ComfyUI Node
Sol-Attn Progressive Sampler
Runs low-resolution sampling, spatial latent upscaling, and high-resolution sampling inside one native sampler call. Connect a model patched by Model Sparse Attention (sol-attn) to the guider, then use this SAMPLER with SamplerCustomAdvanced.
Sol-Attn Progressive Sampler
- SAMPLER
◄sampler_nameeuler►
◄lowres_steps6►
◄lowres_scale0.50►
◄lowres_noise_mode2D orthonormal Haar LL noise downsampling►
◄latent_upsampleMinimax H3 3D►
◄upscaler_modelnone►
◄upscaler_precisioncheckpoint►
◄upscaler_temporal_chunkingtrue►
◄upscaler_layerwise_cpu_offloadfalse►
CategoryANe5s Nodes/Sol-Sampler
Inputs (9)
| Name | Type | Default | Description |
|---|---|---|---|
| sampler_name | COMBO | euler | Native ComfyUI sampler used for both resolution stages. |
| lowres_steps | INT | 61–10000 | Number of denoiser evaluations at low resolution. At least one step is reserved for the high-resolution stage. |
| lowres_scale | FLOAT | 0.500.25–1 | Set to 1.0 to run one native sampling pass without latent upscaling or a second stage. |
| lowres_noise_mode | COMBO | 2D orthonormal Haar LL noise downsampling | How high-resolution noise is mapped to the low-resolution stage. 2D orthonormal Haar LL noise downsampling requires exactly half-size spatial dimensions and preserves unit variance for iid standard Gaussian noise. |
| latent_upsample | COMBO | Minimax H3 3D | How to upscale the low-resolution denoised latent. If Minimax H3 3D is selected with upscaler_model=none, bilinear interpolation is used. Low-resolution noise uses lowres_noise_mode; masks and previews use bilinear/nearest interpolation. |
| upscaler_modelopt | COMBO | none | MiniMax H3 3D checkpoint in models/latent_upscale_models. Select none to use bilinear interpolation instead of the H3 model. |
| upscaler_precisionopt | COMBO | checkpoint | Inference dtype for the learned upscaler. |
| upscaler_temporal_chunkingopt | BOOLEAN | true | Split video latents into overlapping temporal chunks when they exceed the H3 upscaler's chunk size, then blend overlaps to reduce peak VRAM. Audio latents are not chunked. |
| upscaler_layerwise_cpu_offloadopt | BOOLEAN | false | Keep H3 upscaler weights on CPU and move each module to the GPU only for its forward pass. This reduces peak VRAM but adds CPU/GPU transfer overhead. |
Outputs (1)
| Name | Type | Description |
|---|---|---|
| SAMPLER | SAMPLER | — |