Latent Tile Sampler Custom (Mikey)
Tiled denoising for people building a custom sampler graph
- model
- positive
- negative
- sampler
- sigmas
- latent_image
- samples
This one's for a different crowd than Mikey Sampler Tiled or its Advanced sibling. Those two are all-in-one convenience nodes with dropdowns for sampler names and schedulers. This node is the raw building block instead - it plugs into the world of SamplerCustom-style graphs, where you've already got a KSamplerSelect feeding a SAMPLER type and a scheduler node (like BasicScheduler) feeding a SIGMAS type, and you want tiled denoising bolted directly into that existing custom sampling setup rather than tacked on as a separate post-process step.
If you're not already building graphs that way, you probably don't need this node - reach for Mikey Sampler Tiled or the Advanced version instead, which handle the sampler/scheduler selection for you with friendlier dropdowns. But if you're the kind of person experimenting with custom sigma schedules, alternative samplers, or a hand-tuned SamplerCustomAdvanced chain, this node lets you keep all of that while still getting tile-based sampling on a latent, rather than having to abandon your custom setup to use one of the packaged Mikey samplers.
The inputs that matter: model, positive, and negative are standard. sampler takes a SAMPLER-typed input - not a dropdown string, an actual typed connection, meaning it has to come from something like KSamplerSelect. Same story for sigmas, which needs a SIGMAS-typed input from a scheduler node rather than a plain number. latent_image is your starting latent. add_noise (boolean, default true) and noise_seed control noise injection the same way KSampler Advanced does. cfg is your classifier-free guidance scale (default 8). And tile_size (256–4096, default 1024, step 64) is the one setting unique to this node - it should roughly match a resolution your model is comfortable denoising at: 512 for SD 1.5-family models, 1024 for SDXL. The single output is samples - a LATENT, ready for a VAE Decode or further processing.
Installing it
Ships with the full Mikey Nodes pack. Through ComfyUI Manager, search "Mikey Nodes." Manually:
cd ComfyUI/custom_nodes
git clone https://github.com/bash-j/mikey_nodes
Restart ComfyUI. No extra models required for this node - it's a sampling-loop implementation, not something that needs downloaded weights of its own.
Common issues
The single biggest trip-up is trying to feed sampler or sigmas a plain value instead of a proper typed connection - these are strict types in ComfyUI, and a beginner coming from KSampler (where sampler_name and scheduler are just dropdown strings) often doesn't realize they now need a KSamplerSelect node and a scheduler node (BasicScheduler, KarrasScheduler, etc.) upstream to produce those connections. If ComfyUI is complaining about a missing or mismatched input type on this node, that's almost always it.
The other one is tile_size mismatched against your model family - set it too small relative to what the model was trained at and the tiling seams become visible or detail looks chopped up; set it larger than the model comfortably handles and you may not get any real tiling benefit at all. When in doubt, match it to your model's native resolution rather than guessing.
Inputs (10)
| Name | Type | Default | Description |
|---|---|---|---|
| model | MODEL | — | |
| add_noise | BOOLEAN | true | — |
| noise_seed | INT | 00–18446744073709550000 | — |
| cfg | FLOAT | 8.00–100 | — |
| positive | CONDITIONING | — | |
| negative | CONDITIONING | — | |
| sampler | SAMPLER | — | |
| sigmas | SIGMAS | — | |
| latent_image | LATENT | — | |
| tile_size | INT | 1024256–4096 | — |
Outputs (1)
| Name | Type | Description |
|---|---|---|
| samples | LATENT | — |