SETTINGS | — |