Nodes/civitai-comfy-nodes/flux2Klein / editImage
ComfyUI Node

flux2Klein / editImage

The cheap way to edit with Flux 2

By civitai·Created 2 months ago·Updated about a month ago· 42
flux2Klein / editImage
  • loras
  • images
  • api_config
  • images
  • errors
  • workflow_id
  • raw_json
prompt
output_format
image_metadata
width1024
height1024
seed0
quantity1
cfg_scale5.00
steps20
sample_method
schedule
negative_prompt
model_version4b

Klein unified generation and editing into one small model, and this node exposes the editing half on Civitai's cloud. You hand it an image plus a plain-language instruction and get back an edited image - no strength slider to fight, no separate edit checkpoint to download, no VRAM on your end. Because it's Klein rather than Dev, it's also the budget way to instruction-edit with Flux 2: the same "change this about the picture" job for a fraction of the 32B price.

The honest tradeoff is quality versus cost. Community testing settled on Klein as the default local editor for most of 2026 - it does object removal, garment swaps, relighting and re-posing well enough that Dev's edge only matters for the hardest scenes. So this node is less "the best editor on the cloud" and more "the best editing value on the cloud": pick 4b for the cheapest pass, 9b when the edit needs more headroom.

How it works

Standard pack client: submit engine: sdcpp, ecosystem: flux2Klein, operation: editImage with your images, prompt, and modelVersion, long-poll, get the edited IMAGE out.

The inputs that matter

  • prompt (required) - the edit instruction. Klein's Qwen encoder handles natural language well; be specific about what changes.
  • images (optional, IMAGE) - the source(s). Optional in the schema, essential in practice: an edit with no image is just a generation you'll be billed for.
  • model_version - 4b / 4b-base / 9b / 9b-base, default 4b. Distilled variants for editing; base versions are training foundations.
  • steps / cfg_scale - 20 / 5 defaults. Edits often land fine at fewer steps - and fewer steps means less Buzz.

Install

Pack install: ComfyUI Manager → search "Civitai Comfy Nodes" (publisher civitai) → Install → restart. Or clone https://github.com/civitai/civitai-comfy-nodes.git into custom_nodes, pip install -r civitai-comfy-nodes/requirements.txt (just requests), restart.

Common issues

Credentials, as with every node here: Civitai Auth node into api_config or the CIVITAI_API_TOKEN env var (from civitai.com/user/account); headless installs need the env var.

Two stumbles to avoid. First, forgetting images - the "edit" quietly becomes a fresh generation. Second, expecting the 4B to match the 9B on complex edits: 4B is fast and cheap but its anatomy and detail fall apart on hard instructions, which is exactly when you step up to 9b - and pay for it.

Klein's biggest editing weakness (shared with the whole family) is that the model re-emits the whole frame, so pixels you didn't mention drift a little. If you need surgical, unmoving detail, you'll be layering masks around this regardless - but for most "change the X" jobs it just works.

CategoryCivitai/Image/flux2Klein

Inputs (16)

NameTypeDefaultDescription
promptSTRING
output_formatoptCOMBO4 options: , jpeg, png, webP
image_metadataoptSTRINGExternal metadata that will be stored with the image
widthoptINT1024512–2048
heightoptINT1024512–2048
seedoptINT00–4294967295
quantityoptINT11–4
cfg_scaleoptFLOAT5.001–20
stepsoptINT204–50
sample_methodoptCOMBO16 options: , euler, heun, dpm2, dpm++2s_a, dpm++2m, +10
scheduleoptCOMBO12 options: , simple, discrete, karras, exponential, ays, +6
negative_promptoptSTRING
lorasoptCIVITAI_LORAS
model_versionoptCOMBO4b4 options: 4b, 4b-base, 9b, 9b-base
imagesoptIMAGE
api_configoptCIVITAI_CONFIGOptional Civitai Auth connection; defaults to CIVITAI_API_TOKEN or stored OAuth login.

Outputs (4)

NameTypeDescription
imagesIMAGE
errorsSTRING
workflow_idSTRING
raw_jsonSTRING