FLUX

FLUX.2 klein

FLUX.2 klein needs very few steps and is fast. It comes in 4B and 9B sizes. Besides generating, it can be used for Edit by instruction in Refine.

Required files

File	Goes in	4B	9B
Main model	diffusion_models	flux-2-klein-4b or similar	the 9B main model
VAE	vae	flux2-vae	flux2-vae
Text encoder	text_encoders	qwen_3_4b	qwen_3_8b_fp8mixed

qwen_3_4b is the same file Z-Image uses, so the asset library may label it Z-Image. The 9B main model can only be downloaded after signing in to Hugging Face and accepting its licence.

Recommended settings (distilled version)

Item	Default
Steps	4
CFG	1
Sampler	Euler
shift	9.89 (left at the default, it adjusts automatically to the resolution and steps)

Measured on an RTX 5080, the 4B version takes about 2 seconds per 1024×1024 image.

FLUX.1

FLUX.1 comes as dev, schnell and Kontext. Kontext is used for Edit by instruction in Refine.

Required files

FLUX.1 model files come in two forms, and both work:

• All-in-one files: the main model, T5, CLIP-L and VAE in a single file (the form common on model sites); put it in checkpoints. Just choose it; if matching VAE and text encoder files are in your folders, those are used first.
• Main model only: put it in diffusion_models and add these files:

File	Goes in	File name
VAE	vae	ae (shared with Z-Image)
Text encoder	text_encoders	t5xxl_fp16 or the fp8 version
CLIP-L	text_encoders	clip_l (used automatically; no need to choose it)

The generate panel’s “Encoder” slot refers to T5.

Recommended settings

Version	Steps	CFG	Sampler
dev	20	3.5	Euler
schnell	4	1	Euler
Kontext	28	2.5	Euler

For dev and Kontext, CFG is the official guidance value, usually 2–4. Each step is computed once and the negative prompt isn’t used.

Measured on an RTX 5080, the dev fp8 all-in-one file takes about 25 seconds at 896×1152 (the main model doesn’t fit entirely in 16 GB of graphics memory, so it’s streamed automatically).

Kontext file names

Kontext has the same weight shapes as ordinary FLUX.1 dev, so CoCo goes by the file name: it must contain kontext to appear in the Edit by instruction model list.

LoRA

LoRAs for both FLUX.1 and FLUX.2 klein work; CoCo identifies their architecture from the file contents.

Not supported yet

Reference images (ControlNet), the FLUX.1 Fill / Canny / Depth versions, and Chroma.

Licences

Model	Licence
FLUX.2 klein 4B	Apache 2.0
FLUX.2 klein 9B	Non-commercial licence
FLUX.1 dev, Kontext	Non-commercial licence

See each model’s page for the full terms.