I just finished an ambitious project building out a unique Krea2 tiled diffusion approach.
The 'secret sauce' is using the qwen3/krea2's vision encoding ability and apply it tile by tile rather than the whole image - giving extremely good cohesion and adherence with the right settings (denoise 0.1-0.2). No text prompt is needed, but another unique ability - you can provide PER TILE text prompt in addition to global - only rely on the image itself with no text - or rely on text only, global and/or per tile.
You'll need to install my custom nodes (git clone in your custom nodes folder:
https://github.com/RandyHaylor/krea2-tiled-diffusion-comfyui
=== Krea 2 Tiled Diffusion ===
Tiled high-res detail pass for Krea 2
- Two custom nodes
Excellent adherence to the original
- Each tile is shown the region it is about to draw
- Detail is added, composition is not reinvented
- vision_weight 1.0 keeps it close
* Up to 8.0 tightens it further
* Both usable
No ControlNet, no tile preprocessors, no seam repair
- No masking, no regional conditioning nodes
- No second model
- Load image, upscale, encode, sample
* Two of our nodes and ComfyUI's own
Tiles recombine inside each sampling step
- Others finish each tile, then blend
* Tiles already decided content independently
* Blending cannot reconcile them
- Here all tiles run the same step
* Recombined under a raised cosine weight before it returns
* Neighbours cannot diverge
Each tile is shown only its own region
- The encoder looks at that tile's crop, not the full image
- Show a tile the whole picture and it draws the whole picture
* Usual cause of duplicated subjects
Prompt need not describe the image
- Subject comes from what the encoder sees
- Complete working prompt for a figure, two katanas, footwear, text blocks
* a black and white image in a high quality drawing style
- Global prompt applies to EVERY tile
* Keep to medium, palette, style
- Nine optional per-tile fields
* Text in one field affects only that tile
Memory scales with the tile, not the canvas
- Model never given the full latent
- Bigger output = bigger grid, not bigger card
- 3x3 reaches high resolutions on cards that cannot hold the whole latent
Latent in, latent out
- Latent must already be at target resolution
- Neither node resizes, crops, or uses a VAE
- Inserts into a hires chain with no pixel round trip
=== Settings ===
Detail pass
- 2x2, 256px overlap, 4 steps, 0.10 denoise
Rebuild
- 2x2, 256px overlap, 8 steps, 0.75 denoise
Reduced VRAM
- 3x3
All three
- euler, simple, cfg 1
=== Models ===
Tested with
- krea2TurboRawINT8 — INT8, approx. 12 GB
- qwen3vl4bInt8W4a8_int8convrot
Also suitable
- Smaller quantised qwen3vl_4b text encoders
- Smallest v1.2 identity edit LoRA rank
=== Links ===
Required custom nodes
- https://github.com/RandyHaylor/krea2-tiled-diffusion-comfyui
Description
FAQ
Comments (5)
Very cool! So happy to have a simple tiled diffusion workflow for krea2.
thank you! Appreciate you sharing!
K2 can directly generate highly detailed 2K images; instead of using the traditional method (2 sampling) to add detail, you utilized an I2I approach—leveraging Qwen to re-describe the scene and perform tiled, region-based redrawing?
This uses krea2-image-edit and a custom node to gather and encode vlm tokens for each planned tile region to achieve very high resolutions or use less vram for the higher resolutions k2 does support. It was a fair bit of testing and work and came together well. Also you can add custom tags per tile - something I came up with. The tiled diffusion is also a hand rolled process, and it uses the overlap plus per tile vlm tokens to achieve accurate high detail at resolutions past 2k
I've found that this can add even finer small details - and more detail with more coherence than simply pushing a single high res image - there is a speed cost, but nice to have access to hgh quality >2k gens on a consumer gpu