Workflow FLUX 3 Klein · two-pass HD… v1.0 (Archived)

FLUX 3 Klein workflow by me

Two-pass HD pipeline for FLUX 3 Image Klein, the best of my workflows for text

in images (signs, posters, album covers...).

━━━━━━━━━━━━━━━━━━━━━━━━━━━━

WHAT'S INSIDE

━━━━━━━━━━━━━━━━━━━━━━━━━━━━

- FLUX 3 Klein int8 with the official timestep shift (alpha 16)

- Optional BFL Prompt Upsampler: turns a short prompt into a detailed one

- Optional layout bounding boxes

- Two-pass HD + optional Ultimate SD Upscale 2x

- Resolution picker: just choose a ratio + megapixels

- Same clean design as my other workflows: numbered groups, group toggles, readable node titles

Built and tested on an RTX 4070 Laptop 8GB / 16GB RAM.

Flow: prompt -> (prompt upsampler) -> pass 1 -> latent 1.5x -> pass 2 HD -> color match -> save HD

-> USD 2x (group ②) -> color match -> save USD

Toggle each group on/off with the bars above the groups.

Settings:

- Timestep shift: ModelSamplingFlux 2.77 (alpha 16, official value)

- Pass 1: euler / simple, 16 steps, CFG 1

- Pass 2: latent upscale 1.5x (bislerp), 16 steps, denoise 0.45

- Color match (mkl 0.5) keeps the colors of pass 1

- USD: 2x with 4x Remacri, 4 steps, denoise 0.15, color match mkl 0.7

- Default size: 3:4 portrait, 1.0 MP. 2 MP also looks great (change it in "Resolution")

- Negative prompt has no effect (CFG 1)

Prompt Upsampler:

Uses Qwen3.5 9B to rewrite your prompt. It's slow with 16 GB of RAM: switch it

off in the node if you write detailed prompts yourself.

Models (accept the license on Hugging Face first):

- flux3-image-klein-int8-convrot.safetensors (diffusion model)

- qwen3vl_4b_int8_convrot.safetensors (text encoder, type flux3)

- qwen3.5_9b_flux3_prompt_upsampler_w4a8.safetensors (prompt upsampler, optional)

- flux2-vae.safetensors (VAE)

- 4x_foolhardy_Remacri.pth (upscaler)

Custom nodes:

Needs a recent ComfyUI (FLUX 3 support).

Description