ComfyUI Node
Tiled Image KSampler with Tagger
Perform tiled image sampling with dynamic tagger-based prompts for each tile.
Tiled Image KSampler with Tagger
- model
- clip
- tiled_samples
- clip_negative
- ref_latents
- tiled_latents
◄common_positive►
◄common_negativebad quality, worst quality, worst detail, sketch►
◄tagger_text►
◄seed0►
◄steps16►
◄cfg7.0►
◄sampler_nameeuler_ancestral►
◄schedulerbeta►
◄denoise0.35►
◄modeNormal►
◄noise_boost0.00►
◄noise_injection_methodadaptive►
CategoryMira/SubPack/Image Tiled Upscaler
Inputs (17)
| Name | Type | Default | Description |
|---|---|---|---|
| model | MODEL | — | |
| clip | CLIP | — | |
| tiled_samples | LATENT | Tiled latents input from VAE. | |
| common_positive | STRING | Common positive prompt for all tiles. | |
| common_negative | STRING | bad quality, worst quality, worst detail, sketch | Common negative prompt for all tiles. |
| tagger_text | STRING | Tagger output text mapping for tiles, one line per tile. | |
| seed | INT | 00–18446744073709550000 | — |
| steps | INT | 161–100 | — |
| cfg | FLOAT | 7.00–32 | — |
| sampler_name | COMBO | euler_ancestral | 44 options: euler, euler_cfg_pp, euler_ancestral, euler_ancestral_cfg_pp, heun, heunpp2, +38 |
| scheduler | COMBO | beta | 9 options: simple, sgm_uniform, karras, exponential, ddim_uniform, beta, +3 |
| denoise | FLOAT | 0.350–1 | — |
| mode | COMBO | Normal | Normal: ignore ref_latents and noise boost. (SDXL/Z Image) Reference: use ref_latents with optional noise boost. (Flux.2) |
| noise_boost | FLOAT | 0.000–1 | Extra noise injected into ref_latents before conditioning to encourage detail generation. Perturbs the reference that guides generation, effective even at denoise=1.0. 0.0 = no boost (original behavior), 0.1~0.3 = subtle detail enhancement, 0.3~0.6 = moderate (recommended for upscale with ref_latents), 0.6~1.0 = aggressive (more creative, may deviate from original). Only effective when ref_latents is connected. |
| noise_injection_method | COMBO | adaptive | Noise injection method: uniform: Standard Gaussian noise, uniform across all regions. high_frequency: Emphasizes high-frequency detail noise, better for textures. adaptive: Adds more noise to flat/blurry regions, less to detailed areas. |
| clip_negativeopt | CLIP | — | |
| ref_latentsopt | LATENT | — |
Outputs (1)
| Name | Type | Description |
|---|---|---|
| tiled_latents | LATENT | — |