ComfyUI_MiraSubPack

Tile by pixel budget, not by fixed size — the VRAM-first way to crop

The front door of the Mira tiled upscaler — crop a big image into overlap-feathered tiles

Stitch two images that don't quite line up — SIFT alignment plus a clean blend

Sample every tile with its own prompt — the KSampler that runs the tiled loop

Caption every tile through a vision LLM, and feed the result straight back into the sampler

JoyCaption Beta One with the training wheels off — one prompt box, that's it

JoyCaption Beta One, run through your own llama.cpp — the dataset-captioning workhorse

Upscale your latents properly, then tile them — the 'advanced' node that earns the name

Pick a target size, set a VRAM budget — this node does the math and the downscaling

Pack the tile-grid numbers into the descriptor the Mira nodes actually eat

Unpack a tile-grid descriptor into plain numbers you can read and wire

Stitch tiled upscales back together without the seam lines

Merge tiled latents in latent space, so you never leave it for a stitch

Caption images with Qwen3VL without loading a single model into ComfyUI

The text encoder QwenImage-Edit needs, without hunting for the right fork

Multi-image QwenImage-Edit encoding, with the reference-latent knobs the pros use

Kill the seam lines — pull every sampled tile back toward the source's colors

The other half of the VRAM fix — decode tiled latents one at a time

A VAEEncode that won't OOM on your batch — encode tiles one at a time