FLUX 3 Klein Text-to-Image

FLUX 3 Image Klein official text-to-image workflow. A single detailed prompt renders a native 4K image with strong prompt adherence, from photoreal scenes to illustration, without any LoRA or fine-tune.

Main features:

- Built on FLUX 3 Image Klein, the 5.6B open-weights generator from Black Forest Labs (flux3_image_klein_base_int8_convrot + qwen3vl_4b text encoder + flux2 VAE)
- Native 4K output: the graph samples at 1024x1024 latents by default and the ResolutionSelector node lets you widen to any aspect while staying inside the model's native resolution envelope
- Base model path: 50 steps, CFG 3.5, euler/simple, denoise 1.0, fixed seed for reproducible comparisons
- CreateBoundingBoxes + Flux3LayoutPrompt are wired in, so the same graph can be switched from plain prompting to layout-guided prompting without rebuilding it
- No LoRA, no upscale pass: what you see is the raw model

Suggested workflow:

1. Type a full scene description into the text node (subject, setting, light, composition, style). Short prompts degrade badly on FLUX 3 - the model prefers detail.
2. Pick a resolution with ResolutionSelector, or edit the EmptyFlux2LatentImage width/height directly. Keep both divisible by 16.
3. Run the graph and save. For comparisons, keep the seed fixed and change one variable at a time.

Dependencies:

- flux3_image_klein_base_int8_convrot.safetensors (base UNET)
- qwen3.5_9b_flux3_prompt_upsampler_w4a8.safetensors (prompt upsampler)
- qwen3vl_4b_int8_convrot.safetensors (text encoder / CLIP)
- flux2-vae.safetensors (VAE)
- CreateBoundingBoxes, Flux3LayoutPrompt, ResolutionSelector (custom nodes)

RunningHub workflow page: https://www.runninghub.ai/zh-cn/user-center/1846601088051884033/webapp?inviteCode=rh-v1111

Bilibili tutorial video: https://www.bilibili.com/video/BV1btJb6LEGS/

RunningHub registration with invite code rh-v1111: https://www.runninghub.ai/?inviteCode=rh-v1111