ComfyUI Node
GPT Image Latent Refiner
Reduces dot noise, stippling, grime, and repeating micro-textures with an independently trained latent residual model. Qwen is the recommended default; Flux2 preserves more source detail; SDXL changes details more strongly but may work better for photorealistic portraits.
GPT Image Latent Refiner
- image
- image
◄profileqwen►
◄strength1.00►
◄deviceauto►
◄tile_vaetrue►
CategoryGPT Image/refinement
Inputs (5)
| Name | Type | Default | Description |
|---|---|---|---|
| image | IMAGE | — | |
| profile | COMBO | qwen | qwen: recommended balance; flux2: stronger source preservation; sdxl: stronger reconstruction drift, but may work better for photorealistic portraits. |
| strength | FLOAT | 1.000–2 | 0 bypasses the learned correction; 1 is the trained strength. |
| device | COMBO | auto | 3 options: auto, cuda, cpu |
| tile_vae | BOOLEAN | true | Use VAE tiling to reduce peak VRAM on large images. |
Outputs (1)
| Name | Type | Description |
|---|---|---|
| image | IMAGE | — |