The nearest->bilinear wash fix (5f61727) changes what palette-wash recipes
render, so every wash-affected demo is re-rendered through the live V10
nodes on the same seeds: the palette sphere is now a smooth gradient, the
toward-lamp an organic rose/sage piece, the final-only timing lamp a clean
two-tone split. Guide prose that described the old mosaic aesthetic is
reworded to describe the new renders, the V10 paper's treatment row and the
V9 paper's wash pseudocode now document the bilinear upsample (with a note
on why nearest was wrong), and the focus-validation reference images join
the recipe-lab refs.
ComfyUI Krea Reference
Give each reference image a job before it reaches Krea.
Krea Reference is a small ComfyUI custom-node package for directing Krea 2 with multiple image references. Instead of one source image vaguely influencing the whole result, each image gets a clear role: preserve the subject, borrow a visual style, copy lighting, suggest a material, follow a layout, or avoid copying text and logos.
The demo images above were generated with the included workflows. Each individual PNG in docs/assets/krea-v9/demos has ComfyUI workflow metadata embedded, so you can drag a demo PNG into ComfyUI to inspect the exact setup.
What It Helps You Do
| Goal | How to set it up |
|---|---|
| Put image 2's style onto image 1 | Image 1: keep the same subject at 0.80. Image 2: suggest the visual style at 0.45 to 0.65. |
| Keep a product or character recognizable | Use keep the same subject for the identity image, then add lower-strength style, lighting, or material cards. |
| Borrow lighting without copying the subject | Use copy lighting and mood around 0.30 to 0.45. |
| Borrow a surface or finish | Use suggest material or texture around 0.25 to 0.40. |
| Follow a pose, crop, or composition | Use copy pose and layout or copy big shapes only at gentle to medium strength. |
| Use a reference that contains text or logos | Add avoid copying text/logos at very low strength, usually around 0.03. |
How The Workflow Thinks
Each guide card answers two questions:
- What should Krea borrow from this image?
- How strongly should this image guide the result?
The stack encoder combines the written prompt and all connected guide cards:
Load Image -> KG Krea 2 Image Guide Card V9 -> KG Krea 2 Reference Stack Encoder V9 -> KSampler positive input
That makes multi-reference workflows easier to reason about. A product image can be responsible for the subject, an abstract image can be responsible for style, a third image can be responsible for lighting, and a fourth can protect against text/logo copying.
Included Nodes
| Node | Purpose |
|---|---|
KG Krea 2 Image Guide Card V9 |
Describes one reference image. Choose a recipe or use manual tuning. |
KG Krea 2 Reference Stack Encoder V9 |
Combines the final prompt and up to 12 guide cards into Krea conditioning. |
KG Krea 2 Image Guide Card V10 |
The V9 card plus four more recipes, guide direction, per-card timing, and layer dials. |
KG Krea 2 Reference Stack Encoder V10 |
The V9 stack plus balance, study reuse, a stack report, and a prepared-reference preview. |
The nodes expose plain-language controls for prompt strength, image strength feel, image detail level, framing, timing, and text/logo guard behavior.
What V10 Adds
V10 extends the same architecture without touching V9; the two versions cross-connect, and saved V9 workflows keep working unchanged.
| V10 control | What it does |
|---|---|
| Four more quick recipes | suggest the color palette, use the background/setting, copy the camera framing, and mood board only - jobs that previously required manual tuning. |
| Your own recipes | Drop schema-validated YAML/JSON files into custom_recipes/ and they appear in Use image for as first-class choices. |
| Aspect-focus recipes | A recipe's focus field names which aspect of its image to study - "the clothing and garment style, not the person" - so recipes can target clothing, props, or other object-bound aspects. |
| Starter recipes included | Three ready-made custom recipes load out of the box: borrow the weather, borrow the clothing style, and cinematic color grade. |
| Recipe Builder | web/recipe-builder.html (also served at /extensions/<pack folder>/recipe-builder.html): three plain-language questions become a validated recipe file - no schema knowledge needed. |
Guide direction |
A card can steer away from its image: a counter-example for a palette, composition, subject, or style you do not want. |
When this card guides |
Per-card timing: whole image, early layout only, or final details only. |
Structure/Finish layers pull |
Manual-mode dials over the structure-vs-finish conditioning layers. |
Balance strong cards |
Keeps several simultaneously hot cards from fighting by budgeting their total departure. |
Reuse image studies |
Caches image studies by content, so strength and timing tweaks re-run without any encoder passes. |
stack_report output |
A plain-language account of what every card requested, what it got, and why. |
prepared_references output |
A contact sheet of exactly what the vision encoder studied after treatments. |
The V10 demo images above were generated with the V10 recipes and have the matching V10 workflow embedded - drag one into ComfyUI to inspect the exact setup. Each demo's full journey (input images, settings, prompt, result) is documented in the V10 user guide.
