Flux.3 Image Klein (NVFP4) Tests: Moonlit Sorceress, Cafe Rain, Cherry Blossom Sword

Test set made with FLUX.3 Image Klein (Base, NVFP4) in ComfyUI. Same model, same workflow, three anime scenes and one natural-light still life, to see how it handles both styles.

Models used:

- Diffusion model: flux3_image_klein_nvfp4.safetensors (Klein Base, NVFP4, ~5.3 GB)

- Text encoder: qwen3vl_4b_int8_convrot.safetensors (CLIPLoader type: flux3)

- VAE: flux2-vae.safetensors

- Source: black-forest-labs/flux-3-image-klein-comfy on Hugging Face (accept the license to download)

Settings:

- Canvas: 1024x1344 (portrait) / 1344x768 (wide), multiples of 16

- Sampler: euler, scheduler: simple (er_sde worth trying)

- Steps: 40 (official default is 25; 50 gives slightly more detail)

- CFG: 3.5 for anime, 2.5-3 for photo-style images

- Negative prompt: empty (default), batch size 1

- Nodes: UNETLoader > CLIPLoader (flux3) > CLIPTextEncode > EmptyFlux2LatentImage > KSampler > VAEDecode

- No LoRAs, no upscaler, no inpainting