Download
1 variant available
This checkpoint includes a config file, download and place it along side the checkpoint.
Initial relase

100 1 2 3 4 5 6 7 8 90 1 2 3 4 5 6 7 8 9
230 1 2 3 4 5 6 7 8 90 1 2 3 4 5 6 7 8 9

(The other comparison images are at the end of the post.)
Daruniawaesf 4K–16K Context-Tile Generator
A modular high-resolution ComfyUI workflow designed to generate coherent 4K, 8K, and 16K images through staged context-aware diffusion refinement instead of relying only on conventional pixel upscaling.
The workflow provides a simplified front panel for prompts and essential settings, while model loading, sampling, context-tile refinement, VAE routing, optional acceleration, metadata generation, and output processing remain organized inside a backend subgraph.
Created by Daruniawaesf.
## V-1.0 Initial Relase Highlights
V-1.0 redesigns the high-resolution conditioning system to reduce object duplication, semantic drift, repeated subjects, and composition changes during tile refinement.
### HR Conditioning Modes
Controlled directly from Master Settings:
- Detail Only — default: Sends only the dedicated high-resolution refinement prompt to the tile passes. Recommended for preserving the base composition.
- Global Only: Reuses the global positive prompt during refinement.
- Global + Detail: Combines both prompts and reproduces the more creative behavior
Only one selected conditioning text is encoded before one HR guider. The workflow does not add a second positive diffusion prediction or a DualCFG branch.
### Refinement Profiles
- Fidelity — default: Maximum structural and compositional stability.
- Balanced: Moderate detail generation with controlled variation.
- Creative: Stronger reinterpretation and denoise schedule.
Profile denoise schedules:
| Profile | 2K | 4K | 8K | 16K |
|---|---:|---:|---:|---:|
| Fidelity | 0.22 | 0.16 | 0.10 | 0.07 |
| Balanced | 0.26 | 0.20 | 0.14 | 0.09 |
| Creative | 0.32 | 0.26 | 0.18 | 0.12 |
Fast Mode changes step counts only and does not silently modify the selected refinement profile.
## Two Workflow Variants
### Main Checkpoint Version
Uses ComfyUI's standard checkpoint loader and supports compatible checkpoint-based models. The selected checkpoint supplies its own model, text encoder, and VAE components.
### Krea2 Version
Designed for Krea2-compatible diffusion models.
Required components:
- Model: Any compatible Krea2 model
- Text encoder: qwen3vl_4b
- Model type: krea2
- VAE: qwen_image_vae
The Main and Krea2 editions share the same interface, refinement profiles, HR conditioning modes, context-tile pipeline, output controls, and visual organization.
## Resolution Pipeline
The 4K, 8K, and native 16K paths use generative context-tile diffusion refinement. They are not simple raw interpolation passes.
For systems with limited memory:
- Enable Q3
- Enable 16K Memory Saver
- The workflow performs generative refinement through 8K and then uses one capped RTX VSR pass to reach 16K.
For systems with sufficient RAM and VRAM:
- Enable Q3
- Disable 16K Memory Saver
- The workflow performs native 16K context-tile diffusion refinement.
Native 16K processing can require a very large amount of system RAM and VRAM. Hardware requirements vary by checkpoint, aspect ratio, tile settings, and model architecture.
## Main Controls
- T2I and I2I modes
- Base width and height
- Base seed
- Control After Generate: fixed, randomize, increment, or decrement
- Sampler and scheduler
- Base, 2K, 4K, 8K, and 16K pass steps
- HR Conditioning Mode
- Refinement Profile
- Fast Mode
- Optional second and third quality stages
- Tiled or regular base VAE decoding
- Optional SageAttention acceleration
- Optional RTX processing
- 16K Memory Saver
- Lossless PNG output
- Configurable PNG compression
- Embedded workflow and generation metadata
## Output and Metadata
The final image is saved as lossless PNG with embedded workflow information.
PNG compression changes encoding time and file size only. It does not reduce image quality.
Generation metadata records the selected checkpoint or Krea2 UNET, HR conditioning mode, refinement profile, sampling configuration, pass settings, and resolution pipeline.
A reduced front-panel preview is used to avoid sending the complete 8K or 16K image to the interface unnecessarily.
## Compatibility
This release uses the Classic/Legacy ComfyUI renderer.
ComfyUI Nodes 2.0 beta is not required or recommended for this release.
The workflow retains the external Legacy seed controller so Control After Generate remains available without converting the Master Settings subgraph to Nodes 2.0.
## 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
- [SageAttention](https://github.com/thu-ml/SageAttention)
SageAttention is required only when the SageAttention option is enabled in Master Settings. It is disabled by default because compatibility and performance gains depend on the GPU, PyTorch environment, model architecture, and attention implementation.
## Recommended Starting Settings
For maximum scene preservation:
- HR Conditioning Mode: Detail Only
- Refinement Profile: Fidelity
- Tiled Base VAE Decode: Enable when normal VAE decoding exceeds available VRAM
- 16K Memory Saver: Enable on systems that cannot process native 16K diffusion
- SageAttention: Test per system before using it as a permanent default
Results and resource requirements depend heavily on the selected checkpoint. No workflow can completely prevent semantic drift when denoise strength, prompt conditioning, or tile refinement settings are pushed beyond the selected model's capabilities.
## License Notice
The permissions selected on this Civitai resource apply to the original workflow JSON, layout, configuration, and release files created by Daruniawaesf.
Third-party checkpoints, Krea2 models, custom nodes, SageAttention, and other dependencies remain subject to their respective licenses.
-----------------------------------------------------------------------------------------------------------------------------------------





