KLYN / EXISTING COMFYUI / MODERN IMAGE WORKFLOWS

No model weights, Python runtime or second ComfyUI installation are bundled.
In KLYN API/provider settings, choose ComfyUI and set Server URL to your existing
running server (for example http://127.0.0.1:8188). Another port is supported.

Modern model profiles:
FLUX.2 Klein, Qwen-Image 2.1, Z-Image and Custom image workflow.
Klein 4B distilled FP8 has a bundled single-reference image-edit workflow:
flux2_klein_4b_edit_api.json. Leave its workflow path empty to select it automatically.
It uses four Euler steps, CFG 1 and up to 1 megapixel. Editing is driven by the
source reference and prompt; the generic strength slider does not change denoise.
An explicitly configured custom workflow still takes precedence.
Qwen-Image 2.1 also has a built-in single-reference editing workflow:
qwen_image_2_1_edit_api.json. Leave its workflow path empty. It needs the BF16
model, qwen3vl_8b BF16 encoder, VAE and the TextEncodeQwenImage21 core node.
It uses 25 Euler steps and CFG 1. The latent returned by the reference encoder
is passed directly to the sampler to avoid a shifted edit. Custom JSON overrides it.
Z-Image and Custom image workflow require an image-editing workflow you have already
tested on that server, with its model files and nodes installed there.
For the named models, Download beside the model selector can fetch the pinned
model, text encoder and VAE into your chosen existing ComfyUI models folders.
The dialog shows the total size, progress and cancellation; matching existing
files are reused after SHA-256 verification. Custom image workflow has no fixed
model bundle. Downloading weights does not install nodes, configure the workflow
or establish that the model will fit your GPU.

1. In ComfyUI, export the working graph using Save (API Format).
2. In that API JSON, replace the positive text value with "%PROMPT%" and the
   source LoadImage filename with "%INPUT_IMAGE%". Bind these explicitly to the
   intended inputs; unrelated references are not modified by KLYN.
3. Optionally set seed to "%SEED%", denoise to "%DENOISE%", dimensions to
   "%WIDTH%"/"%HEIGHT%" and output prefix to "%OUTPUT_PREFIX%".
   KLYN substitutes exact placeholders with their proper JSON types.
4. Keep the actual installed model, text encoder and VAE names in loader inputs.
   Include PreviewImage or SaveImage for returning an image through /history.
5. Set the matching workflow file path in KLYN provider settings and select the
   corresponding ComfyUI model profile in the image model selector.

KLYN checks API format, installed node types, loader options and graph links
before uploading the scene image. It never silently sends a modern profile to
SDXL or FLUX.1. API/cloud nodes and custom nodes may have their own effects,
terms or charges: use your trusted local image workflow.

Official model setup references checked 16 October 2026:
https://docs.comfy.org/tutorials/flux/flux-2-klein
https://huggingface.co/Comfy-Org/Qwen-Image-2.1
https://docs.comfy.org/tutorials/image/z-image/z-image

LOCAL ENHANCEMENT
Enhance detail above Advanced uses the built-in SeedVR2 7B INT8 image-restoration
workflow in seedvr2_7b_enhance_api.json on the configured ComfyUI server.
It needs two files from Comfy-Org/SeedVR2, downloaded separately and verified.
Choose original size, 2x or a 4096-pixel long edge. Results are saved as a new
PNG at native enhancement size, without fitting them back to source dimensions.
Small details can change. No cloud nodes, custom nodes or API key are used.
Both Qwen and SeedVR2 require a sufficiently recent ComfyUI core. Downloading
weights does not update the server. See the release scope for actual test limits.
