FLUX 3 Image Klein 12GB — 5.6B int8 Text to Image + Edit with up to 4 Reference Images + BFL Prompt Upsampler

☕ One coffee = one happy creator: https://buymeacoffee.com/the_frizzy1

No video on this workflow yet — the one above is how to use my workflows, if you are new to them.

FLUX 3 Image Klein from Black Forest Labs, 5.6B distilled, on a 12 GB card: text to image with optional bounding boxes, and editing with 1 to 4 reference images, with BFL's own prompt upsampler on by default. Two workflow files, both BFL's official int8 graphs with my notes and a 1 MP default for 12 GB.

What you need

✔ Custom nodes — None. Core ComfyUI only.
✔ ComfyUI version — 0.41 or newer (FLUX 3 Image support). 0.43.1 tested, with sage attention.
✔ VRAM — 12 GB — tested, not estimated. Peak 11.4 GB, whole card as nvidia-smi reports it
✔ System RAM — 4.1 GB peak for ComfyUI, in a 32 GB machine
✔ Tested on — RTX 3060 12 GB · 32 GB RAM · ComfyUI 0.43.1 · sage attention
✔ Licence — Workflow MIT · FLUX 3 Image Klein: FLUX Kommunity Licence · Qwen3-VL-4B and the prompt upsampler: Apache-2.0

Models to download

Click a file name to download it. Put it in the folder shown, inside ComfyUI/models/.

Required — 4 files, 17.8 GB

diffusion_models/

File: flux3-image-klein-int8-convrot.safetensors — 5.8 GB (both files)
Download:
https://huggingface.co/black-forest-labs/flux-3-image-klein-comfy/resolve/main/diffusion_models/flux3-image-klein-int8-convrot.safetensors

Put it here:
ComfyUI/models/diffusion_models/

text_encoders/

File: qwen3vl_4b_int8_convrot.safetensors — 4.9 GB (both files)
Download:
https://huggingface.co/black-forest-labs/flux-3-image-klein-comfy/resolve/main/text_encoders/qwen3vl_4b_int8_convrot.safetensors

Put it here:
ComfyUI/models/text_encoders/

File: qwen3.5_9b_flux3_prompt_upsampler_w4a8.safetensors — 6.7 GB (prompt upsampler)
Download:
https://huggingface.co/black-forest-labs/flux-3-image-klein-comfy/resolve/main/text_encoders/qwen3.5_9b_flux3_prompt_upsampler_w4a8.safetensors

Put it here:
ComfyUI/models/text_encoders/

vae/

File: flux2-vae.safetensors — 0.3 GB (both files)
Download:
https://huggingface.co/black-forest-labs/flux-3-image-klein-comfy/resolve/main/vae/flux2-vae.safetensors

Put it here:
ComfyUI/models/vae/

The repo is gated: log in on Hugging Face and accept BFL's terms first, or nothing downloads.

Not tested by me: the bf16 model (11.1 GB) for 16 GB and bigger cards, and the nvfp4 model (5.3 GB) for RTX 50 cards.

Speed on my RTX 3060

Wall is queue to finished image. VRAM is the whole card as nvidia-smi reports it, about 0.7 GB of that is Windows.

• Text to image, first run after starting ComfyUI (loads the files) — 1 MP, 50 s wall, 11.0 GB peak vram, not measured peak ram (comfyui)

• Text to image, new prompt (upsampler + image) — 1 MP, 25 s wall, 11.2 GB peak vram, 3.8 GB peak ram (comfyui)

• Text to image, same prompt, new seed (upsampled prompt reused) — 1 MP, 14 s wall, 11.1 GB peak vram, 3.8 GB peak ram (comfyui)

• Text to image, same prompt — 4 MP, 1 min 16 s wall, 11.2 GB peak vram, 4.0 GB peak ram (comfyui)

• Edit, one image, new prompt — 1 MP, 41 s wall, 11.3 GB peak vram, 4.0 GB peak ram (comfyui)

• Edit, two images, new prompt — 1 MP, 1 min 5 s wall, 11.4 GB peak vram, 4.1 GB peak ram (comfyui)

Why the upsampler stays on

BFL's model card says to always run one: Klein was trained on long scene captions, and a short prompt typed as it is gives generic pictures. The upsampler (Qwen3.5 9B, w4a8) writes that caption for you, and in edit mode it looks at your images first. You see what it wrote in Upsampled prompt. Turn it off with enable upsampler only if you write the long caption yourself.

What did not work

• 4 MP runs on 12 GB, but the sky in my test image showed banding. My 1 MP results looked clean, so 1 MP is the setting I would stay on.

• Images 3 and 4 in the edit graph do nothing until you select them and press Ctrl+M. They ship muted.

• Short prompts give generic pictures. If your results look flat, check that enable upsampler is still on.

• RAM on the cold run was not measured, so that cell stays empty. I did not guess it.

• The licence is the other catch: FLUX Kommunity Licence, non-commercial, with commercial use of outputs only under US$5M revenue. Read it on the model page before you sell anything.

Get the workflow

• GitHub: https://github.com/Thefrizzy1/The_frizzy1-Workflows/tree/main/workflows/flux/flux3-klein-12gb

• Hugging Face mirror: https://huggingface.co/The-frizzy1/The_frizzy1-Workflows/tree/main/workflows/flux/flux3-klein-12gb

More videos on this stuff: https://www.youtube.com/playlist?list=UUGKDBMhTGqvlBZE9cF5d50Q