ComfyUI Node
HiDiffusion Apply to MODEL (Native)
Apply HiDiffusion (RAUNet + optional MSW-MSA) to a NATIVE ComfyUI MODEL. Insert between CheckpointLoaderSimple and KSampler; everything downstream stays native. No diffusers/transformers required. Notes: (1) HiDiffusion targets 2K/4K output - at 1024px the difference is small; (2) when using ControlNet together with RAUNet, the control residuals may not match the pooled feature size - disable RAUNet in that case; (3) turning both switches OFF yields a bit-identical stock MODEL, useful as a parity baseline.
HiDiffusion Apply to MODEL (Native)
- model
- model
◄model_archauto►
◄apply_raunettrue►
◄apply_window_attnfalse►
◄aggressive_raunettrue►
◄t1_ratio-1.00►
◄t2_ratio-1.00►
◄controlnet_t2ifalse►
◄window_shiftrandom►
◄window_shift_index0►
◄aggressive_start_ratio-1.00►
CategoryHidiffusion_Pro/Native
Inputs (11)
| Name | Type | Default | Description |
|---|---|---|---|
| model | MODEL | — | |
| model_arch | COMBO | auto | auto detects a stock SDXL (9 input blocks) or SD1.5/SD2.x (12 input blocks) UNet at runtime. |
| apply_raunet | BOOLEAN | true | Resolution-Aware UNet: early denoising steps run part of the UNet at half resolution to suppress object duplication at 2K/4K output. |
| apply_window_attn | BOOLEAN | false | MSW-MSA window self-attention on 10 (SDXL) / 5 (SD1.5) transformer blocks. Upstream draws a random shift every step; usually only useful at very high resolution, so it is OFF by default (matching the old plugin's default). |
| aggressive_raunet | BOOLEAN | true | Upstream 'aggressive RAUNet' (SDXL, latent<512): pool in [16%, T1) of steps and use stride-4 downsampling in the first 16%. Leaving this ON reproduces the original project's behaviour. |
| t1_ratio | FLOAT | -1.00-1–1 | -1 = upstream auto preset, chosen from the latent size. Override with 0..1. |
| t2_ratio | FLOAT | -1.00-1–1 | -1 = upstream auto preset. |
| controlnet_t2i | BOOLEAN | false | Use upstream's text-to-image+ControlNet preset (T1=0.5) instead of the plain SDXL one (T1=0.4). |
| window_shift | COMBO | random | random reproduces upstream torch.rand(1); fixed pins one of the 4 shift candidates for reproducible output and numerical parity tests. |
| window_shift_index | INT | 00–3 | 0..3, only used when window_shift=fixed. |
| aggressive_start_ratioopt | FLOAT | -1.00-1–1 | -1 = upstream 8/50 = 0.16. |
Outputs (1)
| Name | Type | Description |
|---|---|---|
| model | MODEL | — |