Daruniawaesf 4K–16K Upscale Context-Tile Generator — V-1.0 Relase Krea2 — V-1.1 SDXL

Author notes from daruniawaesf4K16K_v11SDXL.json · selected release

## Central controls - **Control After Generate:** the dedicated Legacy-safe core controller provides fixed, randomize, increment, or decrement. - **HR Conditioning Mode:** Detail Only (default), Global Only, or Global + Detail. - **Refinement Profile:** Fidelity (default), Balanced, or Creative. - **Fidelity denoise:** 0.22 / 0.16 / 0.10 / 0.07. - **Balanced denoise:** 0.26 / 0.20 / 0.14 / 0.09. - **Creative denoise:** 0.32 / 0.26 / 0.18 / 0.12; matches V1 - **Per-pass steps:** Base / 2K / 4K / 8K / native 16K remain independent. - **Fast Mode:** controls step overrides only. - **SageAttention:** optional, auto kernel, compile off, disabled by default. - **Tiled Base VAE Decode:** affects the initial latent decode only. - **Q3 + Memory Saver:** 8K generative result → RTX VSR 16K; recommended. - **Q3 without Memory Saver:** native 16K context diffusion; high RAM. ## Output safety The full result is saved once as lossless PNG. PNG Compression changes encoding time and file size, never image pixels. Save off hard-bypasses output and temporary PNG writing. The canvas preview receives a 12.5% copy. ## Adaptive Context controls - **Legacy Fixed (default):** V1-compatible 64/64 on 2K, 4K, 8K, and 16K. - **Adaptive Balanced:** 64/64 → 64/64 → 64/96 → 96/128. - **Adaptive Seam-Safe:** 64/96 → 64/128 → 96/160 → 128/192. - Format is anchor/overlap. Tile dimensions remain 1280×1280. - Choose Balanced for general work; use Seam-Safe for smooth gradients or visible tile transitions. - Square 2K/4K/8K/16K tile counts: Legacy 4/16/64/256; Balanced 4/16/81/400; Seam-Safe 4/25/121/676. - The selected profile and all resolved values are written into embedded HR pipeline metadata.

Author notes from daruniawaesf4K16K_v11SDXL.json · selected release

# Simplified Front Panel • V1.1 For T2I, enter prompts and keep **I2I Mode=false**. For I2I, load an image and enable **I2I Mode**. **HR Conditioning Mode** defaults to **Detail Only**, so tile passes receive refinement language without replaying the full scene narrative. **Refinement Profile** defaults to **Fidelity**. Global Only and Global + Detail remain available, while Creative reproduces the V1 denoise schedule. Every HR mode uses one selected text, one CLIP encode, and one HR guider; no DualCFG diffusion branch is added. Base dimensions are proportionally normalized to a 1024 px long side, every final edge is capped at 16384 px, and output is lossless PNG. This workflow uses the Classic/Legacy renderer and does not require Nodes 2.0 beta. ## Required custom nodes - [Context-Anchored Tile Refine](https://github.com/Blakeem/ComfyUI-ContextAnchoredTileRefine) - [ComfyUI-DaSiWa-Nodes](https://github.com/darksidewalker/ComfyUI-DaSiWa-Nodes) - [rgthree-comfy](https://github.com/rgthree/rgthree-comfy) - [ComfyUI-KJNodes](https://github.com/kijai/ComfyUI-KJNodes) ## Optional acceleration dependency - [SageAttention](https://github.com/thu-ml/SageAttention) — required only when **SageAttention=true** ## Adaptive Context Profile • V1.1 **Legacy Fixed** is the default and preserves V1 exactly: every generative pass uses a 1280×1280 tile with anchor 64 and overlap 64. **Adaptive Balanced** increases context economically: 2K 64/64, 4K 64/64, 8K 64/96, 16K 96/128. **Adaptive Seam-Safe** uses 2K 64/96, 4K 64/128, 8K 96/160, and 16K 128/192 for smooth skies, fog, gradients, and long cross-tile structures. Values are shown as **anchor/overlap** in pixels. With the current grid solver, square 2K/4K/8K/16K workloads are approximately Legacy 4/16/64/256 tiles, Balanced 4/16/81/400, and Seam-Safe 4/25/121/676. Non-square images use fewer total tiles. The maximum tile remains fixed at 1280×1280, so peak model-facing tile dimensions do not increase.

Author notes from daruniawaesf4K16K_v11SDXL.json · selected release

## DARUNIAWAESF High-Resolution Engine • V13 V13 selects Detail Only, Global Only, or Global + Detail before one HR CLIP encode. Fidelity, Balanced, and Creative profiles route deterministic per-pass denoise values without adding a second positive diffusion prediction. Checkpoint loading, optional SageAttention, sampling, context refinement, VAE routing, RTX processing, hard-save gating, and metadata remain inside this Legacy-compatible backend subgraph. **The 4K/8K/native 16K paths remain generative context-tile diffusion, not raw interpolation.** V13 adds one Legacy-safe CustomCombo and deterministic integer routing for per-pass context anchor and overlap. It does not add a sampler, guider, CLIP encode, diffusion pass, or model branch. Legacy Fixed reproduces the V12 64/64 geometry exactly.

Author notes from daruniawaesf4K16K_v11SDXL.json · selected release

**16K Memory Saver=true:** Q3 keeps the generative chain through 8K, skips the 16K image-space diffusion canvas, and forces one RTX VSR pass to a maximum 16K long side. This is the recommended low-memory path. **16K Memory Saver=false:** Q3 performs native 16K context diffusion and requires high system RAM. Optional RTX remains available, but its scale is automatically capped so no final edge exceeds 16384 px.

Author notes from daruniawaesf4K16K_v11SDXL.json · selected release

The saver is locked to lossless PNG so exact 16384 px outputs are valid and workflow metadata remains embedded. Compression uses the balanced level-3 mapping. When `Save PNG to Output=false`, a lazy empty-image branch makes the saver exit before encoding, so no output or temporary PNG is written; the downscaled front-end preview remains available.