Per-node details: V10 guide card and V10 stack encoder. Try krea-v10-full-showcase-workflow.json for all of it in one graph.
Install
Clone this repo into your ComfyUI custom_nodes directory and restart ComfyUI:
cd /path/to/ComfyUI/custom_nodes
git clone https://github.com/kgilper/krea-reference.git
This repo does not include model weights. Use your own Krea-compatible model, CLIP, and VAE files in ComfyUI.
Try It In Five Minutes
- Copy example_assets/krea-reference-examples into your ComfyUI
input/folder. - Load example_workflows/krea-v9-full-showcase-workflow.json.
- Queue it once with the included synthetic reference images.
- Replace the Load Image nodes with your own test images.
- On each guide card, choose
Use image forand adjustHow strongly this image guides.
The full showcase demonstrates five roles:
| Slot | Recipe | Starting Strength |
|---|---|---|
| Content anchor | keep the same subject |
0.80 |
| Visual style | suggest the visual style |
0.65 |
| Material/texture | suggest material or texture |
0.35 |
| Lighting/mood | manual tuning with lighting controls |
0.45 |
| Text/logo guard | avoid copying text/logos |
0.03 |
Recipe Examples
These examples show the effect of common guide-card recipes. Open the linked PNG or drag it into ComfyUI to load the embedded workflow.
Example Workflows
| Workflow | Best for |
|---|---|
| krea-v10-full-showcase-workflow.json | First V10 run. Six cards including the new palette, environment, and framing recipes, per-card timing, gentle balance, and both feedback outputs. |
| krea-v10-counter-example-workflow.json | The V10 away from this image direction: keep the subject, push a style out. |
| krea-v10-reference-stack-workflow.json | Compact V10 starter graph with the report and prepared-references previews wired. |
| krea-v10-starter-recipe-workflow.json | A shipped custom recipe in action: cinematic color grade (from the auto-loaded starter pack) grading the scene from a style reference. |
| krea-v9-full-showcase-workflow.json | First V9 run. Shows content, style, material, lighting, and text/logo guard cards together. |
| krea-v9-no-prompt-style-transfer-workflow.json | Applying image 2's style to image 1 with no written prompt. |
| krea-v9-reference-stack-workflow.json | Compact starter graph for building your own multi-reference workflow. |
The examples intentionally avoid LoRA, model-enhancer, and switch-node plumbing so the Krea reference nodes are easy to inspect.
Good Starting Values
| Strength | Meaning | Good uses |
|---|---|---|
0.00 |
Off | Keep a card connected but inactive. |
0.03 to 0.08 |
Tiny nudge | Text/logo guard, shape hints, stubborn prompts. |
0.10 to 0.25 |
Gentle guidance | Layout, material, mood-board influence. |
0.25 to 0.45 |
Strong guidance | Lighting, texture, pose, broad style. |
0.50 to 0.90 |
Very strong | Content anchor or deliberate style transfer. Watch for over-copying. |
Tips:
- Start strengths low and raise slowly.
- Use
0.80only when the main subject must stay stable. - Lower
Image detail levelif a style image starts copying the wrong subject. - Raise
Written prompt strengthwhen the text prompt should win over references. - Use text/logo guard whenever a reference includes readable marks.
Documentation
| Start here | What it contains |
|---|---|
| V10 visual HTML guide | Visual walkthrough of everything V10 adds: new recipes, direction, timing, balance, reuse, and the feedback outputs. |
| V10 Markdown user guide | Same V10 guide in Markdown form for GitHub reading. |
| V9 visual HTML guide | Full visual walkthrough of the core recipes with embedded-workflow demo PNGs. |
| V9 Markdown user guide | Same guide in Markdown form for GitHub reading. |
| Documentation landing page | Short navigation by task. |
| Node documentation index | Per-node input and output details. |
| V9 technical paper | How and why the nodes work: architecture, math, verification, extension and porting guides. |
| V10 technical companion | The V10 mechanics on top of that architecture: direction math, balance, the study cache, verification, limitations. |
| Example workflows | What each bundled workflow is for. |
| Testing guide | Maintainer checks for contract tests and workflow validation. |
Development Checks
The contract tests run without launching ComfyUI:
python -m unittest discover -s tests -p "test_*.py" -v
python -m compileall -q kg_krea_v9 kg_krea_v10 __init__.py
License
MIT. See LICENSE.








