feat: Add tiled multi-diffusion upscaling support for FLUX and Z-Image models

docs/features/upscaling.md

@@ -0,0 +1,46 @@
+---
+title: Tiled Upscaling (Multi-Diffusion)
+---
+
+# Tiled Upscaling (Multi-Diffusion)
+
+Invoke's Upscale tab uses a tiled multi-diffusion pipeline to enlarge images while adding fine detail. It works with **SD1.5 / SDXL**, **FLUX**, and **Z-Image** models.
+
+## How It Works
+
+The pipeline runs in five stages:
+
+1. **Spandrel upscale** — A neural-network upscaler (e.g. RealESRGAN, SwinIR) scales the image to the target resolution.
+2. **Sharpening** — An unsharp mask refines edges introduced by the upscaler.
+3. **Encode to latents** — The sharpened image is encoded via the model's VAE.
+4. **Tiled multi-diffusion denoise** — The latent is split into overlapping tiles. Each tile is denoised independently, then tiles are blended back together with gradient weights for seamless results. A **Tile ControlNet** guides each tile to preserve the original structure.
+5. **Decode** — The refined latents are decoded back to a full-resolution image.
+
+## Required Models per Architecture
+
+| Component | SD1.5 / SDXL | FLUX | Z-Image |
+|---|---|---|---|
+| **Main model** | Any SD1.5 or SDXL checkpoint | FLUX.1 Dev / Schnell | Z-Image model |
+| **Spandrel upscaler** | Any Spandrel model | Any Spandrel model | Any Spandrel model |
+| **Tile ControlNet** | SD1.5 or SDXL Tile ControlNet | FLUX Tile or Union ControlNet | Z-Image Tile ControlNet |
+| **Text encoder** | Built-in CLIP | T5 Encoder + CLIP Embed | Qwen3 Encoder (standalone or via Qwen3 Source) |
+| **VAE** | Built-in / model default | FLUX VAE | FLUX VAE |
+
+## Parameters
+
+- **Scale** — Target scale factor for the upscale (e.g. 2× or 4×).
+- **Creativity** — Controls how much new detail the denoiser adds. Higher values produce more creative (but less faithful) results.
+- **Structure** — Controls how strongly the Tile ControlNet preserves the original composition. Higher values keep the image closer to the input.
+- **Tile Size** — Size of each tile in pixels. Larger tiles use more VRAM but may produce more coherent results.
+- **Tile Overlap** — How much adjacent tiles overlap, in pixels. More overlap improves blending at the cost of speed.
+
+## VRAM Requirements
+
+Tiled upscaling requires the full model plus enough VRAM for a single tile's denoising pass. On a 24 GB GPU, all three architectures work comfortably. Low-VRAM mode (`enable_partial_loading: true` in `invokeai.yaml`) can help on smaller GPUs — see [Low-VRAM mode](low-vram.md).
+
+## Tips
+
+- Start with the default **Creativity** and **Structure** settings, then adjust to taste.
+- Use a descriptive positive prompt — the denoiser uses it to guide detail generation in each tile.
+- If you see visible tile seams, increase **Tile Overlap**.
+- For maximum quality, use a high-quality Spandrel upscaler as the initial upscale step.