Tiled Image Generator Advanced

TiledImageGeneratorAdvanced is the exact same tiling machinery as its little sibling - same JSON tile prompts, same neighbor-overlap stitching, same two outputs - except it drops every KSampler-style input and accepts the full custom sampling chain instead: noise, guider, sampler, and sigmas. It exists because Flux and friends don't sample the way SDXL does, and a stock KSampler isn't how you run them.

Here's the deal. Flux wants a BasicGuider (with a FluxGuidance on top), a KSamplerSelect, a BasicScheduler or LyingSigmaSampler, and a RandomNoise - the exact chain you'd otherwise feed a SamplerCustomAdvanced. This node takes those four objects and runs them per tile, which is the whole point. Same goes for z-image Turbo and other flow-matching models: if you're already building a custom guider/sampler/sigmas chain for your model, this is the version of the tile generator that slots into it.

What you give up: the standard node's global_positive and global_negative inputs don't exist here. There's no global style anchor appended to every tile, and no negative prompt at all - the tile prompts in json_tile_prompts are the only text the node encodes (via the clip input). If you want a universal negative or a style anchor, bake it into the tile prompts or handle conditioning upstream, because this node won't do it for you. That's the price of samplers that define their own conditioning setup.

The inputs that matter are the same grid math as the standard node: json_tile_prompts, grid_width/grid_height, tile_width/tile_height, overlap_percent (15–25% recommended), control_strength, seed, and the seamlessX/seamlessY wrap toggles. Coherence wiring is identical too: SDXL/Flux use a tile ControlNet on the optional controlnet input (remember to set the Union control type to tile or repaint via SetUnionControlNetType, or you get a checkerboard), while z-image/Qwen use a DiffSynth/Fun inpaint patch on model_patch. No controlnet or model_patch means independent tiles with visible seams.

One thing the source gets right that a naive per-tile sampler gets wrong: RandomNoise-style noise objects reuse a single stored seed, which would give every same-shaped tile bit-identical noise. The node retargets the noise seed per tile (and bumps the base seed by one per tile) so neighbors decorrelate instead of repeating.

Outputs are composite_image and individual_tiles, same as the standard node.

Install is identical to its sibling - one pack, same clone:

cd ComfyUI/custom_nodes
git clone https://github.com/rickyars/comfyui-llm-tile.git

Restart ComfyUI (or find it as "Tiled Image Generator" in ComfyUI Manager). Heavy dependencies? None beyond what your model needs - requirements.txt is just numpy, Pillow, and torch.

Honest take: if you're on SDXL or SD 1.5 and using standard samplers, you want the plain TiledImageGenerator - it has the global prompt inputs and simpler wiring, and the Advanced node's flexibility buys you nothing there. Reach for this one when your model doesn't fit a KSampler: Flux, z-image, anything where you're already assembling a guider/sampler/sigmas chain. Same tile prompts, same seams, different sampling guts.